provsql.sql

Go to the documentation of this file.

/**
 * @file
 * @brief ProvSQL PL/pgSQL extension code
 *
 * This file contains the PL/pgSQL code of the ProvSQL extension. This
 * extension requires the standard UUID-ossp extension.
 */
 
/**
  * @brief <tt>provsql</tt> schema
  *
  * All types and functions introduced by ProvSQL are defined in the
  * provsql schema, requiring prefixing them by <tt>provsql.</tt> or
  * using PostgreSQL's <tt>search_path</tt> variable with a command such
  * as \code{.sql}SET search_path TO public, provsql;\endcode
  */
CREATE SCHEMA provsql;
 
SET search_path TO provsql;
 
/**
 * @brief Provenance circuit gate types
 *
 * Each gate in the provenance circuit has a type that determines
 * its semantics during semiring evaluation.
 */
CREATE TYPE PROVENANCE_GATE AS
  ENUM(
    'input',   -- Input (variable) gate of the circuit
    'plus',    -- Semiring plus
    'times',   -- Semiring times
    'monus',   -- M-Semiring monus
    'project', -- Project gate (for where provenance)
    'zero',    -- Semiring zero
    'one',     -- Semiring one
    'eq',      -- Equijoin gate (for where provenance)
    'agg',     -- Aggregation operator (for aggregate provenance)
    'semimod', -- Semimodule scalar multiplication (for aggregate provenance)
    'cmp',     -- Comparison of aggregate values (HAVING-clause provenance)
    'delta',   -- δ-semiring operator (see Amsterdamer, Deutch, Tannen, PODS 2011)
    'value',   -- Scalar value (for aggregate provenance)
    'mulinput',-- Multivalued input (for Boolean provenance)
    'update',  -- Update operation
    'rv',      -- Continuous random-variable leaf
    'arith',   -- n-ary arithmetic gate over scalar-valued children
    'mixture', -- Probabilistic mixture of two scalar RV roots with a Bernoulli weight
    'assumed', -- Structural assumption marker over a single child: the
                      -- wrapped sub-circuit was computed under the
                      -- assumption named by the gate's extra label --
                      -- 'BOOLEAN' (e.g. the safe-query rewrite; the
                      -- default when the label is absent) or
                      -- 'absorptive' (cyclic recursion truncated at the
                      -- absorptive value fixpoint).  Transparent for
                      -- evaluation semirings satisfying the assumption,
                      -- fatal error for the rest, rendered as an
                      -- explicit element in PROV-XML export.
    'annotation',     -- Transparent single-child wrapper carrying a
                      -- query-level annotation string in @c extra
                      -- (e.g. the inversion-free tractability
                      -- certificate / per-input order key).  Identity
                      -- for EVERY evaluator; its UUID folds in @c extra
                      -- so distinct annotations over the same child are
                      -- distinct gates.
    'conditioned',    -- Conditioning marker: two children
                      -- [target, evidence].  Evaluated only in the
                      -- measure interpretation: probability_evaluate
                      -- returns P(target ∧ evidence) / P(evidence); the
                      -- RV / AGG_TOKEN evaluators return the restricted
                      -- distribution.  For the UUID carrier it is a
                      -- TERMINAL gate (never a child of a semiring gate);
                      -- nested conditioning folds into a conjunction of
                      -- evidence.  Refused by every general sr_* semiring
                      -- (normalization is not a semiring operation).
    'mobius',         -- Signed Möbius combination over child islands: one
                      -- INTEGER coefficient per child in @c extra (the
                      -- gate_arith precedent), probability_evaluate returns
                      -- Σ_i coeff_i · P(child_i).  The one new primitive of
                      -- the safe-UCQ Möbius-inversion route, evaluated only
                      -- in the measure interpretation; refused by every
                      -- general sr_* semiring (a signed combination is not a
                      -- semiring operation).
    'case',           -- N-ary guarded selection over scalar (RV) children:
                      -- wires [guard_1, value_1, ..., guard_k, value_k,
                      -- default], first-match semantics (the value of the
                      -- first guard event that holds, else the default).
                      -- Backs a CASE expression over random variables (and
                      -- abs / clamp / ReLU as sugar).  RV/measure-carrier;
                      -- refused by every general sr_* semiring.
    'observe'         -- Latent-variable observation (likelihood-weighting
                      -- evidence): one wire -> an observed bare gate_rv
                      -- leaf, the datum in extra.  Contributes a
                      -- continuous density factor (the leaf's pdf at the
                      -- datum) rather than a Boolean truth value,
                      -- composing into an evidence circuit by gate_times
                      -- exactly like a conditioning event.  Evaluated only
                      -- by the importance-sampling weight walk; refused by
                      -- every Boolean / semiring evaluator.
    );
 
/** @defgroup gate_manipulation Circuit gate manipulation
 *  Low-level functions for creating and querying provenance circuit gates.
 *  @{
 */
 
/**
 * @brief Create a new gate in the provenance circuit
 *
 * @param token UUID identifying the new gate
 * @param type gate type (see PROVENANCE_GATE)
 * @param children optional array of child gate UUIDs
 */
CREATE OR REPLACE FUNCTION create_gate(
  token UUID,
  type PROVENANCE_GATE,
  children UUID[] DEFAULT NULL)
  RETURNS VOID AS
  'provsql','create_gate' LANGUAGE C PARALLEL SAFE;
/**
 * @brief Return the gate type of a provenance token
 *
 * Returns @c 'input' for any token not yet materialized in the circuit,
 * since input is the default semantics of an unmaterialized provenance token.
 */
CREATE OR REPLACE FUNCTION get_gate_type(
  token UUID)
  RETURNS PROVENANCE_GATE AS
  'provsql','get_gate_type' LANGUAGE C IMMUTABLE PARALLEL SAFE;
/** @brief Return the children of a provenance gate */
CREATE OR REPLACE FUNCTION get_children(
  token UUID)
  RETURNS UUID[] AS
  'provsql','get_children' LANGUAGE C IMMUTABLE PARALLEL SAFE;
/**
 * @brief Set the probability of an input gate
 *
 * @param token UUID of the input gate
 * @param p probability value in [0,1]
 */
CREATE OR REPLACE FUNCTION set_prob(
  token UUID, p DOUBLE PRECISION)
  RETURNS VOID AS
  'provsql','set_prob' LANGUAGE C PARALLEL SAFE;
/** @brief Get the probability associated with an input gate */
CREATE OR REPLACE FUNCTION get_prob(
  token UUID)
  RETURNS DOUBLE PRECISION AS
  'provsql','get_prob' LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief Set additional INTEGER values on provenance circuit gate
 *
 * This function sets two INTEGER values associated to a circuit gate, used in
 * different ways by different gate types:
 *   - for mulinput, info1 indicates the value of this multivalued variable
 *   - for eq, info1 and info2 indicate the attribute index of the
       equijoin in, respectively, the first and second columns
 *   - for agg, info1 is the oid of the aggregate function and info2 the
       oid of the aggregate result type
 *   - for cmp, info1 is the oid of the comparison operator
 *
 * @param token UUID of the circuit gate
 * @param info1 first INTEGER value
 * @param info2 second INTEGER value
 */
CREATE OR REPLACE FUNCTION set_infos(
  token UUID, info1 INT, info2 INT DEFAULT NULL)
  RETURNS VOID AS
  'provsql','set_infos' LANGUAGE C PARALLEL SAFE;
 
/** @brief Get the INTEGER info values associated with a circuit gate */
CREATE OR REPLACE FUNCTION get_infos(
  token UUID, OUT info1 INT, OUT info2 INT)
  RETURNS RECORD AS
  'provsql','get_infos' LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief Wrap @p token in a fresh @c gate_assumed carrying @p assumption
 *        as its label, and return the wrapper's UUID.
 *
 * Public primitive callable from any rewrite or driver that needs to
 * flag a sub-circuit as sound only under an evaluation assumption:
 *
 * - @c 'BOOLEAN' -- the sub-circuit only preserves the Boolean function
 *   of the lineage (e.g. the safe-query rewrite collapses derivation
 *   multiplicities); transparent for semirings admitting a homomorphism
 *   from Boolean functions.
 * - @c 'absorptive' -- the sub-circuit was truncated at the absorptive
 *   value fixpoint (cyclic recursive query); transparent for absorptive
 *   semirings (probability, BOOLEAN, min-plus over nonnegative
 *   costs...), fatal for the rest (counting, why-provenance).
 *
 * Incompatible evaluators raise a @c CircuitException.  Always kept as
 * an explicit node in PROV-XML export.
 *
 * The wrapper UUID is content-derived via @c uuid_generate_v5 on the
 * assumption and the child, so identical children always wrap to the
 * same outer UUID per assumption.  No-op (returns NULL) on a NULL
 * input.
 */
CREATE OR REPLACE FUNCTION provenance_assume(token UUID, assumption TEXT)
  RETURNS UUID AS
$$
DECLARE
  wrapped UUID;
BEGIN
  IF token IS NULL THEN
    RETURN NULL;
  END IF;
  IF assumption NOT IN ('BOOLEAN', 'absorptive') THEN
    RAISE EXCEPTION 'provenance_assume: unknown assumption %', assumption;
  END IF;
  wrapped := public.uuid_generate_v5(uuid_ns_provsql(),
                                     concat('assumed', assumption, token));
  PERFORM create_gate(wrapped, 'assumed', ARRAY[token]);
  PERFORM set_extra(wrapped, assumption);
  RETURN wrapped;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Wrap @p token in a Boolean-assumption marker (compatibility
 *        name; see @c provenance_assume).
 *
 * This is the entry point the safe-query (read-once) rewriter calls on
 * every per-row root it produces, so the wrapper additionally carries
 * @c PROVSQL_ROUTE_SQ_REWRITE in @c info1: the assumption kind alone does
 * not identify the route (@c provenance_assume(t, 'BOOLEAN') is public),
 * and the probability dispatcher reads the tag back to report
 * @c sq-rewrite rather than the generic @c independent.  Build an untagged
 * Boolean-assumption wrapper with @c provenance_assume directly.
 */
CREATE OR REPLACE FUNCTION assume_boolean(token UUID) RETURNS UUID AS
$$
DECLARE
  wrapped UUID;
BEGIN
  wrapped := provenance_assume(token, 'BOOLEAN');
  IF wrapped IS NOT NULL THEN
    -- 1 = PROVSQL_ROUTE_SQ_REWRITE (see provsql_route in src/provsql_utils.h)
    PERFORM set_infos(wrapped, 1, 0);
  END IF;
  RETURN wrapped;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Wrap @p token in a fresh transparent @c gate_annotation carrying
 *        @p extra, and return the wrapper's UUID.
 *
 * Unlike every other gate, the annotation wrapper's UUID folds in @p extra
 * (not just the child): @c uuid_generate_v5 over @c concat('annotation',
 * token, extra).  This is deliberate -- two annotations over the same child
 * with different @p extra must be distinct gates (e.g. the same input tuple
 * carrying different per-occurrence order keys, or two queries attaching
 * different certificates to a shared root).  The wrapper is transparent
 * (identity) for EVERY evaluator; @p extra is inert metadata read only by the
 * code that placed it.  No-op (returns NULL) on a NULL input.
 */
CREATE OR REPLACE FUNCTION annotate(token UUID, extra TEXT) RETURNS UUID AS
$$
DECLARE
  annotated UUID;
BEGIN
  IF token IS NULL THEN
    RETURN NULL;
  END IF;
  annotated := public.uuid_generate_v5(uuid_ns_provsql(),
                                       concat('annotation', token, extra));
  PERFORM create_gate(annotated, 'annotation', ARRAY[token]);
  PERFORM set_extra(annotated, extra);
  RETURN annotated;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Peel every transparent @c gate_annotation wrapper off @p token,
 *        returning the first non-annotation gate underneath.
 *
 * The dual of @c annotate, for consumers keyed to gate *identity* rather
 * than gate value: a provenance mapping matches input-gate UUIDs, and the
 * reachability edge classifier matches token shapes, so both must see the
 * wrapped gate, not the wrapper (e.g. the inversion-free certificate /
 * order marker the planner attaches to a certified query's row roots).
 * Identity on a token with no annotation wrapper; NULL on NULL.
 */
CREATE OR REPLACE FUNCTION strip_annotations(token UUID) RETURNS UUID AS
$$
WITH RECURSIVE peel(g) AS (
  SELECT token
  UNION ALL
  SELECT (provsql.get_children(p.g))[1] FROM peel p
  WHERE provsql.get_gate_type(p.g) = 'annotation'
)
SELECT g FROM peel WHERE provsql.get_gate_type(g) <> 'annotation' LIMIT 1;
$$ LANGUAGE sql STABLE PARALLEL SAFE;
 
/**
 * @brief Condition a provenance token (a Boolean event) on another.
 *
 * Builds the terminal @c gate_conditioned that the measure evaluators read
 * as @c "P(target ∧ evidence) / P(evidence)".  This is the backing function
 * of the binary @c | operator (@c "target | evidence", value-level
 * conditioning of the UUID carrier).
 *
 * The gate stores three children @c [target, evidence, joint] with
 * @c joint @c = @c times(target, @c evidence); evaluation is then the plain
 * ratio @c P(joint)/P(evidence), and content-addressing makes a base tuple
 * shared by @p target and @p evidence the same input gate in both circuits,
 * so the conditional is exact and correlation-aware.
 *
 * Conventions:
 *  - Conditioning on a certain or absent event is a no-op: @c evidence NULL
 *    or @c gate_one() returns @p target unchanged (@c "P(X|true)=P(X)").
 *  - A @p target with no provenance defaults to the certain event 1, so
 *    @c "1 | c" is the well-defined certain-row posterior.
 *  - Nested conditioning folds (sequential Bayesian update):
 *    @c "(X | A) | B = X | (A ∧ B)" -- the gate never nests, it stays one
 *    level deep with the evidence accumulated by @c times.
 *
 * The result is TERMINAL: a conditioned token may not become a child of a
 * @c plus / @c times / @c monus / @c agg gate (those constructors refuse
 * it); the only operation it admits is more conditioning.
 */
CREATE OR REPLACE FUNCTION cond(target UUID, evidence UUID) RETURNS UUID AS
$$
DECLARE
  tgt UUID;
  ev  UUID;
  jnt UUID;
  result UUID;
  ch UUID[];
BEGIN
  -- P(X | true) = P(X): conditioning on a certain / absent event is inert.
  IF evidence IS NULL OR evidence = gate_one() THEN
    RETURN target;
  END IF;
 
  -- A row with no provenance defaults to the certain event 1.
  tgt := coalesce(target, gate_one());
 
  IF get_gate_type(tgt) = 'conditioned' THEN
    -- Sequential update (X | A) | B = X | (A ∧ B): fold B into both the
    -- evidence and the joint of the inner gate so the result stays a single
    -- gate_conditioned over the ORIGINAL target.
    ch  := get_children(tgt);
    tgt := ch[1];                              -- original target X
    ev  := provenance_times(ch[2], evidence);  -- A ∧ B
    jnt := provenance_times(ch[3], evidence);  -- (X ∧ A) ∧ B
  ELSE
    ev  := evidence;
    jnt := provenance_times(tgt, evidence);    -- X ∧ C
  END IF;
 
  result := public.uuid_generate_v5(uuid_ns_provsql(),
                                    concat('conditioned', tgt, ev, jnt));
  PERFORM create_gate(result, 'conditioned', ARRAY[tgt, ev, jnt]);
  RETURN result;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Binary @c | : value-level conditioning, @c "target | evidence".
 *
 * Carrier-parametric in its left operand; the UUID form builds the terminal
 * @c gate_conditioned via @c cond.  Does not collide with core PostgreSQL's
 * INTEGER bitwise @c | (different argument types).
 */
CREATE OPERATOR | (LEFTARG=UUID, RIGHTARG=UUID, PROCEDURE=cond);
 
/**
 * @brief Placeholder for @c "X | (predicate)" on a UUID event.
 *
 * Lets the conditioning event be written as a natural Boolean combination of
 * random_variable / aggregate comparisons (e.g. @c "event | (sensor > 3)")
 * instead of a hand-built gate.  Never executes: the ProvSQL planner hook
 * converts the Boolean operand into a condition gate and emits @c cond.
 */
CREATE OR REPLACE FUNCTION cond_predicate(target UUID, predicate BOOLEAN)
  RETURNS UUID AS
$$
BEGIN
  RAISE EXCEPTION 'UUID | (predicate) must be rewritten by the ProvSQL '
    'planner hook: the right operand must be a Boolean combination of '
    'random_variable / aggregate comparisons (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR | (LEFTARG=UUID, RIGHTARG=BOOLEAN, PROCEDURE=cond_predicate);
 
/**
 * @brief Placeholder for @c "(predicate) | (predicate)" on two events.
 *
 * Conditions one comparison event on another when both operands are written
 * as comparisons rather than pre-built tokens (e.g.
 * @c "probability((x >= 2000) | (x >= 1000))"): an @c random_variable /
 * @c AGG_TOKEN comparison is statically @c BOOLEAN-typed, so neither the
 * @c "UUID | UUID" (@c cond) nor the @c "UUID | BOOLEAN" (@c cond_predicate)
 * operator resolves.  Never executes: the ProvSQL planner hook lowers each
 * Boolean operand to its event gate and emits @c cond(target, evidence), so
 * the result carries the correlation-aware @c Pr(A ∧ B) / Pr(B).  Returns
 * @c UUID, so @c "A | B" is a first-class event token in every position
 * (a @c probability(UUID) argument, a projected column, a further @c "|").
 */
CREATE OR REPLACE FUNCTION predicate_cond_predicate(target BOOLEAN, evidence BOOLEAN)
  RETURNS UUID AS
$$
BEGIN
  RAISE EXCEPTION '(predicate) | (predicate) must be rewritten by the ProvSQL '
    'planner hook: both operands must be Boolean combinations of '
    'random_variable / aggregate comparisons (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR | (LEFTARG=BOOLEAN, RIGHTARG=BOOLEAN, PROCEDURE=predicate_cond_predicate);
 
/**
 * @brief Deterministic indicator gate for an ordinary (regular) comparison.
 *
 * The predicate-provenance of an ordinary comparison (both sides of regular
 * type, e.g. @c "region = 'north'") is the deterministic indicator
 * @c "χ(cond)": @c gate_one() when the comparison holds on the row,
 * @c gate_zero() otherwise (Definition in the HAVING-provenance semantics).
 * The planner emits this for a regular comparison appearing inside a MIXED
 * conditioning predicate (one that also has a random_variable / aggregate
 * comparison); @c cond is evaluated per row, so the indicator is the row's
 * own truth value, combined by @c ⊗ / @c ⊕ with the probabilistic gates.
 */
CREATE OR REPLACE FUNCTION regular_indicator(cond BOOLEAN) RETURNS UUID AS
$$
  SELECT CASE WHEN cond THEN provsql.gate_one() ELSE provsql.gate_zero() END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Whole-tuple output conditioning directive: @c "given(evidence)".
 *
 * Written as a term in the select list, @c given(c) conditions the OUTPUT
 * provenance of the current query's rows on @p c:
 *
 * @code
 *   SELECT a, b, given((SELECT provenance() FROM tests
 *                       WHERE patient_id = s.id AND result = 'positive'))
 *   FROM source s;
 *   -- visible columns: a, b   (the given(...) term is stripped)
 *   -- per-row output provenance: provenance() | <that row's evidence>
 * @endcode
 *
 * The query rewriter recognises the marker, STRIPS it from the visible
 * projection, and wraps each output row's provenance expression in
 * @c cond(row_provenance, c) -- deriving a new conditioned relation, never
 * mutating any stored provenance.  @p c is evaluated per output row and may
 * correlate with the row's columns, so each tuple is conditioned on its own
 * evidence.  When the rewriter is inactive the call is a harmless identity
 * (it returns @p evidence as an ordinary column).
 *
 * When @b executed rather than stripped -- i.e. nested inside an expression,
 * the idiom @c "and_agg(given(Y = d))" that folds one observation per row
 * into a latent-variable evidence circuit -- a point-equality @c "Y = d" on
 * a bare random-variable leaf is turned into likelihood-weighting evidence
 * (@c evidence_as_observation); any other evidence passes through unchanged.
 */
CREATE OR REPLACE FUNCTION given(evidence UUID) RETURNS UUID AS
$$
BEGIN
  RETURN provsql.evidence_as_observation(evidence);
END
$$ LANGUAGE plpgsql VOLATILE PARALLEL SAFE
   SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/**
 * @brief Prefix unary @c | : alias for @c given, @c "| evidence".
 *
 * Disambiguated from the binary @c | by the absence of a left operand
 * (@c "a, | c" parses @c "| c" as the prefix form).  PostgreSQL keeps
 * prefix operators on every supported version (postfix operators were
 * removed in PG14), so @c "| c" is safe across the CI matrix.
 */
CREATE OPERATOR | (RIGHTARG=UUID, PROCEDURE=given);
 
/**
 * @brief Conditioning-evidence from a predicate: @c "given(predicate)"
 *        (also the prefix @c "| (predicate)").
 *
 * Two uses of the same marker:
 * - whole-tuple output conditioning written as a select-list term, the
 *   natural-predicate spelling of @c given (@c "SELECT a, given(sensor > 3)");
 * - per-row evidence for a latent-variable posterior, folded with
 *   @c and_agg -- @c "and_agg(given(normal(mu,1) = x))" turns each row's
 *   observation into likelihood-weighting evidence.
 *
 * Never executes: the planner converts the Boolean operand into a condition
 * gate and emits @c given(gate); a point-equality @c "Y = d" on a bare
 * random-variable leaf then becomes an observation (see @c given(UUID) /
 * @c evidence_as_observation).
 */
CREATE OR REPLACE FUNCTION given(predicate BOOLEAN) RETURNS UUID AS
$$
BEGIN
  RAISE EXCEPTION 'given(predicate) / prefix | (predicate) must be rewritten '
    'by the ProvSQL planner hook: the operand must be a Boolean combination '
    'of random_variable / aggregate comparisons (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR | (RIGHTARG=BOOLEAN, PROCEDURE=given);
 
/**
 * @brief Event negation: @c "! event" / @c "provenance_not(event)".
 *
 * The complement of a Boolean provenance event: @c "!x" holds in exactly the
 * worlds where @p x does not.  It is sugar for @c "monus(one, x)" -- an
 * ordinary m-semiring expression (Boolean @c NOT, probability @c "1 - P(x)"),
 * NOT a measure-only marker -- so it composes like any @c monus, and a
 * conditioned / terminal token is refused as its child (so @c "!(x | c)"
 * errors, as conditioning cannot be buried under further algebra).
 *
 * The motivating use is conditioning on the NON-occurrence of an arbitrary
 * violation query @p W (a denial constraint), where @p W itself is built with
 * ordinary idioms and needs no hand-rolled gates:
 *
 * @code
 *   -- W = "some pair of overlapping same-room bookings is present"
 *   WITH w AS (SELECT provenance() AS tok
 *              FROM bookings a JOIN bookings b
 *                ON a.id < b.id AND a.room = b.room
 *                   AND a.lo < b.hi AND b.lo < a.hi
 *              GROUP BY ())
 *   SELECT probability_evaluate((SELECT provenance() FROM bookings WHERE id=1)
 *                               | !w.tok)             -- P(booking 1 | no overlap)
 *   FROM w;
 * @endcode
 *
 * Named @c provenance_not, after the @c "provenance_times / _plus / _monus"
 * family; the prefix @c ! operator is the ergonomic form (SQL's reserved
 * @c NOT keyword cannot serve as a function name).
 */
CREATE OR REPLACE FUNCTION provenance_not(event UUID) RETURNS UUID AS
$$
  SELECT provsql.provenance_monus(provsql.gate_one(), event);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE
   SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Prefix unary @c ! : alias for @c provenance_not, @c "! event".
 *
 * Prefix operators are kept on every supported PostgreSQL version (postfix
 * operators were removed in PG14), and core PG defines no prefix @c !  on
 * @c UUID, so @c "! event" is safe across the CI matrix.
 */
CREATE OPERATOR ! (RIGHTARG=UUID, PROCEDURE=provenance_not);
 
/**
 * @brief Build a per-input order-key string for the inversion-free path.
 *
 * Emitted by the planner per certified atom: @c K-prefixed, length-prefixed
 * @c "K<factor> <octet_length(root)>:<root><octet_length(sec)>:<sec>", parsed
 * back at evaluation by @c safe_cert_key_parse.  @p root / @p sec are the
 * tuple's root- and secondary-class column values (TEXT-cast by the caller);
 * the byte-length prefixes keep the values unambiguous for @em any column type,
 * including TEXT containing spaces, colons or digits.  @p factor is the atom's
 * factor id (or -1 for the shared self-join guard).  @c IMMUTABLE so the planner
 * can fold it and the marker dedups by content-addressing.
 */
CREATE OR REPLACE FUNCTION inversion_free_key(root TEXT, sec TEXT, factor INT)
  RETURNS TEXT AS
$$ SELECT 'K' || factor::TEXT || ' '
       || octet_length(root) || ':' || root
       || octet_length(sec)  || ':' || sec $$
  LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Set extra TEXT information on provenance circuit gate
 *
 * This function sets TEXT-encoded data associated to a circuit gate, used in
 * different ways by different gate types:
 *   - for project, it is a TEXT-encoded ARRAY of two-element ARRAYs that
 *     indicate mappings between input attribute (first element) and output
 *     attribute (second element)
 *   - for value and agg, it is the TEXT-encoded (base for value, computed
 *     for agg) scalar value
 *
 * @param token UUID of the circuit gate
 * @param data TEXT-encoded information
 */
CREATE OR REPLACE FUNCTION set_extra(
  token UUID, data TEXT)
  RETURNS VOID AS
  'provsql','set_extra' LANGUAGE C PARALLEL SAFE STRICT;
/** @brief Get the TEXT-encoded extra data associated with a circuit gate */
CREATE OR REPLACE FUNCTION get_extra(token UUID)
  RETURNS TEXT AS
  'provsql','get_extra' LANGUAGE C STABLE PARALLEL SAFE RETURNS NULL ON NULL INPUT;
 
/**
 * @brief Return the total number of materialized gates in the provenance circuit
 *
 * Input gates for provenance-tracked table rows are created lazily on
 * first reference; rows that have never appeared in a query result are
 * not counted.
 */
CREATE OR REPLACE FUNCTION get_nb_gates() RETURNS BIGINT AS
  'provsql', 'get_nb_gates' LANGUAGE C PARALLEL SAFE;
 
/** @} */
 
/** @defgroup table_management Provenance table management
 *  Functions for enabling, disabling, and configuring provenance
 *  tracking on user tables.
 *  @{
 */
 
 
/**
 * @brief Trigger function for DELETE statement provenance tracking
 *
 * Records the deletion and applies monus to provenance tokens of
 * deleted rows. This is the version for PostgreSQL < 14.
 */
CREATE OR REPLACE FUNCTION delete_statement_trigger()
  RETURNS TRIGGER AS
$$
DECLARE
  query_text TEXT;
  delete_token UUID;
  old_token UUID;
  new_token UUID;
  r RECORD;
BEGIN
  delete_token := public.uuid_generate_v4();
 
  PERFORM create_gate(delete_token, 'input');
 
  SELECT query
  INTO query_text
  FROM pg_stat_activity
  WHERE pid = pg_backend_pid();
 
  INSERT INTO delete_provenance (delete_token, query, deleted_by, deleted_at)
  VALUES (delete_token, query_text, current_user, CURRENT_TIMESTAMP);
 
  EXECUTE format('INSERT INTO %I.%I SELECT * FROM OLD_TABLE;', TG_TABLE_SCHEMA, TG_TABLE_NAME);
 
  FOR r IN (SELECT * FROM OLD_TABLE) LOOP
    old_token := r.provsql;
    new_token := provenance_monus(old_token, delete_token);
 
    EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2;', TG_TABLE_SCHEMA, TG_TABLE_NAME)
    USING new_token, old_token;
  END LOOP;
 
  RETURN NULL;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp SECURITY DEFINER;
 
 
/**
 * @brief Record per-relation provenance metadata used by the
 *        safe-query optimisation.
 *
 * Stores a @c (relid, kind, block_key) RECORD in the persistent
 * mmap-backed table-info store.  @p kind is one of:
 *   - @c 'tid' -- independent input leaves (post-@c add_provenance default)
 *   - @c 'bid' -- block-correlated leaves; rows sharing the same value
 *                 of @p block_key are mutually exclusive.  An empty
 *                 @p block_key means the whole table is one block.
 *   - @c 'opaque' -- arbitrary correlations from a derived source
 *                 (CREATE TABLE AS SELECT, INSERT INTO SELECT,
 *                 UPDATE under provsql.update_provenance); the
 *                 safe-query rewriter must bail on these.
 *
 * @param relid     pg_class OID of the relation.
 * @param kind      One of @c 'tid' / @c 'bid' / @c 'opaque'.
 * @param block_key Block-key column numbers (only meaningful for
 *                  @c 'bid'; ignored otherwise but conventionally
 *                  passed empty).
 */
CREATE OR REPLACE FUNCTION set_table_info(
  relid OID, kind TEXT, block_key INT2[] DEFAULT ARRAY[]::INT2[])
  RETURNS VOID AS
  'provsql','set_table_info' LANGUAGE C PARALLEL SAFE;
 
/** @brief Remove per-relation provenance metadata.  No-op when missing. */
CREATE OR REPLACE FUNCTION remove_table_info(relid OID)
  RETURNS VOID AS
  'provsql','remove_table_info' LANGUAGE C PARALLEL SAFE;
 
/**
 * @brief Read per-relation provenance metadata.
 *
 * Returns NULL if no RECORD exists.  @c kind is one of @c 'tid' /
 * @c 'bid' / @c 'opaque'; @c block_key is the (possibly empty) array
 * of block-key column numbers, only meaningful when @c kind = @c 'bid'.
 * Used by the planner-time hierarchy detector to gate the safe-query
 * rewrite.
 */
CREATE OR REPLACE FUNCTION get_table_info(
  relid OID, OUT kind TEXT, OUT block_key INT2[])
  RETURNS RECORD AS
  'provsql','get_table_info' LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief Record the base-relation ancestor set of a tracked relation.
 *
 * Base tables created with @c add_provenance / @c repair_key carry
 * @c {self}; CTAS-derived tables inherit the union of their sources'
 * ancestor sets.  The safe-query rewriter consults the registry to
 * enforce that joined FROM entries have disjoint base ancestors
 * before firing the read-once factoring.
 *
 * The worker preserves the relation's existing @c kind / @c block_key
 * half on update; it silently no-ops when no kind RECORD exists for
 * @p relid (callers should run @c add_provenance / @c repair_key
 * first).  The ancestor list is capped at 64 entries (clear error if
 * exceeded).
 *
 * @param relid      pg_class OID of the relation.
 * @param ancestors  Sorted, deduplicated base-relation OIDs.
 */
CREATE OR REPLACE FUNCTION set_ancestors(
  relid OID, ancestors OID[] DEFAULT ARRAY[]::OID[])
  RETURNS VOID AS
  'provsql','set_ancestors' LANGUAGE C PARALLEL SAFE;
 
/** @brief Clear the ancestor half of a per-relation RECORD (keeps kind/block_key).
 *  No-op when missing. */
CREATE OR REPLACE FUNCTION remove_ancestors(relid OID)
  RETURNS VOID AS
  'provsql','remove_ancestors' LANGUAGE C PARALLEL SAFE;
 
/**
 * @brief Read the base-relation ancestor set of a tracked relation.
 *
 * Returns @c NULL when no ancestor RECORD exists for @p relid (or the
 * RECORD is empty -- both cases make the safe-query rewriter take
 * its conservative refuse path, so they collapse here).
 */
CREATE OR REPLACE FUNCTION get_ancestors(relid OID)
  RETURNS OID[] AS
  'provsql','get_ancestors' LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief BEFORE INSERT OR UPDATE OF provsql row trigger installed by
 *        @c add_provenance.
 *
 * Two jobs:
 *
 *  1. Fill @c NEW.provsql with a fresh @c uuid_generate_v4 leaf when
 *     the user did not supply one (a column DEFAULT would not do here:
 *     it fires before the trigger sees the row, so we could not tell
 *     "user omitted the column" from "user supplied a value").
 *  2. When the user does supply a non-NULL @c provsql on @c INSERT,
 *     or changes it on @c UPDATE, flip the table's per-table
 *     metadata to @c OPAQUE.  The user is free to write whatever
 *     UUIDs they want (cross-table reuse, compound tokens minted
 *     via @c create_gate, ...); the cost is that the safe-query
 *     rewriter then refuses to fire on this table, because TID
 *     independence can no longer be assumed.
 */
CREATE OR REPLACE FUNCTION provenance_guard()
  RETURNS TRIGGER AS $$
DECLARE
  _m RECORD;
BEGIN
  IF TG_OP = 'INSERT' THEN
    IF NEW.provsql IS NULL THEN
      -- A genuine insert: mint a fresh atomic input variable. This is the
      -- one place a new input token is born, so it is also where any
      -- maintained mapping on this table is extended (keyed to that token).
      -- Data-modification re-insertions (INSERT ... SELECT * FROM OLD_TABLE)
      -- carry a supplied provsql and take the ELSE branch, so they are
      -- correctly skipped: the validity stays keyed to the original input,
      -- which is exactly the child a later monus/update gate wraps.
      NEW.provsql := public.uuid_generate_v4();
      FOR _m IN SELECT mapping, attribute
                  FROM provsql.provenance_mapping_registry WHERE source = TG_RELID
      LOOP
        EXECUTE format(
          'INSERT INTO %s(value, provenance) SELECT ($1).%I, $2',
          _m.mapping::REGCLASS, _m.attribute)
          USING NEW, NEW.provsql;
      END LOOP;
    ELSE
      PERFORM provsql.set_table_info(TG_RELID, 'opaque');
    END IF;
  ELSIF TG_OP = 'UPDATE' THEN
    IF NEW.provsql IS DISTINCT FROM OLD.provsql THEN
      PERFORM provsql.set_table_info(TG_RELID, 'opaque');
    END IF;
  END IF;
  RETURN NEW;
END;
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER;
 
/**
 * @brief Enable provenance tracking on an existing table
 *
 * Adds a <tt>provsql</tt> UUID column to the table, an index for
 * fast UUID-keyed lookups, and a BEFORE INSERT/UPDATE row trigger
 * (@c provenance_guard) that mints a fresh @c uuid_generate_v4
 * leaf when the user omits the column on INSERT, or flips the
 * table's metadata to @c OPAQUE when the user supplies their own
 * value.  Input gates for existing rows are created lazily when
 * first referenced by a query.
 *
 * @param _tbl the table to add provenance tracking to
 */
CREATE OR REPLACE FUNCTION add_provenance(_tbl REGCLASS)
  RETURNS VOID AS
$$
BEGIN
  -- Idempotence: a second add_provenance on an already-tracked table is
  -- a no-op with a NOTICE, so setup scripts and notebook cells can be
  -- re-run freely.
  IF EXISTS (
    SELECT 1 FROM pg_attribute
    WHERE attrelid = _tbl AND attname = 'provsql' AND NOT attisdropped
  ) THEN
    RAISE NOTICE 'table % already has provenance tracking', _tbl;
    RETURN;
  END IF;
  -- No DEFAULT: the guard trigger mints the UUID, so the trigger can
  -- distinguish "user omitted" (NULL) from "user supplied a value".
  -- No UNIQUE: we no longer rely on it to keep the table TID -- the
  -- guard does that semantically -- and a UNIQUE would reject the
  -- legitimate cross-table UUID copy that just flips the table to
  -- OPAQUE.  We keep a plain index for fast UUID-keyed lookups.
  EXECUTE format('ALTER TABLE %s ADD COLUMN provsql UUID', _tbl);
  EXECUTE format(
    'UPDATE %s SET provsql = public.uuid_generate_v4() WHERE provsql IS NULL',
    _tbl);
  EXECUTE format('CREATE INDEX ON %s(provsql)', _tbl);
  EXECUTE format(
    'CREATE TRIGGER provenance_guard BEFORE INSERT OR UPDATE OF provsql '
    'ON %s FOR EACH ROW EXECUTE PROCEDURE provsql.provenance_guard()',
    _tbl);
  PERFORM provsql.set_table_info(_tbl::oid, 'tid');
  -- Seed the base-ancestor set to {self}: a base TID table's atoms
  -- come from itself and no other relation.  CTAS-derived tables
  -- inherit unions of source ancestor sets; that is handled by the
  -- CTAS hook (a separate slice), not here.
  PERFORM provsql.set_ancestors(_tbl::oid, ARRAY[_tbl::oid]);
END
$$ LANGUAGE plpgsql SECURITY DEFINER;
 
/**
 * @brief Remove provenance tracking from a table
 *
 * Drops the <tt>provsql</tt> column and associated triggers.
 *
 * @param _tbl the table to remove provenance tracking from
 */
CREATE OR REPLACE FUNCTION remove_provenance(_tbl REGCLASS)
  RETURNS VOID AS
$$
DECLARE
BEGIN
  PERFORM provsql.remove_table_info(_tbl::oid);
  -- Idempotence, mirroring add_provenance: removing provenance from a
  -- table that does not have it is a NOTICE-and-no-op, so setup scripts
  -- and notebook cells can be re-run freely.  The metadata strip above
  -- still runs, so a table left half-tracked is cleaned up.
  IF NOT EXISTS (
    SELECT 1 FROM pg_attribute
    WHERE attrelid = _tbl AND attname = 'provsql' AND NOT attisdropped
  ) THEN
    RAISE NOTICE 'table % does not have provenance tracking', _tbl;
    RETURN;
  END IF;
  -- Drop the BEFORE INSERT/UPDATE guard first: it has a column
  -- dependency on provsql (via the OF provsql clause), so the
  -- subsequent DROP COLUMN would otherwise raise.
  BEGIN
    EXECUTE format('DROP TRIGGER provenance_guard on %s', _tbl);
  EXCEPTION WHEN undefined_object THEN
  END;
  EXECUTE format('ALTER TABLE %s DROP COLUMN provsql', _tbl);
  BEGIN
    EXECUTE format('DROP TRIGGER add_gate on %s', _tbl);
  EXCEPTION WHEN undefined_object THEN
  END;
  BEGIN
    EXECUTE format('DROP TRIGGER insert_statement on %s', _tbl);
    EXECUTE format('DROP TRIGGER update_statement on %s', _tbl);
    EXECUTE format('DROP TRIGGER delete_statement on %s', _tbl);
  EXCEPTION WHEN undefined_object THEN
  END;
END
$$ LANGUAGE plpgsql;
 
/**
 * @brief Set up provenance for a table with duplicate key values
 *
 * When a table has duplicate rows for a given key, this function
 * replaces simple input gates with multivalued input (mulinput) gates
 * that model a uniform distribution over duplicates.
 *
 * @param _tbl the table to repair
 * @param key_att the key attribute(s) as a comma-separated string, or
 *        empty string if the whole table is one group
 */
CREATE OR REPLACE FUNCTION repair_key(_tbl REGCLASS, key_att TEXT)
  RETURNS VOID AS
$$
DECLARE
  r RECORD;
  rows_query TEXT;
  block_key_cols INT2[];
BEGIN
  -- Resolve the (possibly comma-separated) key_att TEXT into the
  -- corresponding pg_attribute.attnum values for the safe-query
  -- metadata.  Names are trimmed; quoting is not supported because
  -- repair_key has never accepted quoted identifiers in key_att.
  IF key_att = '' THEN
    block_key_cols := ARRAY[]::INT2[];
  ELSE
    SELECT array_agg(a.attnum ORDER BY t.ord)::INT2[]
      INTO block_key_cols
      FROM unnest(string_to[](key_att, ',')) WITH ORDINALITY AS t(name, ord)
      JOIN pg_attribute a
        ON a.attrelid = _tbl
       AND a.attname  = trim(t.name)
       AND a.attnum   > 0
       AND NOT a.attisdropped;
    IF block_key_cols IS NULL OR array_length(block_key_cols, 1) IS NULL THEN
      RAISE EXCEPTION 'repair_key: could not resolve key columns from "%"', key_att;
    END IF;
    IF array_length(block_key_cols, 1) > 16 THEN
      RAISE EXCEPTION 'repair_key: block key wider than 16 columns is not supported';
    END IF;
  END IF;
 
  -- Same column shape as add_provenance: no UNIQUE, no DEFAULT past
  -- the initial backfill (the guard trigger added after the rename
  -- takes over both jobs once the column has been renamed to its
  -- final name).  The DEFAULT is kept here only so the second pass
  -- below can read provsql_temp from the user-visible rows
  -- without a separate UPDATE.
  EXECUTE format('ALTER TABLE %s ADD COLUMN provsql_temp UUID DEFAULT public.uuid_generate_v4()', _tbl);
 
  -- Build a per-group mapping (key columns + a fresh key_token + the
  -- group size) once, then use it for both the create_gate(key_token,
  -- 'input') first pass and the per-row mulinput second pass.  Going
  -- through a temp table avoids re-running uuid_generate_v4() (which
  -- would produce different UUIDs the second time).  USING (%1$s) on
  -- the second pass handles the multi-column case uniformly.
  -- ON COMMIT DROP plus the explicit DROP TABLE at the end of this
  -- function leave the temp table cleaned up across transactions and
  -- across repeated calls in the same transaction.
  IF key_att = '' THEN
    EXECUTE format(
      'CREATE TEMP TABLE provsql_repair_key_tmp ON COMMIT DROP AS
         SELECT public.uuid_generate_v4() AS provsql_key_token,
                COUNT(*) AS provsql_group_size
           FROM %s', _tbl);
    rows_query := format(
      'SELECT t.provsql_temp,
              k.provsql_key_token AS key_token,
              ROW_NUMBER() OVER (ORDER BY t.ctid) AS within_group,
              k.provsql_group_size AS group_size
         FROM %s t CROSS JOIN provsql_repair_key_tmp k', _tbl);
  ELSE
    EXECUTE format(
      'CREATE TEMP TABLE provsql_repair_key_tmp ON COMMIT DROP AS
         SELECT %1$s,
                public.uuid_generate_v4() AS provsql_key_token,
                COUNT(*) AS provsql_group_size
           FROM %2$s
       GROUP BY %1$s', key_att, _tbl);
    rows_query := format(
      'SELECT t.provsql_temp,
              k.provsql_key_token AS key_token,
              ROW_NUMBER() OVER (PARTITION BY k.provsql_key_token
                                 ORDER BY t.ctid) AS within_group,
              k.provsql_group_size AS group_size
         FROM %2$s t
         JOIN provsql_repair_key_tmp k USING (%1$s)', key_att, _tbl);
  END IF;
 
  -- Pass 1: one input gate per group key.
  FOR r IN SELECT provsql_key_token FROM provsql_repair_key_tmp LOOP
    PERFORM provsql.create_gate(r.provsql_key_token, 'input');
  END LOOP;
 
  -- Pass 2: per row, attach a mulinput gate to its group's key token.
  FOR r IN EXECUTE rows_query LOOP
    PERFORM provsql.create_gate(r.provsql_temp, 'mulinput', ARRAY[r.key_token]);
    PERFORM provsql.set_prob(r.provsql_temp, 1./r.group_size);
    PERFORM provsql.set_infos(r.provsql_temp, r.within_group::INT);
  END LOOP;
 
  DROP TABLE provsql_repair_key_tmp;
 
  EXECUTE format('ALTER TABLE %s ALTER COLUMN provsql_temp DROP DEFAULT', _tbl);
  EXECUTE format('ALTER TABLE %s RENAME COLUMN provsql_temp TO provsql', _tbl);
  EXECUTE format('CREATE INDEX ON %s(provsql)', _tbl);
  EXECUTE format(
    'CREATE TRIGGER provenance_guard BEFORE INSERT OR UPDATE OF provsql '
    'ON %s FOR EACH ROW EXECUTE PROCEDURE provsql.provenance_guard()',
    _tbl);
  PERFORM provsql.set_table_info(_tbl::oid, 'bid', block_key_cols);
  -- Base BID tables also have themselves as their sole ancestor.  Same
  -- rationale as the @c add_provenance branch above.
  PERFORM provsql.set_ancestors(_tbl::oid, ARRAY[_tbl::oid]);
END
$$ LANGUAGE plpgsql;
 
/**
 * @brief Event trigger that purges per-table provenance metadata when
 *        a tracked relation is dropped outside of remove_provenance().
 *
 * Plain DROP TABLE bypasses remove_provenance() and would otherwise
 * leave a stale entry in the table-info store keyed by a now-recycled
 * OID, with confusing consequences for the safe-query rewriter the
 * next time the OID is reused.  This trigger forwards every dropped
 * relation OID to provsql.remove_table_info(), which is a no-op for
 * relations that were not tracked.
 */
CREATE OR REPLACE FUNCTION cleanup_table_info()
  RETURNS event_trigger AS
$$
DECLARE
  r RECORD;
BEGIN
  FOR r IN
    SELECT objid FROM pg_event_trigger_dropped_objects()
     WHERE object_type IN ('table', 'foreign table', 'materialized view')
  LOOP
    PERFORM provsql.remove_table_info(r.objid);
    -- Forget any maintained mapping whose source or mapping table is gone.
    DELETE FROM provsql.provenance_mapping_registry
     WHERE source = r.objid OR mapping = r.objid;
  END LOOP;
END
$$ LANGUAGE plpgsql;
 
DROP EVENT TRIGGER IF EXISTS provsql_cleanup_table_info;
-- @c EXECUTE @c PROCEDURE (rather than the PG 11+ @c EXECUTE
-- @c FUNCTION alias) so the extension installs on PG 10 too.
CREATE EVENT TRIGGER provsql_cleanup_table_info ON sql_drop
  EXECUTE PROCEDURE provsql.cleanup_table_info();
 
/**
 * @brief Registry of maintained provenance mappings
 *
 * Each row records that mapping table @c mapping is kept current for the
 * @c attribute column of the provenance-tracked @c source table: every
 * genuine insert into @c source appends @c (value, provenance) to it (see
 * @c provenance_guard).  Keyed on the mapping table, indexed on the source
 * so the guard can look up a table's mappings cheaply.  Entries are removed
 * when either table is dropped (see @c cleanup_table_info).
 */
CREATE TABLE IF NOT EXISTS provsql.provenance_mapping_registry(
  mapping   oid PRIMARY KEY,
  source    oid NOT NULL,
  attribute name NOT NULL
);
CREATE INDEX IF NOT EXISTS provenance_mapping_registry_source_idx
  ON provsql.provenance_mapping_registry(source);
 
/**
 * @brief Create a provenance mapping table from an attribute
 *
 * Creates a new table mapping provenance tokens to values of a given
 * attribute, for use with semiring evaluation functions.
 * Idempotent: if the mapping table already exists, raises a NOTICE and
 * changes nothing (drop it first to rebuild).
 *
 * @param newtbl name of the mapping table to create
 * @param oldtbl source table with provenance tracking
 * @param att attribute whose values populate the mapping
 * @param preserve_case if true, quote the table name to preserve case
 * @param maintained if true, register the mapping so later inserts into
 *        @c oldtbl keep it current, and it stays correct after data
 *        modification (deletes/updates rewrite a row's provsql, but the
 *        validity stays keyed to the original input token).  @c att must
 *        then be a plain column name.  When false (the default) the table is
 *        a one-off snapshot.
 */
CREATE OR REPLACE FUNCTION create_provenance_mapping(
  newtbl TEXT,
  oldtbl REGCLASS,
  att TEXT,
  preserve_case BOOL DEFAULT 'f',
  maintained BOOL DEFAULT false
) RETURNS VOID AS
$$
DECLARE
BEGIN
  -- Idempotence: when the mapping table already exists, leave it alone
  -- with a NOTICE (re-runnable setup scripts / notebook cells). Drop it
  -- first to rebuild a stale mapping.
  IF (CASE WHEN preserve_case THEN to_regclass(format('%I', newtbl))
           ELSE to_regclass(newtbl) END) IS NOT NULL THEN
    RAISE NOTICE 'mapping table % already exists', newtbl;
    RETURN;
  END IF;
  -- ON COMMIT DROP only fires at COMMIT: several mapping creations in
  -- one transaction (a notebook cell, a setup script run via psql -1)
  -- would otherwise collide on the leftover temp table. The to_regclass
  -- probe (rather than DROP IF EXISTS) keeps the first call NOTICE-free.
  IF to_regclass('pg_temp.tmp_provsql') IS NOT NULL THEN
    DROP TABLE tmp_provsql;
  END IF;
  EXECUTE format('CREATE TEMP TABLE tmp_provsql ON COMMIT DROP AS TABLE %s', oldtbl);
  ALTER TABLE tmp_provsql RENAME provsql TO provenance;
  -- The mapping is keyed by gate identity (input-token UUIDs), so peel any
  -- transparent annotation wrapper (e.g. the inversion-free certificate a
  -- certified query attaches to its row roots) off the captured tokens.
  UPDATE tmp_provsql SET provenance = provsql.strip_annotations(provenance)
    WHERE provsql.get_gate_type(provenance) = 'annotation';
  IF preserve_case THEN
    EXECUTE format('CREATE TABLE %I AS SELECT %s AS value, provenance FROM tmp_provsql', newtbl, att);
    EXECUTE format('CREATE INDEX ON %I(provenance)', newtbl);
  ELSE
    EXECUTE format('CREATE TABLE %s AS SELECT %s AS value, provenance FROM tmp_provsql', newtbl, att);
    EXECUTE format('CREATE INDEX ON %s(provenance)', newtbl);
  END IF;
  IF maintained THEN
    -- Register so genuine inserts into oldtbl keep the mapping current
    -- (see provenance_guard); keyed to the input token, so it survives the
    -- provsql rewrites that data modification performs.
    INSERT INTO provsql.provenance_mapping_registry(mapping, source, attribute)
      VALUES (
        (CASE WHEN preserve_case THEN to_regclass(format('%I', newtbl))
              ELSE to_regclass(newtbl) END)::oid,
        oldtbl::oid, att)
      ON CONFLICT (mapping)
        DO UPDATE SET source = EXCLUDED.source, attribute = EXCLUDED.attribute;
  END IF;
END
$$ LANGUAGE plpgsql;
 
/** @} */
 
/** @defgroup internal_constants Internal constants
 *  UUID namespace and identity element functions used for
 *  deterministic gate generation.
 *  @{
 */
 
/** @brief Return the ProvSQL UUID namespace (used for deterministic gate UUIDs) */
CREATE OR REPLACE FUNCTION uuid_ns_provsql() RETURNS UUID AS
$$
 -- uuid_generate_v5(uuid_ns_url(),'http://pierre.senellart.com/software/provsql/')
 SELECT '920d4f02-8718-5319-9532-d4ab83a64489'::UUID
$$ LANGUAGE SQL IMMUTABLE PARALLEL SAFE;
 
/** @brief Return the UUID of the semiring zero gate */
CREATE OR REPLACE FUNCTION gate_zero() RETURNS UUID AS
$$
  SELECT public.uuid_generate_v5(provsql.uuid_ns_provsql(),'zero');
$$ LANGUAGE SQL IMMUTABLE PARALLEL SAFE;
 
/** @brief Return the UUID of the semiring one gate */
CREATE OR REPLACE FUNCTION gate_one() RETURNS UUID AS
$$
  SELECT public.uuid_generate_v5(provsql.uuid_ns_provsql(),'one');
$$ LANGUAGE SQL IMMUTABLE PARALLEL SAFE;
 
/** @brief Return the epsilon threshold used for probability comparisons */
CREATE OR REPLACE FUNCTION epsilon() RETURNS DOUBLE PRECISION AS
$$
  SELECT CAST(0.001 AS DOUBLE PRECISION)
$$ LANGUAGE SQL IMMUTABLE PARALLEL SAFE;
 
/** @} */
 
/** @defgroup semiring_operations Semiring operations
 *  Functions that build provenance circuit gates for semiring operations.
 *  These are called internally by the query rewriter.
 *
 *  They are declared @c IMMUTABLE: each derives its gate UUID
 *  deterministically from its arguments (a @c uuid5 content address) and
 *  the @c create_gate write at that address is idempotent, so the token a
 *  call returns is a pure function of its inputs.  The marking matters for
 *  parallelism: PL/pgSQL runs a non-volatile function's inner SPI
 *  read-only, so the per-row builders the rewriter injects into a scan do
 *  not call @c CommandCounterIncrement -- which would raise "cannot start
 *  commands during a parallel operation" once the enclosing statement has
 *  gone parallel.  A @c VOLATILE builder both blocks that parallel plan and
 *  loses the query-wide speed-up.
 *  @{
 */
 
/**
 * @brief Create a times (product) gate from multiple provenance tokens
 *
 * Filters out NULL and one-gates; returns gate_one() if all tokens
 * are trivial, or a single token if only one remains.
 *
 * Before creating an ordinary gate, the *times-canonical* address of
 * the surviving multiset -- @c uuid5('times-canonical{sorted tokens}')
 * -- is probed: the reachability rewriter pre-creates there, for
 * self-join conjunctions of reachability tokens, a certified
 * equivalent (the all-members-reachable circuit; see
 * @c plant_reach_cover).  Ordinary creation never writes under that
 * recipe, so a hit is always a deliberate plant; the ordinary
 * order-dependent recipe is used otherwise, so ordinary
 * times gates (and their formula rendering) are untouched.
 */
CREATE OR REPLACE FUNCTION provenance_times(VARIADIC tokens UUID[])
  RETURNS UUID AS
$$
DECLARE
  times_token UUID;
  filtered_tokens UUID[];
  canonical UUID;
BEGIN
  -- A NULL element reads as the ⊗-neutral 1: it is the token slot of an
  -- untracked source (a join against an untracked table), which is
  -- certain.  Contrast provenance_plus / provenance_monus, where NULL
  -- reads as the ⊕- / ⊖-right-neutral 0: each combinator maps NULL to
  -- its own neutral element.  Nothing may therefore hand a NULL to ⊗
  -- meaning "false"; a comparison with a NULL operand goes through
  -- provenance_cmp, which returns gate_zero for it.
  SELECT array_agg(t) FROM unnest(tokens) t WHERE t IS NOT NULL AND t <> gate_one() INTO filtered_tokens;
 
  -- Dispatch on the FILTERED count: a single survivor short-circuits
  -- to that token directly (no useless single-child times gate); zero
  -- survivors collapse to the identity. Using array_length(tokens, 1)
  -- here would miss the [one, cmp] → [cmp] case, leaving the cmp wrapped
  -- in a one-child times when its only sibling was gate_one().
  CASE coalesce(array_length(filtered_tokens, 1), 0)
    WHEN 0 THEN
      times_token:=gate_one();
    WHEN 1 THEN
      times_token:=filtered_tokens[1];
    ELSE
      -- Computed separately from the filtering aggregate above: an
      -- ORDER BY aggregate there would make the planner feed *both*
      -- aggregates sorted input, scrambling the stored children order.
      SELECT uuid_generate_v5(uuid_ns_provsql(),
                              concat('times-canonical', array_agg(t ORDER BY t)))
      FROM unnest(filtered_tokens) t
      INTO canonical;
      IF get_gate_type(canonical) = 'times' THEN
        -- A deliberate pre-creation at the canonical address: same
        -- children, same product.
        times_token := canonical;
      ELSE
        times_token := uuid_generate_v5(uuid_ns_provsql(),concat('times',filtered_tokens));
 
        PERFORM create_gate(times_token, 'times', ARRAY_AGG(t)) FROM UNNEST(filtered_tokens) AS t WHERE t IS NOT NULL;
      END IF;
  END CASE;
 
  RETURN times_token;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Create a monus (difference) gate from two provenance tokens
 *
 * Implements m-semiring monus. Returns token1 if token2 is NULL
 * (used for LEFT OUTER JOIN semantics in the EXCEPT rewriting).
 */
CREATE OR REPLACE FUNCTION provenance_monus(token1 UUID, token2 UUID)
  RETURNS UUID AS
$$
DECLARE
  monus_token UUID;
BEGIN
  IF token1 IS NULL THEN
    RAISE EXCEPTION USING MESSAGE='provenance_monus is called with first argument NULL';
  END IF;
 
  IF token2 IS NULL THEN
    -- The ⊖-right-neutral 0: a NULL second argument is the no-match case
    -- of the difference operator's LEFT OUTER JOIN (nothing to subtract),
    -- so X ⊖ NULL = X ⊖ 0 = X.  Note this is NOT the NULL ≡ 1 reading of
    -- provenance_times; each combinator maps NULL to its own neutral.
    RETURN token1;
  END IF;
 
  IF token1 = token2 THEN
    -- X-X=0
    monus_token:=gate_zero();
  ELSIF token1 = gate_zero() THEN
    -- 0-X=0
    monus_token:=gate_zero();
  ELSIF token2 = gate_zero() THEN
    -- X-0=X
    monus_token:=token1;
  ELSE
    monus_token:=uuid_generate_v5(uuid_ns_provsql(),concat('monus',token1,token2));
    PERFORM create_gate(monus_token, 'monus', ARRAY[token1::UUID, token2::UUID]);
  END IF;
 
  RETURN monus_token;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Create a project gate for where-provenance tracking
 *
 * Records the mapping between input and output attribute positions.
 *
 * @param token child provenance token
 * @param positions array encoding attribute position mappings
 */
CREATE OR REPLACE FUNCTION provenance_project(token UUID, VARIADIC positions INT[])
  RETURNS UUID AS
$$
DECLARE
  project_token UUID;
  rec RECORD;
BEGIN
  project_token:=uuid_generate_v5(uuid_ns_provsql(),concat('project', token, positions));
  PERFORM create_gate(project_token, 'project', ARRAY[token]);
  PERFORM set_extra(project_token, ARRAY_AGG(pair)::TEXT)
  FROM (
    SELECT ARRAY[(CASE WHEN info=0 THEN NULL ELSE info END), idx] AS pair
    FROM unnest(positions) WITH ORDINALITY AS a(info, idx)
    ORDER BY idx
  ) t;
 
  RETURN project_token;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Create an equijoin gate for where-provenance tracking
 *
 * @param token child provenance token
 * @param pos1 attribute index in the first relation
 * @param pos2 attribute index in the second relation
 */
CREATE OR REPLACE FUNCTION provenance_eq(token UUID, pos1 INT, pos2 INT)
  RETURNS UUID AS
$$
DECLARE
  eq_token UUID;
  rec RECORD;
BEGIN
  eq_token:=uuid_generate_v5(uuid_ns_provsql(),concat('eq',token,pos1,',',pos2));
 
  PERFORM create_gate(eq_token, 'eq', ARRAY[token::UUID]);
  PERFORM set_infos(eq_token, pos1, pos2);
  RETURN eq_token;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Create a plus (sum) gate from an array of provenance tokens
 *
 * Filters out NULL and zero-gates; returns gate_zero() if all tokens
 * are trivial, or a single token if only one remains.  Before creating
 * a gate, probes the *canonical* address of the multiset -- a dedicated
 * v5 recipe namespace over the sorted tokens (plus is commutative), in
 * which this function never creates anything, so a gate found there is
 * always a deliberate pre-creation computing the same sum.  That is the
 * bounded-hop reachability route's hook: it plants, at the canonical
 * address of a vertex's per-length tokens, a certified gate over its
 * native within-bound circuit, keeping the natural hop-discarding query
 * on the linear evaluation route.  Absent a canonical gate, the
 * ordinary order-dependent recipe is used, so ordinary plus
 * gates (and their formula rendering) are untouched.
 */
CREATE OR REPLACE FUNCTION provenance_plus(tokens UUID[])
  RETURNS UUID AS
$$
DECLARE
  c INTEGER;
  plus_token UUID;
  filtered_tokens UUID[];
  canonical UUID;
BEGIN
  -- A NULL element reads as the ⊕-neutral 0: it stands for a row absent
  -- from the disjunction (a null-padded antijoin row whose token array
  -- slot is NULL), not for an untracked source.  Contrast provenance_times,
  -- where NULL reads as the ⊗-neutral 1 (untracked source): each
  -- combinator maps NULL to its own neutral element.
  SELECT array_agg(t) FROM unnest(tokens) t
  WHERE t IS NOT NULL AND t <> gate_zero()
  INTO filtered_tokens;
 
  c:=array_length(filtered_tokens, 1);
 
  IF c = 0 THEN
    plus_token := gate_zero();
  ELSIF c = 1 THEN
    plus_token := filtered_tokens[1];
  ELSE
    -- Computed separately from the filtering aggregate above: an ORDER
    -- BY aggregate there would make the planner feed *both* aggregates
    -- sorted input, scrambling the stored (aggregation-order) children.
    SELECT uuid_generate_v5(uuid_ns_provsql(),
                            concat('plus-canonical', array_agg(t ORDER BY t)))
    FROM unnest(filtered_tokens) t
    INTO canonical;
    IF get_gate_type(canonical) = 'plus' THEN
      -- A deliberate pre-creation at the canonical address: same
      -- children, same sum.
      plus_token := canonical;
    ELSE
      plus_token := uuid_generate_v5(
        uuid_ns_provsql(),
        concat('plus', filtered_tokens));
 
      PERFORM create_gate(plus_token, 'plus', filtered_tokens);
    END IF;
  END IF;
 
  RETURN plus_token;
END
$$ LANGUAGE plpgsql STRICT SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Driver for provenance over recursive queries (WITH RECURSIVE).
 *
 * Invoked by the planner hook (@c lower_recursive_cte in @c provsql.c) when it
 * lowers a recursive CTE whose body touches provenance-tracked relations.  The
 * hook deparses the CTE body to SQL and calls this function, which runs naive
 * bottom-up (fixpoint) evaluation: each round re-evaluates the body
 * @c base @c UNION @c recursive over a tracked working table until the
 * provenance tokens stop changing.  Every round goes through ProvSQL's normal
 * rewriting, so the recursive join yields @c times gates, the untracked base
 * branch yields @c gate_one, and the @c UNION yields the @c plus merge of
 * alternative derivations -- no provenance is plumbed by hand here.  The result
 * is left in a tracked temp table named @p work_name, which the hook then scans
 * in place of the CTE.
 *
 * The working tables (@p work_name and a scratch @c _new) are created once and
 * reused across rounds (TRUNCATE + INSERT), so the round count never
 * accumulates relation locks.  Because content-addressed gate UUIDs make
 * structurally identical sub-circuits share, the fixpoint test is an exact
 * relational @c EXCEPT and the circuit stays the shared (polynomial) form.
 *
 * Scope: UNION (set) recursion.  On *acyclic* input the structural fixpoint is
 * reached and the resulting circuit is the universal provenance, sound for any
 * semiring.  On *cyclic* input the circuit never stabilises structurally; when
 * the session's provenance class (@c provsql.provenance) is @c 'absorptive' or
 * @c 'BOOLEAN' we instead stop at the value-fixpoint bound (number of
 * derivable tuples) -- every minimal, tuple-repetition-free derivation is then
 * covered, and the longer ones are absorbed in any absorptive semiring (after
 * Deutch, Milo, Roy & Tannen, ICDT 2014) -- and wrap the resulting tokens in
 * the @c 'absorptive' assumption marker, so that non-absorptive semiring
 * evaluations (counting, why-provenance: genuinely infinite on cyclic data)
 * refuse them while probability, Boolean, formula-as-circuit and min-plus
 * evaluations proceed.  Under the general classes, cyclic input trips the
 * @p max_iter guard.
 *
 * This function has no @c SET @c search_path on purpose: @p body_sql is the
 * caller's deparsed query and must resolve relation names in the caller's path.
 *
 * @param body_sql   the recursive CTE body, e.g.
 *                   @c 'SELECT 1 UNION SELECT e.dst FROM edge e JOIN reach r ON e.src=r.node'
 * @param work_name  the working relation name @p body_sql references (the CTE name)
 * @param colnames   comma-separated user columns, e.g. @c 'node'
 * @param coldef     column definitions for the working table, e.g. @c 'node INTEGER'
 * @param max_iter   safety bound on fixpoint rounds (non-termination guard)
 */
CREATE OR REPLACE FUNCTION eval_recursive(
  body_sql  TEXT,
  work_name TEXT,
  colnames  TEXT,
  coldef    TEXT,
  max_iter  INT DEFAULT 1000)
  RETURNS VOID AS
$$
DECLARE
  changed   BOOLEAN;        -- circuit changed structurally this round
  set_stable BOOLEAN;       -- user-column tuple set unchanged this round
  iters     INT := 0;
  new_count INT;            -- rows in _new this round (INSERT ROW_COUNT)
  -- Under an absorptive semiring the provenance *value* converges on cyclic
  -- data even though the circuit keeps growing structurally.  A minimal
  -- derivation cannot repeat a tuple, so it has depth <= (number of derivable
  -- tuples); after that many naive rounds the value equals the least fixpoint,
  -- and the surplus (longer, cyclic) derivations are absorbed at evaluation
  -- time.  We learn that bound from the tuple-set fixpoint, stop there, and
  -- mark the resulting tokens with the 'absorptive' assumption so evaluation
  -- under a non-absorptive semiring refuses rather than silently returning a
  -- truncated value.
  absorptive_mode BOOLEAN :=
    coalesce(current_setting('provsql.provenance', true), 'semiring')
      IN ('absorptive', 'BOOLEAN');
  truncated BOOLEAN := false; -- exited at the value fixpoint (cyclic data)
  ntuples   INT := NULL;    -- the bound above, set once the tuple set stabilises
BEGIN
  EXECUTE format('DROP TABLE IF EXISTS %I', work_name);
  DROP TABLE IF EXISTS _new;
 
  -- Tracked working table (carries provsql), initially empty, plus a scratch
  -- table of the same shape; both reused across rounds.
  EXECUTE format('CREATE TEMP TABLE %I (%s, provsql UUID)', work_name, coldef);
  EXECUTE format('CREATE TEMP TABLE _new (LIKE %I)', work_name);
 
  LOOP
    iters := iters + 1;
    -- Hard safety bound (also catches genuinely unbounded recursion, e.g. an
    -- unbounded counter, where even the tuple set never stabilises).
    IF iters > max_iter THEN
      RAISE EXCEPTION 'eval_recursive: no fixpoint after % rounds (cyclic data?)', max_iter;
    END IF;
 
    -- One round of naive evaluation: re-run the CTE body over the current
    -- working table.  INSERT targets a tracked table, so ProvSQL fills provsql.
    -- Take the row count from the INSERT itself (counting _new directly would be
    -- an aggregate over a provenance-tracked table -> an AGG_TOKEN).
    EXECUTE 'TRUNCATE _new';
    EXECUTE format('INSERT INTO _new(%s) %s', colnames, body_sql);
    GET DIAGNOSTICS new_count = ROW_COUNT;
 
    -- Exact structural fixpoint test (content-addressed tokens => set equality).
    EXECUTE format(
      'SELECT EXISTS((TABLE _new EXCEPT TABLE %1$I) UNION ALL (TABLE %1$I EXCEPT TABLE _new))',
      work_name) INTO changed;
 
    -- In an absorptive class, learn the round bound from the tuple-set
    -- fixpoint (the set always stabilises after finitely many rounds, even on
    -- cyclic data).
    IF absorptive_mode AND ntuples IS NULL THEN
      EXECUTE format(
        'SELECT NOT EXISTS('
        || '(SELECT %2$s FROM _new EXCEPT SELECT %2$s FROM %1$I) UNION ALL '
        || '(SELECT %2$s FROM %1$I EXCEPT SELECT %2$s FROM _new))',
        work_name, colnames) INTO set_stable;
      IF set_stable THEN
        ntuples := new_count;
      END IF;
    END IF;
 
    -- Copy _new into the working table (tracked -> tracked carries the tokens).
    EXECUTE format('TRUNCATE %I', work_name);
    EXECUTE format('INSERT INTO %1$I(%2$s) SELECT %2$s FROM _new', work_name, colnames);
 
    -- Structural fixpoint: done (acyclic / fully converged) -- sound for any
    -- semiring.
    EXIT WHEN NOT changed;
 
    -- Absorptive class on cyclic data: once the value-fixpoint bound is
    -- reached (plus one confirming round, so that acyclic circuits whose
    -- token depth lags the tuple-set saturation still exit through the
    -- structural test above, untagged) we stop, even though the circuit
    -- is not structurally stable.
    IF absorptive_mode AND ntuples IS NOT NULL AND iters >= ntuples + 1 THEN
      truncated := true;
      EXIT;
    END IF;
  END LOOP;
 
  -- Tokens of a truncated (cyclic) fixpoint are sound only under absorptive
  -- evaluation: RECORD that in the circuit itself.
  IF truncated THEN
    EXECUTE format(
      'UPDATE %I SET provsql = provsql.provenance_assume(provsql, ''absorptive'')',
      work_name);
  END IF;
END
$$ LANGUAGE plpgsql SET client_min_messages = warning;
 
/**
 * @brief Create a comparison gate for HAVING clause provenance
 *
 * @param left_token provenance token for the left operand
 * @param comparison_op OID of the comparison operator
 * @param right_token provenance token for the right operand
 */
CREATE OR REPLACE FUNCTION provenance_cmp(
  left_token  UUID,
  comparison_op OID,
  right_token UUID
)
RETURNS UUID AS
$$
DECLARE
  cmp_token UUID;
BEGIN
  -- A comparison with a NULL operand (a NULL random_variable cell, or an
  -- aggregate that is NULL on the instance) is unknown under SQL's 3VL in
  -- every possible world: the row is annotated zero.  The function must
  -- not be STRICT: a NULL result would read as the neutral token
  -- (provenance_times drops it), silently turning "unknown" into
  -- "certainly true".
  IF left_token IS NULL OR right_token IS NULL OR comparison_op IS NULL THEN
    RETURN gate_zero();
  END IF;
  -- deterministic v5 namespace id
  cmp_token := public.uuid_generate_v5(
    uuid_ns_provsql(),
    concat('cmp', left_token::TEXT, comparison_op::TEXT, right_token::TEXT)
  );
  -- wire it up in the circuit
  PERFORM create_gate(cmp_token, 'cmp', ARRAY[left_token, right_token]);
  PERFORM set_infos(cmp_token, comparison_op::INTEGER);
  RETURN cmp_token;
END
$$ LANGUAGE plpgsql
  SET search_path=provsql,pg_temp,public
  SECURITY DEFINER
  IMMUTABLE
  PARALLEL SAFE;
 
/**
 * @brief The factors of a row annotation an aggregate comparison does not
 *        subsume.
 *
 * A lifted comparison entails the existence of the group it ranges over, so it
 * supersedes that group's @c gate_delta instead of multiplying with it.  This
 * reports which factors of @p tokens survive that supersede: a bare δ over the
 * compared group disappears, a @c times keeps its other factors, and anything
 * else -- an earlier comparison on the same group, an input -- is kept whole.
 */
CREATE FUNCTION cmp_surviving_factors(tokens UUID[], cmp UUID)
  RETURNS UUID[] AS
  'provsql', 'cmp_surviving_factors' LANGUAGE C PARALLEL SAFE STABLE;
 
/**
 * @brief Combine a lifted aggregate comparison with the row annotation it
 *        supersedes only part of.
 *
 * @param cmp     Gate of the lifted comparison.
 * @param tokens  Row-annotation factors at the level owning the comparison.
 * @return  @c cmp multiplied with whatever of @p tokens it does not subsume.
 */
CREATE OR REPLACE FUNCTION provenance_cmp_times(cmp UUID, tokens UUID[])
  RETURNS UUID AS
$$
DECLARE
  kept UUID[];
BEGIN
  kept := provsql.cmp_surviving_factors(tokens, cmp);
  IF kept IS NULL OR array_length(kept, 1) IS NULL THEN
    RETURN cmp;
  END IF;
  RETURN provsql.provenance_times(VARIADIC kept || cmp);
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Create an arithmetic gate over scalar-valued provenance children
 *
 * Builds a deterministic @c gate_arith from an operator tag and an
 * ordered list of children.  The tag is one of the @c provsql_arith_op
 * ENUM values declared in @c src/provsql_utils.h
 * (@c PLUS=0, @c TIMES=1, @c MINUS=2, @c DIV=3, @c NEG=4) and is
 * stored in the gate's @c info1 field.  Children must be UUIDs of
 * scalar-producing gates (@c gate_rv, @c gate_value, or another
 * @c gate_arith).  The token UUID is derived deterministically from
 * @p op and @p children so identical sub-expressions share their gate.
 *
 * @param op        Operator tag (@c provsql_arith_op).
 * @param children  Ordered list of child gate UUIDs.
 * @return  UUID of the (possibly pre-existing) @c gate_arith.
 */
CREATE OR REPLACE FUNCTION provenance_arith(
  op       INTEGER,
  children UUID[]
)
RETURNS UUID AS
$$
DECLARE
  arith_token UUID;
BEGIN
  arith_token := public.uuid_generate_v5(
    uuid_ns_provsql(),
    concat('arith', op::TEXT, children::TEXT)
  );
  PERFORM create_gate(arith_token, 'arith', children);
  PERFORM set_infos(arith_token, op);
  RETURN arith_token;
END
$$ LANGUAGE plpgsql
  SET search_path=provsql,pg_temp,public
  SECURITY DEFINER
  IMMUTABLE
  PARALLEL SAFE
  STRICT;
 
/**
 * @brief Create a guarded-selection gate over scalar (RV) children.
 *
 * Builds a deterministic @c gate_case from the flattened wire list
 * @c [guard_1, value_1, ..., guard_k, value_k, default] (odd length): the
 * value of the first guard event that holds, else the default (first-match
 * semantics).  Each guard is a Boolean event token (a @c gate_cmp or Boolean
 * combination); each value and the default are scalar-producing gates
 * (@c gate_rv, @c gate_value, @c gate_arith, another @c gate_case, ...).  The
 * token UUID is derived deterministically from @p children so identical
 * @c CASE expressions share their gate.
 *
 * @param children  Flattened guard/value wires ending with the default
 *                  (@c array_length must be odd and @c >= 1).
 * @return  UUID of the (possibly pre-existing) @c gate_case.
 */
CREATE OR REPLACE FUNCTION provenance_case(
  children UUID[]
)
RETURNS UUID AS
$$
DECLARE
  case_token UUID;
BEGIN
  IF array_length(children, 1) IS NULL OR array_length(children, 1) % 2 = 0 THEN
    RAISE EXCEPTION 'provenance_case expects an odd number of children '
      '(guard/value pairs followed by a default), got %',
      coalesce(array_length(children, 1), 0);
  END IF;
  case_token := public.uuid_generate_v5(
    uuid_ns_provsql(),
    concat('case', children::TEXT)
  );
  PERFORM create_gate(case_token, 'case', children);
  RETURN case_token;
END
$$ LANGUAGE plpgsql
  SET search_path=provsql,pg_temp,public
  SECURITY DEFINER
  IMMUTABLE
  PARALLEL SAFE
  STRICT;
 
/** @} */
 
/** @defgroup semiring_evaluation Semiring evaluation
 *  Functions for evaluating provenance circuits over semirings,
 *  both user-defined (via function references) and compiled (built-in).
 *  @{
 */
 
/**
 * @brief Evaluate provenance using a compiled (built-in) semiring
 *
 * This C function handles semiring evaluation entirely in C++ for
 * better performance. The semiring is specified by name.
 *
 * @param token provenance token to evaluate
 * @param token2value mapping table from tokens to semiring values
 * @param semiring name of the compiled semiring (e.g., "formula", "counting")
 * @param element_one identity element of the semiring
 */
CREATE OR REPLACE FUNCTION provenance_evaluate_compiled(
  token UUID,
  token2value REGCLASS,
  semiring TEXT,
  element_one ANYELEMENT)
RETURNS ANYELEMENT AS
  'provsql', 'provenance_evaluate_compiled' LANGUAGE C PARALLEL SAFE STABLE;
 
 
/**
 * @brief Evaluate provenance over a user-defined semiring (PL/pgSQL version)
 *
 * Recursively walks the provenance circuit and evaluates each gate
 * using the provided semiring operations. This is the generic version
 * that accepts semiring operations as function references.
 *
 * @param token provenance token to evaluate
 * @param token2value mapping table from tokens to semiring values
 * @param element_one identity element of the semiring
 * @param value_type OID of the semiring value type
 * @param plus_function semiring addition (aggregate)
 * @param times_function semiring multiplication (aggregate)
 * @param monus_function semiring monus (binary), or NULL
 * @param delta_function δ-semiring operator, or NULL
 */
CREATE OR REPLACE FUNCTION provenance_evaluate(
  token UUID,
  token2value REGCLASS,
  element_one ANYELEMENT,
  value_type REGTYPE,
  plus_function REGPROC,
  times_function REGPROC,
  monus_function REGPROC,
  delta_function REGPROC)
  RETURNS ANYELEMENT AS
$$
DECLARE
  gate_type PROVENANCE_GATE;
  result ALIAS FOR $0;
  children UUID[];
--  cmp_value ANYELEMENT;
--  temp_result ANYELEMENT;
  value_text TEXT;
BEGIN
  SELECT get_gate_type(token) INTO gate_type;
 
  IF gate_type IS NULL THEN
    RETURN NULL;
 
  ELSIF gate_type = 'input' THEN
    EXECUTE format('SELECT value FROM %s WHERE provenance=%L', token2value, token)
      INTO result;
    IF result IS NULL THEN
      result := element_one;
    END IF;
  ELSIF gate_type = 'mulinput' THEN
    SELECT concat('{',(get_children(token))[1]::TEXT,'=',(get_infos(token)).info1,'}')
      INTO result;
  ELSIF gate_type='update' THEN
    EXECUTE format('SELECT value FROM %s WHERE provenance=%L',token2value,token) INTO result;
    IF result IS NULL THEN
      result:=element_one;
    END IF;
  ELSIF gate_type = 'plus' THEN
    EXECUTE format('SELECT %s(provsql.provenance_evaluate(t,%L,%L::%s,%L,%L,%L,%L,%L)) FROM unnest(get_children(%L)) AS t',
      plus_function, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function, token)
      INTO result;
 
  ELSIF gate_type = 'times' THEN
    EXECUTE format('SELECT %s(provsql.provenance_evaluate(t,%L,%L::%s,%L,%L,%L,%L,%L)) FROM unnest(get_children(%L)) AS t',
      times_function, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function, token)
      INTO result;
 
  ELSIF gate_type = 'monus' THEN
    IF monus_function IS NULL THEN
      RAISE EXCEPTION USING MESSAGE='Provenance with negation evaluated over a semiring without monus function';
    ELSE
      EXECUTE format('SELECT %s(a1,a2) FROM (SELECT provsql.provenance_evaluate(c[1],%L,%L::%s,%L,%L,%L,%L,%L) AS a1, ' ||
                     'provsql.provenance_evaluate(c[2],%L,%L::%s,%L,%L,%L,%L,%L) AS a2 FROM get_children(%L) c) tmp',
        monus_function, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function,
        token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function, token)
      INTO result;
    END IF;
 
  ELSIF gate_type = 'eq' THEN
    EXECUTE format('SELECT provsql.provenance_evaluate((get_children(%L))[1],%L,%L::%s,%L,%L,%L,%L,%L)',
      token, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function)
      INTO result;
 
/*  elsif gate_type = 'cmp' then
 
    EXECUTE format('SELECT provsql.provenance_evaluate((get_children(%L))[1],%L,%L::%s,%L,%L,%L,%L,%L)',
      token, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function)
      INTO temp_result;
 
    EXECUTE format('SELECT get_extra((get_children(%L))[2])', token)
      INTO cmp_value;
 
    IF temp_result::TEXT = cmp_value::TEXT THEN
      SELECT concat('{',temp_result::TEXT,'=',cmp_value::TEXT,'}')
      INTO result;
    ELSE
      RETURN gate_zero()
      */
 
 
 
  ELSIF gate_type = 'delta' THEN
    IF delta_function IS NULL THEN
      RAISE EXCEPTION USING MESSAGE='Provenance with aggregation evaluated over a semiring without delta function';
    ELSE
      EXECUTE format('SELECT %I(a) FROM (SELECT provsql.provenance_evaluate((get_children(%L))[1],%L,%L::%s,%L,%L,%L,%L,%L) AS a) tmp',
        delta_function, token, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function)
      INTO result;
    END IF;
 
  ELSIF gate_type = 'zero' THEN
    EXECUTE format('SELECT %I(a) FROM (SELECT %L::%I AS a WHERE FALSE) temp', plus_function, element_one, value_type)
      INTO result;
 
  ELSIF gate_type = 'one' THEN
    EXECUTE format('SELECT %L::%I', element_one, value_type)
      INTO result;
 
  ELSIF gate_type = 'project' THEN
    EXECUTE format('SELECT provsql.provenance_evaluate((get_children(%L))[1],%L,%L::%s,%L,%L,%L,%L,%L)',
      token, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function)
      INTO result;
 
  ELSIF gate_type = 'annotation' THEN
    -- Transparent single-child wrapper (carries the inversion-free certificate
    -- / per-input order keys in extra, inert for every semiring): evaluate
    -- through to the child, like 'project'.
    EXECUTE format('SELECT provsql.provenance_evaluate((get_children(%L))[1],%L,%L::%s,%L,%L,%L,%L,%L)',
      token, token2value, element_one, value_type, value_type, plus_function, times_function, monus_function, delta_function)
      INTO result;
 
  ELSE
    RAISE EXCEPTION USING MESSAGE='provenance_evaluate cannot be called on formulas using ' || gate_type || ' gates; use compiled semirings instead';
  END IF;
 
  RETURN result;
END
$$ LANGUAGE plpgsql PARALLEL SAFE STABLE;
 
 
/**
 * @brief Evaluate provenance over a user-defined semiring (C version)
 *
 * Optimized C implementation of provenance_evaluate. Infers the
 * value type from element_one. Monus and delta functions are optional.
 *
 * @param token provenance token to evaluate
 * @param token2value mapping table from tokens to semiring values
 * @param element_one identity element of the semiring
 * @param plus_function semiring addition (aggregate)
 * @param times_function semiring multiplication (aggregate)
 * @param monus_function semiring monus, or NULL if not needed
 * @param delta_function δ-semiring operator, or NULL if not needed
 */
CREATE OR REPLACE FUNCTION provenance_evaluate(
  token UUID,
  token2value REGCLASS,
  element_one ANYELEMENT,
  plus_function REGPROC,
  times_function REGPROC,
  monus_function REGPROC = NULL,
  delta_function REGPROC = NULL)
  RETURNS ANYELEMENT AS
  'provsql','provenance_evaluate' LANGUAGE C STABLE;
 
/** @} */
 
/** @defgroup circuit_introspection Circuit introspection
 *  Functions for examining the structure of provenance circuits,
 *  used by visualization and where-provenance features.
 *  @{
 */
 
/** @brief Row type for sub_circuit_with_desc results */
CREATE TYPE GATE_WITH_DESC AS (f UUID, t UUID, gate_type PROVENANCE_GATE, desc_str CHARACTER VARYING, infos INTEGER[], extra TEXT);
 
/**
 * @brief Return the sub-circuit reachable from a token, with descriptions
 *
 * Recursively traverses the provenance circuit from the given token and
 * returns all edges together with input gate descriptions from the
 * mapping table.
 *
 * @param token root provenance token
 * @param token2desc mapping table providing descriptions for input gates
 */
CREATE OR REPLACE FUNCTION sub_circuit_with_desc(
  token UUID,
  token2desc REGCLASS) RETURNS SETOF GATE_WITH_DESC AS
$$
BEGIN
  RETURN QUERY EXECUTE
    'WITH RECURSIVE transitive_closure(f,t,gate_type) AS (
      SELECT $1,t,provsql.get_gate_type($1) FROM unnest(provsql.get_children($1)) AS t
        UNION ALL
      SELECT p1.t,u,provsql.get_gate_type(p1.t) FROM transitive_closure p1, unnest(provsql.get_children(p1.t)) AS u)
    SELECT *, ARRAY[(get_infos(f)).info1, (get_infos(f)).info2], get_extra(f) FROM (
      SELECT f::UUID,t::UUID,gate_type,NULL FROM transitive_closure
        UNION ALL
      SELECT p2.provenance::UUID as f, NULL::UUID, ''input'', CAST (p2.value AS varchar) FROM transitive_closure p1 JOIN ' || token2desc || ' AS p2
        ON p2.provenance=t
        UNION ALL
      SELECT provenance::UUID as f, NULL::UUID, ''input'', CAST (value AS varchar) FROM ' || token2desc || ' WHERE provenance=$1
    ) t'
  USING token LOOP;
  RETURN;
END
$$ LANGUAGE plpgsql PARALLEL SAFE;
 
/**
 * @brief Identify which table and how many columns a provenance token belongs to
 *
 * Searches all provenance-tracked tables for a row matching the given
 * token and returns the table name and column count.
 *
 * @param token provenance token to look up
 * @param table_name (OUT) the table containing this token
 * @param nb_columns (OUT) number of non-provenance columns in that table
 */
CREATE OR REPLACE FUNCTION identify_token(
  token UUID, OUT table_name REGCLASS, OUT nb_columns INTEGER) AS
$$
DECLARE
  t RECORD;
  result RECORD;
BEGIN
  table_name:=NULL;
  nb_columns:=-1;
  FOR t IN
    SELECT relname,
      (SELECT count(*) FROM pg_attribute a2 WHERE a2.attrelid=a1.attrelid AND attnum>0 AND atttypid<>0)-1 c
    FROM pg_attribute a1 JOIN pg_type ON atttypid=pg_type.oid
                        JOIN pg_class ON attrelid=pg_class.oid
                        JOIN pg_namespace ON relnamespace=pg_namespace.oid
    WHERE typname='UUID' AND relkind='r'
                                     AND nspname<>'provsql'
                                     AND attname='provsql'
  LOOP
    EXECUTE format('SELECT * FROM %I WHERE provsql=%L',t.relname,token) INTO result;
    -- Test result.provsql rather than the whole RECORD: "RECORD IS NOT NULL"
    -- is true only when every field is non-null, so a matched row that has any
    -- NULL data column would be wrongly skipped. The provsql column is the
    -- (non-null) token we matched on, so it is set iff a row was found.
    IF result.provsql IS NOT NULL THEN
      table_name:=t.relname;
      nb_columns:=t.c;
      EXIT;
    END IF;
  END LOOP;
END
$$ LANGUAGE plpgsql STRICT;
 
/**
 * @brief Return the sub-circuit for where-provenance computation
 *
 * Similar to sub_circuit_with_desc but resolves input gates to their
 * source table and column count for where-provenance evaluation.
 */
CREATE OR REPLACE FUNCTION sub_circuit_for_where(token UUID)
  RETURNS TABLE(f UUID, t UUID, gate_type PROVENANCE_GATE, table_name REGCLASS, nb_columns INTEGER, infos INTEGER[], extra TEXT) AS
$$
    WITH RECURSIVE transitive_closure(f,t,idx,gate_type) AS (
      SELECT $1,t,id,provsql.get_gate_type($1) FROM unnest(provsql.get_children($1)) WITH ORDINALITY AS a(t,id)
        UNION ALL
      SELECT p1.t,u,id,provsql.get_gate_type(p1.t) FROM transitive_closure p1, unnest(provsql.get_children(p1.t)) WITH ORDINALITY AS a(u, id)
    ) SELECT f, t, gate_type, table_name, nb_columns, ARRAY[(get_infos(f)).info1, (get_infos(f)).info2], get_extra(f) FROM (
      -- One row per distinct (parent, child, child-position) edge.  The
      -- recursive closure (UNION ALL) re-emits a gate's outgoing edges once per
      -- path that reaches it, so a *shared* non-input gate would otherwise be
      -- reported with duplicate edges; DISTINCT on the (f,t,idx) triple
      -- collapses those while keeping genuine repeated children (same f,t,
      -- different idx, e.g. a self-product).  Without this, a shared
      -- single-child gate (notably an inversion-free order-marker annotation)
      -- gets its child wired k times in the where-circuit -> the locator sets
      -- are duplicated k-fold.
      SELECT DISTINCT f, t::UUID, idx, gate_type, NULL::REGCLASS AS table_name, NULL::INTEGER AS nb_columns FROM transitive_closure
      UNION ALL
        SELECT DISTINCT t, NULL::UUID, NULL::INT, 'input'::PROVENANCE_GATE, (id).table_name, (id).nb_columns FROM transitive_closure JOIN (SELECT t AS prov, provsql.identify_token(t) as id FROM transitive_closure WHERE t NOT IN (SELECT f FROM transitive_closure)) temp ON t=prov
      UNION ALL
        SELECT DISTINCT $1, NULL::UUID, NULL::INT, 'input'::PROVENANCE_GATE, (id).table_name, (id).nb_columns FROM (SELECT provsql.identify_token($1) AS id WHERE $1 NOT IN (SELECT f FROM transitive_closure)) temp
      ) t
    -- order each parent's edges by child position so the where-circuit's TIMES
    -- concatenation reproduces the column order (input rows have idx NULL).
    ORDER BY f, idx
$$
LANGUAGE sql;
 
/**
 * @brief BFS expansion of a provenance circuit, capped at @p max_depth
 *
 * Returns one row per (parent, child) edge in the BFS-bounded subgraph
 * rooted at @p root, plus one row for the root with <tt>parent</tt> and
 * <tt>child_pos</tt> NULL.  Provenance circuits are DAGs, so a child gate
 * may have several parents within the bound; each such edge is reported
 * as a separate row, so callers must deduplicate on <tt>node</tt> if they
 * need a one-row-per-node view.
 *
 * <tt>depth</tt> is the node's longest-path distance from @p root
 * within the depth bound (the standard circuit-depth notion), so for
 * an edge (parent, child) it is the case that
 * <tt>child.depth &gt;= parent.depth + 1</tt>, except at the
 * <tt>max_depth</tt> truncation frontier.  A node at
 * <tt>depth = max_depth</tt> is not
 * expanded; callers can detect a partial expansion by comparing
 * <tt>provsql.get_children</tt> length against the number of outgoing
 * edges reported.
 *
 * <tt>info1</tt> and <tt>info2</tt> are the INTEGER values stored on
 * the gate by <tt>provsql.set_infos</tt>, formatted as TEXT; their
 * meaning is gate-type-specific (see <tt>provsql.set_infos</tt>).
 *
 * @param root root provenance token
 * @param max_depth maximum BFS depth (default 8)
 */
CREATE OR REPLACE FUNCTION circuit_subgraph(root UUID, max_depth INT DEFAULT 8)
  RETURNS TABLE(node UUID, parent UUID, child_pos INT, gate_type TEXT, info1 TEXT, info2 TEXT, depth INT) AS
$$
  WITH RECURSIVE bfs(node, parent, child_pos, depth) AS (
    SELECT root, NULL::UUID, NULL::INT, 0
      UNION ALL
    SELECT c.t, b.node, c.idx::INT, b.depth + 1
    FROM bfs b
    CROSS JOIN LATERAL unnest(provsql.get_children(b.node))
      WITH ORDINALITY AS c(t, idx)
    WHERE b.depth < max_depth
  ),
  -- Each node's canonical depth is its longest-path distance from the
  -- root (the standard circuit-depth notion: the longest chain of
  -- gates separating the node from the output). The recursive CTE
  -- enumerates paths up to @c max_depth, so MAX over those is the
  -- longest path of length at most @c max_depth.
  node_depth AS (
    SELECT node, MAX(depth) AS depth FROM bfs GROUP BY node
  ),
  -- All distinct (parent, child, child_pos) triples seen during the BFS.
  -- A child reached from k parents within the bound contributes k rows.
  -- Self-joins (times(x, x)) contribute one row per child position.
  edges AS (
    SELECT DISTINCT parent, node AS child, child_pos
    FROM bfs WHERE parent IS NOT NULL
  )
  SELECT
    d.node,
    e.parent,
    e.child_pos,
    provsql.get_gate_type(d.node)::TEXT,
    i.info1::TEXT,
    i.info2::TEXT,
    d.depth
  FROM node_depth d
  LEFT JOIN edges e ON e.child = d.node
  LEFT JOIN LATERAL provsql.get_infos(d.node) i ON TRUE
  ORDER BY d.depth, d.node, e.parent;
$$ LANGUAGE sql STABLE PARALLEL SAFE;
 
/**
 * @brief BFS subgraph of the IN-MEMORY simplified circuit rooted at @p root.
 *
 * Same row shape as @ref circuit_subgraph plus an inline @c extra
 * column, but built from the @c GenericCircuit returned by
 * @c getGenericCircuit -- i.e. AFTER @c provsql.simplify_on_load
 * passes (RangeCheck, ...) have rewritten any decidable @c gate_cmp
 * into Bernoulli @c gate_input / @c gate_zero / @c gate_one leaves.
 * Lets a renderer show the user what the evaluator actually sees,
 * without mutating the persisted DAG.
 *
 * Returns @c jsonb (an array of objects) rather than @c SETOF RECORD
 * to keep the C++ implementation free of SRF / @c FuncCallContext
 * boilerplate; callers either consume the array directly or expand
 * it via @c jsonb_array_elements.
 *
 * @param root      Root provenance token.
 * @param max_depth Maximum BFS depth (default 8).
 */
CREATE OR REPLACE FUNCTION simplified_circuit_subgraph(
  root UUID, max_depth INT DEFAULT 8) RETURNS jsonb
  AS 'provsql','simplified_circuit_subgraph'
  LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief Empirical histogram of a scalar sub-circuit
 *
 * Returns a jsonb array of @c {bin_lo, bin_hi, count} objects covering
 * the observed @c [min, max] range of @p bins equal-width samples from
 * the sub-circuit rooted at @p token.  Sample count is taken from
 * @c provsql.rv_mc_samples; pinning @c provsql.monte_carlo_seed makes
 * the result reproducible.
 *
 * Accepted root gate types are the scalar ones: @c gate_value (Dirac
 * at the constant, single bin), @c gate_rv (sampled from the leaf's
 * distribution), and @c gate_arith (sampled by recursing through the
 * arithmetic DAG, with shared @c gate_rv leaves correctly correlated
 * within an iteration).  Any other gate type raises.
 *
 * @param token Root provenance token of a scalar sub-circuit.
 * @param bins  Number of equal-width histogram bins (default 30).
 * @param prov  Conditioning event (defaults to @c gate_one() = no
 *              conditioning).  When non-trivial, the histogram is
 *              over the conditional distribution recovered by
 *              rejection sampling on the joint circuit with @p token.
 */
CREATE OR REPLACE FUNCTION rv_histogram(
  token UUID, bins INT DEFAULT 30, prov UUID DEFAULT gate_one())
  RETURNS jsonb
  AS 'provsql','rv_histogram'
  LANGUAGE C VOLATILE PARALLEL SAFE;
 
/**
 * @brief Sample the closed-form PDF and CDF of a (possibly truncated)
 *        scalar distribution.
 *
 * Returns @c {"pdf": [{x, p}, ...], "cdf": [{x, p}, ...]} with @p samples
 * evenly-spaced points spanning the distribution's natural display
 * range (intersected with the conditioning event's interval when
 * @c prov is non-trivial).  Used by ProvSQL Studio's Distribution
 * profile panel to overlay the analytical curve on the empirical
 * histogram from :sqlfunc:`rv_histogram` -- the simplifier's
 * analytical wins (e.g. @c c·Exp(λ) folding to @c Exp(λ/c)) become
 * visible as a smooth curve riding over the MC-sampled bars.
 *
 * Returns @c NULL when the root sub-circuit is not a closed-form
 * shape (V1: only bare @c gate_rv of Normal / Uniform / Exponential
 * / INTEGER-Erlang).  The frontend reads @c NULL as "skip overlay"
 * without erroring, so the caller can dispatch this in parallel with
 * @c rv_histogram regardless of the underlying shape.
 *
 * @param token   Scalar gate token (random_variable's UUID).
 * @param samples Number of (x, p) points; must be >= 2.
 * @param prov    Conditioning event (defaults to @c gate_one() = no
 *                conditioning).  When non-trivial, the curves are
 *                over the truncated distribution.
 */
CREATE OR REPLACE FUNCTION rv_analytical_curves(
  token UUID, samples INT DEFAULT 100, prov UUID DEFAULT gate_one())
  RETURNS jsonb
  AS 'provsql','rv_analytical_curves'
  LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Draw conditional Monte Carlo samples from a scalar gate.
 *
 * Returns up to @c n samples of the scalar value at @c token; when
 * @c prov is not the trivial @c gate_one() event, draws are accepted
 * only on iterations where @c prov evaluates true (rejection
 * sampling).  Shared @c gate_rv leaves between @c token and @c prov
 * are loaded into a single joint circuit so the indicator's draw
 * and the value's draw share their per-iteration state.
 *
 * @param token  Scalar sub-circuit root.
 * @param n      Number of accepted samples to attempt.
 * @param prov   Conditioning event (defaults to @c gate_one() = no
 *               conditioning).
 *
 * Emits a @c NOTICE when the conditional acceptance rate yields fewer
 * than @c n samples within the @c provsql.rv_mc_samples budget so the
 * caller can choose to widen the budget.
 */
CREATE OR REPLACE FUNCTION rv_sample(
  token UUID, n INTEGER, prov UUID DEFAULT gate_one())
  RETURNS SETOF float8
  AS 'provsql','rv_sample'
  LANGUAGE C VOLATILE PARALLEL SAFE;
 
/**
 * @brief Resolve an input gate UUID back to its source row
 *
 * Searches every provenance-tracked relation for a row whose
 * <tt>provsql</tt> column equals @p UUID and returns the relation's
 * REGCLASS together with the row encoded as JSONB.  Returns zero
 * rows when @p UUID is not the provenance token of any tracked row,
 * including when it identifies an internal gate (<tt>plus</tt>,
 * <tt>times</tt>, ...) rather than an input.
 *
 * Ordinarily exactly one row is returned, but if the same UUID
 * happens to appear as a <tt>provsql</tt> value in several tracked
 * tables, all matches are returned.
 *
 * @param UUID token to resolve
 */
CREATE OR REPLACE FUNCTION resolve_input(UUID UUID)
  RETURNS TABLE(relation REGCLASS, row_data JSONB) AS
$$
DECLARE
  t RECORD;
  rel REGCLASS;
  rd  JSONB;
  -- ProvSQL's rewriter unconditionally appends a provsql column to the
  -- targetlist of any SELECT reading from a tracked relation; capture and
  -- discard it here rather than disabling the rewriter for the whole call.
  ign UUID;
BEGIN
  FOR t IN
    SELECT c.oid::REGCLASS AS regc
    FROM pg_attribute a
         JOIN pg_class c ON a.attrelid = c.oid
         JOIN pg_namespace ns ON c.relnamespace = ns.oid
         JOIN pg_type ty ON a.atttypid = ty.oid
    WHERE a.attname = 'provsql'
      AND ty.typname = 'UUID'
      AND c.relkind = 'r'
      AND ns.nspname <> 'provsql'
      AND a.attnum > 0
  LOOP
    FOR rel, rd, ign IN
      EXECUTE format(
        'SELECT %L::REGCLASS, to_jsonb(t) - ''provsql'', t.provsql FROM %s AS t WHERE provsql = $1',
        t.regc, t.regc)
      USING UUID
    LOOP
      relation := rel;
      row_data := rd;
      RETURN NEXT;
    END LOOP;
  END LOOP;
END
$$ LANGUAGE plpgsql STABLE;
 
/** @} */
 
/** @defgroup agg_token_type Type for the result of aggregate queries
 *
 *  Custom type <tt>AGG_TOKEN</tt> for a provenance semimodule value, to
 *  be used in attributes that are computed as a result of aggregation.
 *  As for provenance tokens, this is simply a UUID, but this UUID is
 *  displayed in a specific way (as the result of the aggregation
 *  followed by a "(*)") to help with readability.
 *
 *  The TEXT output is controlled by the
 *  <tt>provsql.aggtoken_text_as_uuid</tt> GUC. By default it is off and
 *  the cell renders as <tt>"value (*)"</tt>. When set to on (typical
 *  for UI layers such as ProvSQL Studio), the cell renders as the
 *  underlying UUID instead, so the caller can click through to the
 *  provenance circuit; the value side is then recovered via
 *  <tt>provsql.agg_token_value_text(UUID)</tt>.
 *
 *  @{
 */
 
CREATE TYPE AGG_TOKEN;
 
/** @brief Input function for the AGG_TOKEN type (parses TEXT representation) */
CREATE OR REPLACE FUNCTION agg_token_in(CSTRING)
  RETURNS AGG_TOKEN
  AS 'provsql','agg_token_in' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Output function for the AGG_TOKEN type
 *
 * Default: produces the human-friendly @c "value (*)" form, where
 * @c value is the running aggregate state.
 *
 * When the @c provsql.aggtoken_text_as_uuid GUC is on, returns the
 * underlying provenance UUID instead. UI layers (notably ProvSQL
 * Studio) flip this on per session so aggregate cells expose the
 * circuit root UUID for click-through; the @c "value (*)" display
 * string is recovered via @c provsql.agg_token_value_text(UUID).
 *
 * Marked STABLE rather than IMMUTABLE because the chosen output
 * shape now depends on a GUC that the same session can flip at
 * runtime.
 */
CREATE OR REPLACE FUNCTION agg_token_out(AGG_TOKEN)
  RETURNS CSTRING
  AS 'provsql','agg_token_out' LANGUAGE C STABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to its TEXT representation */
CREATE OR REPLACE FUNCTION agg_token_cast(AGG_TOKEN)
  RETURNS TEXT
  AS 'provsql','agg_token_cast' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE TYPE AGG_TOKEN (
  internallength = 117,
  input = agg_token_in,
  output = agg_token_out,
  alignment = char
);
 
/** @brief Extract the UUID from an AGG_TOKEN (implicit cast to UUID) */
CREATE OR REPLACE FUNCTION agg_token_uuid(aggtok AGG_TOKEN)
  RETURNS UUID AS
$$
BEGIN
  RETURN agg_token_cast(aggtok)::UUID;
END
$$ LANGUAGE plpgsql STRICT SET search_path=provsql,pg_temp,public SECURITY DEFINER IMMUTABLE PARALLEL SAFE;
 
/** @brief Implicit PostgreSQL cast from AGG_TOKEN to UUID (delegates to agg_token_uuid()) */
CREATE CAST (AGG_TOKEN AS UUID) WITH FUNCTION agg_token_uuid(AGG_TOKEN) AS IMPLICIT;
 
/**
 * @brief Deterministic truth of a Boolean guard sub-circuit over aggregate
 *        comparisons, evaluated in the actual world (all input tuples present).
 *
 * Guards are the shapes @c having_Expr_to_provenance_cmp mints: @c cmp gates
 * over aggregate-valued children (comparison-operator OID in @c info1),
 * @c times / @c plus combinations (AND / OR, with negation pushed into the
 * comparison operators), and the @c one / @c zero indicators of regular
 * (aggregate-free) conditions.  Uses Kleene three-valued logic: returns
 * @c NULL on any other gate shape, or when an operand's deterministic value
 * cannot be resolved.
 */
CREATE OR REPLACE FUNCTION agg_guard_holds(token UUID)
  RETURNS BOOLEAN AS
$$
DECLARE
  gt PROVENANCE_GATE := get_gate_type(token);
  ch UUID[];
  opname TEXT;
  l NUMERIC;
  r NUMERIC;
  all_true BOOLEAN;
  any_true BOOLEAN;
  any_null BOOLEAN;
BEGIN
  IF gt = 'one' THEN
    RETURN true;
  ELSIF gt = 'zero' THEN
    RETURN false;
  ELSIF gt IN ('times', 'plus') THEN
    SELECT bool_and(h), bool_or(h), bool_or(h IS NULL)
      INTO all_true, any_true, any_null
      FROM (SELECT provsql.agg_guard_holds(c) AS h
            FROM unnest(get_children(token)) AS c) AS s;
    IF gt = 'times' THEN
      -- AND: false dominates unknown (bool_and skips NULL inputs, so it is
      -- false exactly when some child is false).
      RETURN CASE WHEN NOT all_true THEN false
                  WHEN any_null THEN NULL
                  ELSE true END;
    ELSE
      -- OR: true dominates unknown.
      RETURN CASE WHEN any_true THEN true
                  WHEN any_null THEN NULL
                  ELSE false END;
    END IF;
  ELSIF gt = 'cmp' THEN
    ch := get_children(token);
    l := agg_gate_value(ch[1]);
    r := agg_gate_value(ch[2]);
    IF l IS NULL OR r IS NULL THEN
      RETURN NULL;
    END IF;
    SELECT oprname INTO opname
      FROM pg_catalog.pg_operator WHERE oid = (get_infos(token)).info1;
    RETURN CASE opname
      WHEN '<'  THEN l <  r
      WHEN '<=' THEN l <= r
      WHEN '='  THEN l =  r
      WHEN '<>' THEN l <> r
      WHEN '>=' THEN l >= r
      WHEN '>'  THEN l >  r
    END;
  END IF;
  RETURN NULL;
END
$$ LANGUAGE plpgsql STABLE STRICT PARALLEL SAFE
  SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Deterministic (actual-world) scalar value of an aggregate-carrying
 *        gate.
 *
 * Resolves the value an aggregate expression takes on the actual data -- the
 * value an @c AGG_TOKEN display cell carries: @c agg / @c arith gates RECORD
 * it in @c extra (set by aggregate evaluation and @c agg_arith_make), a
 * @c value gate carries its constant, a @c semimod wraps a value gate, a
 * @c conditioned gate has its target's value, and a @c case gate selects the
 * first branch whose guard holds in the actual world (per
 * @c agg_guard_holds), else the default.  Returns @c NULL when the gate is
 * not aggregate-carrying or the value cannot be resolved (e.g. a
 * non-NUMERIC aggregate).
 */
CREATE OR REPLACE FUNCTION agg_gate_value(token UUID)
  RETURNS NUMERIC AS
$$
DECLARE
  gt PROVENANCE_GATE := get_gate_type(token);
  ch UUID[];
  n INTEGER;
  holds BOOLEAN;
BEGIN
  IF gt IN ('agg', 'arith', 'value') THEN
    BEGIN
      RETURN get_extra(token)::NUMERIC;
    EXCEPTION WHEN others THEN
      RETURN NULL;   -- non-NUMERIC aggregate (e.g. min over TEXT)
    END;
  ELSIF gt = 'semimod' THEN
    RETURN agg_gate_value((get_children(token))[2]);
  ELSIF gt = 'conditioned' THEN
    RETURN agg_gate_value((get_children(token))[1]);
  ELSIF gt = 'case' THEN
    ch := get_children(token);
    n := array_length(ch, 1);
    FOR i IN 1 .. (n - 1) / 2 LOOP
      holds := agg_guard_holds(ch[2 * i - 1]);
      IF holds IS NULL THEN
        RETURN NULL;
      ELSIF holds THEN
        RETURN agg_gate_value(ch[2 * i]);
      END IF;
    END LOOP;
    RETURN agg_gate_value(ch[n]);
  END IF;
  RETURN NULL;
END
$$ LANGUAGE plpgsql STABLE STRICT PARALLEL SAFE
  SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Recover the @c "value (*)" display string for an aggregation gate
 *
 * Companion helper to the @c provsql.aggtoken_text_as_uuid GUC. With
 * the GUC on, an @c AGG_TOKEN cell prints as the underlying provenance
 * UUID, which is convenient for tooling that wants to click through to
 * the circuit but loses the human-readable aggregate value. This
 * function takes such a UUID and returns the original @c "value (*)"
 * string by reading the gate's @c extra (set by aggregate evaluation
 * for @c agg gates, and by @c agg_arith_make for the @c arith gates
 * that AGG_TOKEN arithmetic mints); for the other aggregate-carrying
 * gates (@c case, @c conditioned, @c semimod, @c value) the value is
 * resolved through the circuit by @c agg_gate_value. Returns @c NULL
 * if @p token does not resolve to an aggregate-carrying gate.
 *
 * @param token UUID of an @c agg gate (typically obtained from an
 *              @c AGG_TOKEN cell when @c aggtoken_text_as_uuid is on,
 *              or via a manual UUID cast otherwise).
 */
CREATE OR REPLACE FUNCTION agg_token_value_text(token UUID)
  RETURNS TEXT AS
$$
  SELECT CASE
    -- agg gates: extra is set by aggregate evaluation; arith gates
    -- (AGG_TOKEN arithmetic): extra is recorded by agg_arith_make.
    WHEN provsql.get_gate_type(token) IN ('agg', 'arith')
      THEN provsql.get_extra(token) || ' (*)'
    -- other aggregate-carrying gates: resolve the actual-world value
    -- through the circuit.
    WHEN provsql.get_gate_type(token) IN ('case', 'conditioned', 'semimod', 'value')
      THEN provsql.agg_gate_value(token)::TEXT || ' (*)'
    ELSE NULL
  END;
$$ LANGUAGE sql STABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to NUMERIC (extracts the aggregate value, loses provenance) */
CREATE OR REPLACE FUNCTION agg_token_to_numeric(AGG_TOKEN)
  RETURNS NUMERIC
  AS 'provsql','agg_token_to_numeric' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to double precision (extracts the aggregate value, loses provenance) */
CREATE OR REPLACE FUNCTION agg_token_to_float8(AGG_TOKEN)
  RETURNS double precision
  AS 'provsql','agg_token_to_float8' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to INTEGER (extracts the aggregate value, loses provenance) */
CREATE OR REPLACE FUNCTION agg_token_to_int4(AGG_TOKEN)
  RETURNS INTEGER
  AS 'provsql','agg_token_to_int4' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to bigint (extracts the aggregate value, loses provenance) */
CREATE OR REPLACE FUNCTION agg_token_to_int8(AGG_TOKEN)
  RETURNS bigint
  AS 'provsql','agg_token_to_int8' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Cast an AGG_TOKEN to TEXT (extracts the aggregate value, loses provenance) */
CREATE OR REPLACE FUNCTION agg_token_to_text(AGG_TOKEN)
  RETURNS TEXT
  AS 'provsql','agg_token_to_text' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Assignment cast from AGG_TOKEN to NUMERIC (extracts the scalar
 *  value, dropping provenance).  ASSIGNMENT, not IMPLICIT: provenance-
 *  preserving arithmetic on aggregates is provided by the native
 *  AGG_TOKEN operators below, so an implicit NUMERIC coercion would only
 *  silently steal `s + 1` away from them (and reroute it differently
 *  depending on whether provsql is in search_path).  Write `s::NUMERIC`
 *  to opt into the lossy scalar. */
CREATE CAST (AGG_TOKEN AS NUMERIC) WITH FUNCTION agg_token_to_numeric(AGG_TOKEN) AS ASSIGNMENT;
 
-- ---------------------------------------------------------------------
-- Arithmetic on aggregates (AGG_TOKEN)
--
-- Mirrors the random_variable arithmetic surface: the operators build a
-- `gate_arith` over the operand provenance UUIDs (via provenance_arith,
-- info1 = PROVSQL_ARITH_*), so the arithmetic is recorded symbolically
-- in the circuit and can be resolved when a comparison (gate_cmp) over
-- the result is evaluated.  Unlike random_variable (a bare UUID), an
-- AGG_TOKEN also carries a running scalar value, so each operator
-- additionally computes the resulting value and bundles it back with the
-- new gate.
-- ---------------------------------------------------------------------
 
/** @brief Running value of an AGG_TOKEN as NUMERIC, without the
 *  provenance-loss warning the public cast emits (internal use). */
CREATE OR REPLACE FUNCTION agg_token_value(AGG_TOKEN)
  RETURNS NUMERIC
  AS 'provsql','agg_token_value' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Bundle a provenance gate UUID with a running value into an
 *  AGG_TOKEN (inverse of the agg_token_uuid / agg_token_value
 *  accessors). */
CREATE OR REPLACE FUNCTION agg_token_make(tok UUID, val NUMERIC)
  RETURNS AGG_TOKEN AS
$$
  SELECT format('( %s , %s )', tok::TEXT, val::TEXT)::provsql.AGG_TOKEN;
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE
  SET search_path=provsql,pg_temp,public;
 
/** @brief Lift a scalar NUMERIC constant into a gate_value leaf and
 *  return its UUID, so it can be a child of a gate_arith (the agg-side
 *  analogue of as_random for random_variable). */
CREATE OR REPLACE FUNCTION agg_value_gate(v NUMERIC)
  RETURNS UUID AS
$$
DECLARE
  token UUID := public.uuid_generate_v5(
    provsql.uuid_ns_provsql(), concat('value', v::TEXT));
BEGIN
  PERFORM provsql.create_gate(token, 'value');
  PERFORM provsql.set_extra(token, v::TEXT);
  RETURN token;
END
$$ LANGUAGE plpgsql STRICT IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Mint (or reuse) the gate_arith for an AGG_TOKEN arithmetic
 *  result and return the AGG_TOKEN carrying it.
 *
 * Also records the computed scalar in the gate's @c extra -- exactly
 * what aggregate evaluation does for @c agg gates -- so
 * @c agg_token_value_text can recover the @c "value (*)" display from
 * the bare UUID (as ProvSQL Studio does for result cells under
 * @c provsql.aggtoken_text_as_uuid). The gate UUID is deterministic in
 * (op, children), so re-recording the (identical) value is idempotent. */
CREATE OR REPLACE FUNCTION agg_arith_make(op INT, children UUID[], val NUMERIC)
  RETURNS AGG_TOKEN AS
$$
DECLARE
  token UUID := provsql.provenance_arith(op, children);
BEGIN
  PERFORM provsql.set_extra(token, val::TEXT);
  RETURN provsql.agg_token_make(token, val);
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
-- AGG_TOKEN <op> AGG_TOKEN --------------------------------------------
/** @brief AGG_TOKEN + AGG_TOKEN (gate_arith PLUS). */
CREATE OR REPLACE FUNCTION agg_token_plus(a AGG_TOKEN, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(0, ARRAY[(a)::UUID, (b)::UUID],
     provsql.agg_token_value(a) + provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN - AGG_TOKEN (gate_arith MINUS). */
CREATE OR REPLACE FUNCTION agg_token_minus(a AGG_TOKEN, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(2, ARRAY[(a)::UUID, (b)::UUID],
     provsql.agg_token_value(a) - provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN * AGG_TOKEN (gate_arith TIMES). */
CREATE OR REPLACE FUNCTION agg_token_times(a AGG_TOKEN, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(1, ARRAY[(a)::UUID, (b)::UUID],
     provsql.agg_token_value(a) * provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN / AGG_TOKEN (gate_arith DIV). */
CREATE OR REPLACE FUNCTION agg_token_div(a AGG_TOKEN, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(3, ARRAY[(a)::UUID, (b)::UUID],
     provsql.agg_token_value(a) / provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief Unary -AGG_TOKEN (gate_arith NEG). */
CREATE OR REPLACE FUNCTION agg_token_neg(a AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(4, ARRAY[(a)::UUID],
     - provsql.agg_token_value(a)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
-- AGG_TOKEN <op> NUMERIC ----------------------------------------------
/** @brief AGG_TOKEN + NUMERIC (gate_arith PLUS, constant lifted to a value gate). */
CREATE OR REPLACE FUNCTION agg_token_plus_numeric(a AGG_TOKEN, b NUMERIC)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(0, ARRAY[(a)::UUID, provsql.agg_value_gate(b)],
     provsql.agg_token_value(a) + b); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN - NUMERIC. */
CREATE OR REPLACE FUNCTION agg_token_minus_numeric(a AGG_TOKEN, b NUMERIC)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(2, ARRAY[(a)::UUID, provsql.agg_value_gate(b)],
     provsql.agg_token_value(a) - b); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN * NUMERIC. */
CREATE OR REPLACE FUNCTION agg_token_times_numeric(a AGG_TOKEN, b NUMERIC)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(1, ARRAY[(a)::UUID, provsql.agg_value_gate(b)],
     provsql.agg_token_value(a) * b); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief AGG_TOKEN / NUMERIC. */
CREATE OR REPLACE FUNCTION agg_token_div_numeric(a AGG_TOKEN, b NUMERIC)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(3, ARRAY[(a)::UUID, provsql.agg_value_gate(b)],
     provsql.agg_token_value(a) / b); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
-- NUMERIC <op> AGG_TOKEN ----------------------------------------------
/** @brief NUMERIC + AGG_TOKEN. */
CREATE OR REPLACE FUNCTION numeric_plus_agg_token(a NUMERIC, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(0, ARRAY[provsql.agg_value_gate(a), (b)::UUID],
     a + provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief NUMERIC - AGG_TOKEN. */
CREATE OR REPLACE FUNCTION numeric_minus_agg_token(a NUMERIC, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(2, ARRAY[provsql.agg_value_gate(a), (b)::UUID],
     a - provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief NUMERIC * AGG_TOKEN. */
CREATE OR REPLACE FUNCTION numeric_times_agg_token(a NUMERIC, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(1, ARRAY[provsql.agg_value_gate(a), (b)::UUID],
     a * provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/** @brief NUMERIC / AGG_TOKEN. */
CREATE OR REPLACE FUNCTION numeric_div_agg_token(a NUMERIC, b AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$ SELECT provsql.agg_arith_make(3, ARRAY[provsql.agg_value_gate(a), (b)::UUID],
     a / provsql.agg_token_value(b)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
-- Operator declarations -----------------------------------------------
CREATE OPERATOR + (LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_plus,  COMMUTATOR = +);
CREATE OPERATOR - (LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_minus);
CREATE OPERATOR * (LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_times, COMMUTATOR = *);
CREATE OPERATOR / (LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_div);
CREATE OPERATOR - (RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_neg);
 
CREATE OPERATOR + (LEFTARG=AGG_TOKEN, RIGHTARG=NUMERIC, PROCEDURE=agg_token_plus_numeric,  COMMUTATOR = +);
CREATE OPERATOR - (LEFTARG=AGG_TOKEN, RIGHTARG=NUMERIC, PROCEDURE=agg_token_minus_numeric);
CREATE OPERATOR * (LEFTARG=AGG_TOKEN, RIGHTARG=NUMERIC, PROCEDURE=agg_token_times_numeric, COMMUTATOR = *);
CREATE OPERATOR / (LEFTARG=AGG_TOKEN, RIGHTARG=NUMERIC, PROCEDURE=agg_token_div_numeric);
 
CREATE OPERATOR + (LEFTARG=NUMERIC, RIGHTARG=AGG_TOKEN, PROCEDURE=numeric_plus_agg_token,  COMMUTATOR = +);
CREATE OPERATOR - (LEFTARG=NUMERIC, RIGHTARG=AGG_TOKEN, PROCEDURE=numeric_minus_agg_token);
CREATE OPERATOR * (LEFTARG=NUMERIC, RIGHTARG=AGG_TOKEN, PROCEDURE=numeric_times_agg_token, COMMUTATOR = *);
CREATE OPERATOR / (LEFTARG=NUMERIC, RIGHTARG=AGG_TOKEN, PROCEDURE=numeric_div_agg_token);
 
/** @brief Assignment cast from AGG_TOKEN to double precision */
CREATE CAST (AGG_TOKEN AS double precision) WITH FUNCTION agg_token_to_float8(AGG_TOKEN) AS ASSIGNMENT;
/** @brief Assignment cast from AGG_TOKEN to INTEGER */
CREATE CAST (AGG_TOKEN AS INTEGER) WITH FUNCTION agg_token_to_int4(AGG_TOKEN) AS ASSIGNMENT;
/** @brief Assignment cast from AGG_TOKEN to bigint */
CREATE CAST (AGG_TOKEN AS bigint) WITH FUNCTION agg_token_to_int8(AGG_TOKEN) AS ASSIGNMENT;
/** @brief Assignment cast from AGG_TOKEN to TEXT (extracts value, not UUID) */
CREATE CAST (AGG_TOKEN AS TEXT) WITH FUNCTION agg_token_to_text(AGG_TOKEN) AS ASSIGNMENT;
 
/**
 * @brief Condition a discrete aggregate's distribution on an event:
 *        @c "SUM(x) | C".
 *
 * Mirrors @c random_variable_cond for the @c AGG_TOKEN carrier: returns a
 * conditioned @c AGG_TOKEN that flows onward, its provenance token wrapped in
 * the composable two-child @c gate_conditioned @c [agg_target, condition]
 * while its running value is preserved.  The moment / support dispatchers
 * unpack it (@c agg_conditioned_target + @c rv_conditioned_prov) and route
 * through the existing @c agg_raw_moment with the condition conjoined into the
 * @c prov argument, so @c expected(SUM(x)|C) / @c variance(SUM(x)|C) report
 * the conditional aggregate distribution.  Nested conditioning folds.
 */
CREATE OR REPLACE FUNCTION agg_token_cond(a AGG_TOKEN, cond UUID)
  RETURNS AGG_TOKEN AS
$$
DECLARE
  tok UUID;
  ev  UUID;
  result UUID;
  ch UUID[];
BEGIN
  IF cond IS NULL OR cond = gate_one() THEN
    RETURN a;
  END IF;
 
  tok := (a)::UUID;
  IF get_gate_type(tok) = 'conditioned'
     AND array_length(get_children(tok), 1) = 2 THEN
    ch  := get_children(tok);
    tok := ch[1];
    ev  := provenance_times(ch[2], cond);
  ELSE
    ev := cond;
  END IF;
 
  result := public.uuid_generate_v5(uuid_ns_provsql(),
                                    concat('conditioned', tok, ev));
  PERFORM create_gate(result, 'conditioned', ARRAY[tok, ev]);
  RETURN agg_token_make(result, agg_token_value(a));
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
CREATE OPERATOR | (
  LEFTARG   = AGG_TOKEN,
  RIGHTARG  = UUID,
  PROCEDURE = agg_token_cond
);
 
/**
 * @brief Placeholder for @c "SUM(x) | (predicate)" on an AGG_TOKEN.
 *
 * Lets the conditioning event be a natural Boolean predicate (e.g.
 * @c "SUM(x) | (SUM(x) > 5)") instead of a hand-built gate.  Never executes:
 * the planner converts the Boolean operand into a condition gate and emits
 * @c agg_token_cond.
 */
CREATE OR REPLACE FUNCTION agg_token_cond_predicate(
  a AGG_TOKEN, predicate BOOLEAN) RETURNS AGG_TOKEN AS
$$
BEGIN
  RAISE EXCEPTION 'AGG_TOKEN | (predicate) must be rewritten by the ProvSQL '
    'planner hook: the right operand must be a Boolean combination of '
    'aggregate / random_variable comparisons (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR | (
  LEFTARG   = AGG_TOKEN,
  RIGHTARG  = BOOLEAN,
  PROCEDURE = agg_token_cond_predicate
);
 
/**
 * @brief Unpack the target of a conditioned @c AGG_TOKEN.
 *
 * For a @c "SUM(x) | C" whose provenance token is the two-child
 * @c gate_conditioned @c [agg_target, condition] returns the AGG_TOKEN over
 * @c agg_target (same running value); for any other AGG_TOKEN returns it
 * unchanged.  The conditioning event itself is recovered separately via
 * @c rv_conditioned_prov on the token's UUID.
 */
CREATE OR REPLACE FUNCTION agg_conditioned_target(a AGG_TOKEN)
  RETURNS AGG_TOKEN AS
$$
  SELECT CASE
    WHEN provsql.get_gate_type((a)::UUID) = 'conditioned'
         AND array_length(provsql.get_children((a)::UUID), 1) = 2
    THEN provsql.agg_token_make(
           (provsql.get_children((a)::UUID))[1], provsql.agg_token_value(a))
    ELSE a
  END;
$$ LANGUAGE sql STABLE PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Placeholder comparison of AGG_TOKEN with NUMERIC
 *
 * This function is never actually called; it exists so the SQL parser
 * accepts comparison operators between AGG_TOKEN and NUMERIC values.
 * The ProvSQL query rewriter replaces these comparisons at plan time.
 */
CREATE OR REPLACE FUNCTION agg_token_comp_numeric(a AGG_TOKEN, b NUMERIC)
RETURNS BOOLEAN
LANGUAGE plpgsql
IMMUTABLE STRICT PARALLEL SAFE
AS $$
BEGIN
  RAISE EXCEPTION 'Comparison AGG_TOKEN-NUMERIC not implemented, should be replaced by ProvSQL behavior';
END;
$$;
 
/**
 * @brief Placeholder comparison of NUMERIC with AGG_TOKEN
 *
 * Symmetric to agg_token_comp_numeric; never actually called.
 * The ProvSQL query rewriter replaces these comparisons at plan time.
 */
CREATE OR REPLACE FUNCTION numeric_comp_agg_token(a NUMERIC, b AGG_TOKEN)
RETURNS BOOLEAN
LANGUAGE plpgsql
IMMUTABLE STRICT PARALLEL SAFE
AS $$
BEGIN
  RAISE EXCEPTION 'Comparison NUMERIC-AGG_TOKEN not implemented, should be replaced by ProvSQL behavior';
END;
$$;
 
/** @brief SQL operator AGG_TOKEN < NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR < (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = >,
  NEGATOR    = >=
);
/** @brief SQL operator NUMERIC < AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR < (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = >,
  NEGATOR    = >=
);
 
/** @brief SQL operator AGG_TOKEN <= NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <= (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = >=,
  NEGATOR    = >
);
/** @brief SQL operator NUMERIC <= AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <= (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = >=,
  NEGATOR    = >
);
 
/** @brief SQL operator AGG_TOKEN = NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR = (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = =,
  NEGATOR    = <>
);
/** @brief SQL operator NUMERIC = AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR = (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = =,
  NEGATOR    = <>
);
 
/** @brief SQL operator AGG_TOKEN <> NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <> (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = <>,
  NEGATOR    = =
);
/** @brief SQL operator NUMERIC <> AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <> (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = <>,
  NEGATOR    = =
);
 
/** @brief SQL operator AGG_TOKEN >= NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR >= (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = <=,
  NEGATOR    = <
);
/** @brief SQL operator NUMERIC >= AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR >= (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = <=,
  NEGATOR    = <
);
 
/** @brief SQL operator AGG_TOKEN > NUMERIC (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR > (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = NUMERIC,
  PROCEDURE  = agg_token_comp_numeric,
  COMMUTATOR = <,
  NEGATOR    = <=
);
/** @brief SQL operator NUMERIC > AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR > (
  LEFTARG    = NUMERIC,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = numeric_comp_agg_token,
  COMMUTATOR = <,
  NEGATOR    = <=
);
 
/**
 * @brief Placeholder comparison of two AGG_TOKEN values (the diagonal)
 *
 * Never actually called; lets the parser accept AGG_TOKEN \<op\> AGG_TOKEN
 * (e.g. sum(x) > sum(y) on materialised tokens), which the ProvSQL
 * rewriter lowers to a gate_cmp at plan time.  Declaring this diagonal
 * also disambiguates `s = s2` (otherwise "operator is not unique",
 * because both AGG_TOKEN -> UUID and AGG_TOKEN -> NUMERIC casts apply).
 */
CREATE OR REPLACE FUNCTION agg_token_comp_agg_token(a AGG_TOKEN, b AGG_TOKEN)
RETURNS BOOLEAN
LANGUAGE plpgsql
IMMUTABLE STRICT PARALLEL SAFE
AS $$
BEGIN
  RAISE EXCEPTION 'Comparison AGG_TOKEN-AGG_TOKEN not implemented, should be replaced by ProvSQL behavior';
END;
$$;
 
/** @brief SQL operator AGG_TOKEN < AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR < (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = >, NEGATOR = >=
);
/** @brief SQL operator AGG_TOKEN <= AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR <= (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = >=, NEGATOR = >
);
/** @brief SQL operator AGG_TOKEN > AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR > (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = <, NEGATOR = <=
);
/** @brief SQL operator AGG_TOKEN >= AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR >= (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = <=, NEGATOR = <
);
/** @brief SQL operator AGG_TOKEN = AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR = (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = =, NEGATOR = <>
);
/** @brief SQL operator AGG_TOKEN <> AGG_TOKEN (placeholder rewritten at plan time) */
CREATE OPERATOR <> (
  LEFTARG=AGG_TOKEN, RIGHTARG=AGG_TOKEN, PROCEDURE=agg_token_comp_agg_token,
  COMMUTATOR = <>, NEGATOR = =
);
 
/**
 * @brief Placeholder comparison of AGG_TOKEN with TEXT
 *
 * This function is never actually called; it exists so the SQL parser
 * accepts comparison operators between AGG_TOKEN and TEXT values.
 * The ProvSQL query rewriter replaces these comparisons at plan time.
 */
CREATE OR REPLACE FUNCTION agg_token_comp_text(a AGG_TOKEN, b TEXT)
RETURNS BOOLEAN
LANGUAGE plpgsql
IMMUTABLE STRICT PARALLEL SAFE
AS $$
BEGIN
  RAISE EXCEPTION 'Comparison AGG_TOKEN-TEXT not implemented, should be replaced by ProvSQL behavior';
END;
$$;
 
/**
 * @brief Placeholder comparison of TEXT with AGG_TOKEN
 *
 * Symmetric to agg_token_comp_text; never actually called.
 * The ProvSQL query rewriter replaces these comparisons at plan time.
 */
CREATE OR REPLACE FUNCTION text_comp_agg_token(a TEXT, b AGG_TOKEN)
RETURNS BOOLEAN
LANGUAGE plpgsql
IMMUTABLE STRICT PARALLEL SAFE
AS $$
BEGIN
  RAISE EXCEPTION 'Comparison TEXT-AGG_TOKEN not implemented, should be replaced by ProvSQL behavior';
END;
$$;
 
/** @brief SQL operator AGG_TOKEN = TEXT (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR = (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = TEXT,
  PROCEDURE  = agg_token_comp_text,
  COMMUTATOR = =,
  NEGATOR    = <>
);
/** @brief SQL operator TEXT = AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR = (
  LEFTARG    = TEXT,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = text_comp_agg_token,
  COMMUTATOR = =,
  NEGATOR    = <>
);
 
/** @brief SQL operator AGG_TOKEN <> TEXT (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <> (
  LEFTARG    = AGG_TOKEN,
  RIGHTARG   = TEXT,
  PROCEDURE  = agg_token_comp_text,
  COMMUTATOR = <>,
  NEGATOR    = =
);
/** @brief SQL operator TEXT <> AGG_TOKEN (placeholder rewritten by ProvSQL at plan time) */
CREATE OPERATOR <> (
  LEFTARG    = TEXT,
  RIGHTARG   = AGG_TOKEN,
  PROCEDURE  = text_comp_agg_token,
  COMMUTATOR = <>,
  NEGATOR    = =
);
 
/** @} */
 
/** @defgroup random_variable_type Type for continuous random variables
 *
 *  Custom type <tt>random_variable</tt>: a thin wrapper around a
 *  provenance gate UUID, used to expose continuous probabilistic
 *  c-tables in SQL.  The UUID indexes either a <tt>gate_rv</tt>
 *  (an actual distribution) or a <tt>gate_value</tt> (a
 *  zero-variance constant produced by <tt>provsql.as_random</tt>).
 *  Binary-coercible with <tt>UUID</tt> (same 16-byte layout), so an
 *  <tt>rv</tt>-typed expression flows directly into any function
 *  expecting a UUID at zero runtime cost.
 *
 *  Constructors live in this group: <tt>provsql.normal(μ, σ)</tt>,
 *  <tt>provsql.uniform(a, b)</tt>, <tt>provsql.exponential(λ)</tt>,
 *  <tt>provsql.erlang(k, λ)</tt>, <tt>provsql.gamma(k, λ)</tt>,
 *  <tt>provsql.chi_squared(k)</tt>, <tt>provsql.lognormal(μ, σ)</tt>,
 *  <tt>provsql.weibull(k, λ)</tt>, <tt>provsql.pareto(xₘ, α)</tt>,
 *  <tt>provsql.beta(α, β)</tt>,
 *  the discrete count constructors (<tt>provsql.poisson(λ)</tt>,
 *  <tt>provsql.binomial(n, p)</tt>, <tt>provsql.geometric(p)</tt>,
 *  <tt>provsql.hypergeometric(N, K, n)</tt>,
 *  <tt>provsql.negative_binomial(r, p)</tt>, all lowering to
 *  @c categorical via @c categorical_from_log_pmf),
 *  and <tt>provsql.as_random(c)</tt>.
 *  Operator overloads
 *  (<tt>+ - * /</tt> and the six comparators) are defined further
 *  below, alongside direct <tt>rv_cmp_*</tt> UUID constructors for
 *  callers that want a <tt>gate_cmp</tt> token without going through
 *  the planner hook.
 *  @{
 */
 
CREATE TYPE random_variable;
 
/** @brief Input function for the random_variable type */
CREATE OR REPLACE FUNCTION random_variable_in(CSTRING)
  RETURNS random_variable
  AS 'provsql','random_variable_in' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Output function for the random_variable type */
CREATE OR REPLACE FUNCTION random_variable_out(random_variable)
  RETURNS CSTRING
  AS 'provsql','random_variable_out' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE TYPE random_variable (
  internallength = 16,
  input  = random_variable_in,
  output = random_variable_out,
  alignment = char
);
 
/** @brief Build a random_variable from a UUID (internal). */
CREATE OR REPLACE FUNCTION random_variable_make(tok UUID)
  RETURNS random_variable
  AS 'provsql','random_variable_make' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Binary-coercible cast random_variable -> UUID.
 *  A random_variable is byte-for-byte a pg_uuid_t (alignment char,
 *  length 16), so WITHOUT FUNCTION lets PostgreSQL reinterpret the
 *  bytes at zero runtime cost.  The cast is ASSIGNMENT (not IMPLICIT):
 *  an implicit cross-domain cast would silently reroute a comparison
 *  such as `v < w` to `UUID < UUID` (raw byte comparison) whenever
 *  `provsql` is not in search_path, since operators are resolved
 *  through search_path but casts are not.  Demoting to ASSIGNMENT
 *  turns that silent wrong result into a clean parse error.  Passing a
 *  random_variable to a UUID-taking function now needs an explicit
 *  `v::UUID` (function resolution never applies assignment casts). */
CREATE CAST (random_variable AS UUID) WITHOUT FUNCTION AS ASSIGNMENT;
CREATE CAST (UUID AS random_variable) WITHOUT FUNCTION;
 
/**
 * @brief Coerce an @c AGG_TOKEN to a @c random_variable (its circuit token).
 *
 * An aggregate over probabilistic tuples IS a random variable: its
 * @c AGG_TOKEN carries the provenance circuit of the aggregate distribution.
 * Exposing that as a @c random_variable lets a comparison / conditioning
 * predicate mix the two -- e.g. conditioning a latent leaf on a count,
 * @c "R | (poisson(lambda) = C)" with @c C a @c count(*) AGG_TOKEN -- resolve
 * to the ordinary @c random_variable comparison operators (which the planner
 * hook rewrites into a @c gate_cmp).  IMPLICIT so the mixed comparison
 * type-checks without an explicit cast; the polymorphic dispatchers keep
 * their exact @c AGG_TOKEN overloads (an exact match beats the cast).
 */
CREATE OR REPLACE FUNCTION agg_token_to_random_variable(a AGG_TOKEN)
  RETURNS random_variable AS
$$ SELECT provsql.random_variable_make(provsql.agg_token_uuid($1)); $$
  LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (AGG_TOKEN AS random_variable)
  WITH FUNCTION agg_token_to_random_variable(AGG_TOKEN) AS IMPLICIT;
 
/**
 * @brief Internal: true iff @p x is a finite (non-NaN, non-±∞) float8.
 *
 * PostgreSQL's <tt>isnan</tt> is defined for <tt>NUMERIC</tt> only,
 * not for <tt>double precision</tt>; we use the inequality form,
 * which works because PG defines <tt>NaN = NaN</tt> as <tt>TRUE</tt>
 * for floats (so <tt>NaN <> 'NaN'::float8</tt> is <tt>FALSE</tt>).
 */
CREATE OR REPLACE FUNCTION is_finite_float8(x double precision)
  RETURNS BOOL AS
$$
  SELECT $1 <> 'NaN'::float8 AND $1 <> 'Infinity'::float8 AND $1 <> '-Infinity'::float8;
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/*
 * Latent (token-valued) distribution parameters.
 *
 * A distribution parameter may be a scalar provenance token -- another
 * random_variable (or an AGG_TOKEN cast to UUID) -- rather than a
 * concrete double.  The parameter is then a random variable itself,
 * making the leaf a compound (hierarchical) distribution: e.g.
 * normal(M, 1) with M ~ normal(0, 10).  The token constructors below
 * wire such parameters as children of the gate_rv, encoding each wired
 * slot as "$i" in the extra TEXT (a literal slot keeps its decimal TEXT,
 * so an all-literal call is byte-identical to the plain NUMERIC
 * constructor).  Only the Monte Carlo sampler resolves the wires (per
 * iteration); every analytic path recognises the wired form and falls
 * through to MC.
 */
 
/**
 * @brief Internal: build a two-parameter latent @c gate_rv.
 *
 * Each parameter is supplied as EITHER a token (@p pN_tok, a scalar
 * gate @c UUID) OR a literal (@p pN_lit); exactly one is non-NULL per
 * parameter.  Token parameters are appended to the gate's wire vector
 * in order and referenced as @c "$i" in the @c extra TEXT; literal
 * parameters keep their decimal TEXT.  Not @c STRICT: the NULLs are the
 * literal-vs-token sentinels.
 */
CREATE OR REPLACE FUNCTION rv_parametric2(
    family TEXT,
    p1_tok UUID, p1_lit double precision,
    p2_tok UUID, p2_lit double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
  wires UUID[] := ARRAY[]::UUID[];
  s1 TEXT;
  s2 TEXT;
BEGIN
  IF p1_tok IS NOT NULL THEN
    wires := wires || p1_tok;
    s1 := '$' || (array_length(wires, 1) - 1);
  ELSE
    IF NOT provsql.is_finite_float8(p1_lit) THEN
      RAISE EXCEPTION 'provsql.%: literal parameter must be finite (got %)',
        family, p1_lit;
    END IF;
    s1 := p1_lit::TEXT;
  END IF;
  IF p2_tok IS NOT NULL THEN
    wires := wires || p2_tok;
    s2 := '$' || (array_length(wires, 1) - 1);
  ELSE
    IF NOT provsql.is_finite_float8(p2_lit) THEN
      RAISE EXCEPTION 'provsql.%: literal parameter must be finite (got %)',
        family, p2_lit;
    END IF;
    s2 := p2_lit::TEXT;
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv', wires);
  PERFORM provsql.set_extra(token, family || ':' || s1 || ',' || s2);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql VOLATILE PARALLEL SAFE;
 
/**
 * @brief Internal: build a one-parameter latent @c gate_rv (rate/scale).
 */
CREATE OR REPLACE FUNCTION rv_parametric1(family TEXT, p_tok UUID)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv', ARRAY[p_tok]);
  PERFORM provsql.set_extra(token, family || ':$0');
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/*
 * Token-accepting constructor overloads.  For each NUMERIC family the
 * three mixed-arity forms (token, literal), (literal, token), (token,
 * token) let any parameter be a random_variable; an AGG_TOKEN parameter
 * is passed as @c (agg)::UUID::random_variable.  The all-literal call
 * still resolves to the plain NUMERIC constructor (an exact match beats
 * the implicit NUMERIC->random_variable cast), so the literal fast path
 * is unchanged.  STRICT: a NULL parameter yields a NULL random_variable.
 */
 
-- normal(mu, sigma)
CREATE OR REPLACE FUNCTION normal(mu random_variable, sigma double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('normal', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION normal(mu double precision, sigma random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('normal', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION normal(mu random_variable, sigma random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('normal', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- logistic(mu, s)
CREATE OR REPLACE FUNCTION logistic(mu random_variable, s double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('logistic', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION logistic(mu double precision, s random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('logistic', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION logistic(mu random_variable, s random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('logistic', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- uniform(a, b)
CREATE OR REPLACE FUNCTION uniform(a random_variable, b double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('uniform', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION uniform(a double precision, b random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('uniform', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION uniform(a random_variable, b random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('uniform', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- exponential(lambda)
CREATE OR REPLACE FUNCTION exponential(lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric1('exponential', ($1)::UUID); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- gamma(k, lambda)
CREATE OR REPLACE FUNCTION gamma(k random_variable, lambda double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('gamma', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION gamma(k double precision, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('gamma', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION gamma(k random_variable, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('gamma', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- lognormal(mu, sigma)
CREATE OR REPLACE FUNCTION lognormal(mu random_variable, sigma double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('lognormal', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION lognormal(mu double precision, sigma random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('lognormal', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION lognormal(mu random_variable, sigma random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('lognormal', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- weibull(k, lambda)
CREATE OR REPLACE FUNCTION weibull(k random_variable, lambda double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('weibull', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION weibull(k double precision, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('weibull', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION weibull(k random_variable, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('weibull', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- pareto(xm, alpha)
CREATE OR REPLACE FUNCTION pareto(xm random_variable, alpha double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('pareto', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION pareto(xm double precision, alpha random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('pareto', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION pareto(xm random_variable, alpha random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('pareto', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- beta(alpha, beta)
CREATE OR REPLACE FUNCTION beta(alpha random_variable, beta double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('beta', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION beta(alpha double precision, beta random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('beta', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION beta(alpha random_variable, beta random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('beta', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- inverse_gamma(alpha, beta)
CREATE OR REPLACE FUNCTION inverse_gamma(alpha random_variable, beta double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gamma', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION inverse_gamma(alpha double precision, beta random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gamma', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION inverse_gamma(alpha random_variable, beta random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gamma', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
-- inverse_gaussian(mu, lambda)
CREATE OR REPLACE FUNCTION inverse_gaussian(mu random_variable, lambda double precision)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gaussian', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION inverse_gaussian(mu double precision, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gaussian', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION inverse_gaussian(mu random_variable, lambda random_variable)
  RETURNS random_variable AS $$ SELECT provsql.rv_parametric2('inverse_gaussian', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a normal-distribution random variable
 *
 * Creates a fresh <tt>gate_rv</tt> with @c "normal:μ,σ" stored in
 * the gate's @c extra field, and returns a <tt>random_variable</tt>
 * pointing at it.
 *
 * Validation:
 * - @p mu and @p sigma must be finite (no @c NaN, no @c ±Infinity).
 * - @p sigma must be non-negative.
 * - When @p sigma is zero the distribution degenerates to the Dirac
 *   at @p mu; the call is silently routed through @c as_random(mu),
 *   producing a @c gate_value rather than a zero-variance @c gate_rv.
 *   This keeps the sampler / moment / boundcheck paths free of σ=0
 *   special cases and lets <tt>normal(x, 0)</tt> share its gate with
 *   <tt>as_random(x)</tt>.
 *
 * @warning The <tt>VOLATILE</tt> marking is load-bearing and must
 * not be weakened.  Each call mints a fresh <tt>uuid_generate_v4</tt>
 * token because two calls to <tt>normal(0, 1)</tt> are *independent*
 * random variables; if PostgreSQL were allowed to fold the function
 * (which it would under <tt>STABLE</tt> / <tt>IMMUTABLE</tt>), two
 * calls in the same query would share a UUID and collapse into a
 * single dependent RV, silently breaking the c-table semantics.
 * Same warning applies to @c uniform and @c exponential below.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Normal_distribution">Wikipedia: Normal distribution</a>
 */
CREATE OR REPLACE FUNCTION normal(mu double precision, sigma double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(mu) OR NOT provsql.is_finite_float8(sigma) THEN
    RAISE EXCEPTION 'provsql.normal: parameters must be finite (got mu=%, sigma=%)', mu, sigma;
  END IF;
  IF sigma < 0 THEN
    RAISE EXCEPTION 'provsql.normal: sigma must be non-negative (got %)', sigma;
  END IF;
  IF sigma = 0 THEN
    RETURN provsql.as_random(mu);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'normal:' || mu || ',' || sigma);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a logistic-distribution random variable Logistic(μ, s)
 *
 * The location-scale family whose CDF is the logistic sigmoid; a threshold
 * event over a Logistic(0, 1) noise realises the logit link exactly
 * (@c P(eps < score) = 1/(1 + exp(-score))), the natural link for a
 * log-odds / latent-utility selection model.
 *
 * Validation:
 * - @p mu and @p s must be finite.
 * - @p s (the scale) must be non-negative; <tt>s = 0</tt> is the Dirac at
 *   @p mu, routed through @c as_random(mu) as with @c normal's sigma = 0.
 *
 * @param mu  location (the mean and median).
 * @param s   scale (> 0); the variance is @f$\pi^2 s^2 / 3@f$.
 * @return    a @c random_variable token for Logistic(μ, s).
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Logistic_distribution">Wikipedia: Logistic distribution</a>
 */
CREATE OR REPLACE FUNCTION logistic(mu double precision, s double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(mu) OR NOT provsql.is_finite_float8(s) THEN
    RAISE EXCEPTION 'provsql.logistic: parameters must be finite (got mu=%, s=%)', mu, s;
  END IF;
  IF s < 0 THEN
    RAISE EXCEPTION 'provsql.logistic: scale s must be non-negative (got %)', s;
  END IF;
  IF s = 0 THEN
    RETURN provsql.as_random(mu);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'logistic:' || mu || ',' || s);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a uniform-distribution random variable on [a, b]
 *
 * Validation:
 * - @p a and @p b must be finite.
 * - @p a must be ≤ @p b (reversed bounds are rejected).
 * - When <tt>a = b</tt> the distribution is the Dirac at @p a; the
 *   call is silently routed through @c as_random(a) for the same
 *   reason as @c normal with @p sigma = 0.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Continuous_uniform_distribution">Wikipedia: Continuous uniform distribution</a>
 */
CREATE OR REPLACE FUNCTION uniform(a double precision, b double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(a) OR NOT provsql.is_finite_float8(b) THEN
    RAISE EXCEPTION 'provsql.uniform: bounds must be finite (got a=%, b=%)', a, b;
  END IF;
  IF a > b THEN
    RAISE EXCEPTION 'provsql.uniform: a must be <= b (got a=%, b=%)', a, b;
  END IF;
  IF a = b THEN
    RETURN provsql.as_random(a);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'uniform:' || a || ',' || b);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct an exponential-distribution random variable with rate λ
 *
 * Validation:
 * - @p lambda must be finite and strictly positive.  No degenerate
 *   form exists for the exponential distribution, so there is no
 *   silent route through @c as_random.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Exponential_distribution">Wikipedia: Exponential distribution</a>
 */
CREATE OR REPLACE FUNCTION exponential(lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(lambda) THEN
    RAISE EXCEPTION 'provsql.exponential: lambda must be finite (got %)', lambda;
  END IF;
  IF lambda <= 0 THEN
    RAISE EXCEPTION 'provsql.exponential: lambda must be strictly positive (got %)', lambda;
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'exponential:' || lambda);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct an Erlang-distribution random variable, sum of
 *        @p k i.i.d. exponentials with shared rate @p lambda
 *
 * The Erlang distribution is the sum of @p k independent
 * <tt>Exp(λ)</tt> random variables (equivalently the gamma with
 * INTEGER shape).  It is the natural closure of i.i.d.
 * exponentials under addition, and is materialised here as a single
 * <tt>gate_rv</tt> so the analytic CDF and closed-form moments fire
 * directly (rather than the sampler having to draw and sum @p k
 * exponential leaves per Monte-Carlo iteration).
 *
 * Validation:
 * - @p k must be ≥ 1.  The degenerate @c k=1 case is silently routed
 *   through @c exponential so <tt>erlang(1, λ)</tt> shares its gate
 *   with <tt>exponential(λ)</tt>.
 * - @p lambda must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Erlang_distribution">Wikipedia: Erlang distribution</a>
 */
CREATE OR REPLACE FUNCTION erlang(k INTEGER, lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF k < 1 THEN
    RAISE EXCEPTION 'provsql.erlang: k must be >= 1 (got %)', k;
  END IF;
  IF NOT provsql.is_finite_float8(lambda) THEN
    RAISE EXCEPTION 'provsql.erlang: lambda must be finite (got %)', lambda;
  END IF;
  IF lambda <= 0 THEN
    RAISE EXCEPTION 'provsql.erlang: lambda must be strictly positive (got %)', lambda;
  END IF;
  IF k = 1 THEN
    RETURN provsql.exponential(lambda);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'erlang:' || k || ',' || lambda);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a gamma-distribution random variable with shape @p k
 *        (any positive real) and rate @p lambda
 *
 * The gamma distribution generalises Erlang to non-INTEGER shape; its
 * CDF is the regularised lower incomplete gamma, evaluated in closed
 * form by the analytic passes.  Sums of independent gammas with the
 * same rate fold to a single gamma in the simplifier.
 *
 * Validation:
 * - @p k must be finite and strictly positive.  An INTEGER @p k (in
 *   @c INTEGER range) is silently routed through @c erlang -- the gamma
 *   with INTEGER shape *is* Erlang -- so <tt>gamma(2, λ)</tt> shares
 *   its gate encoding and closure interplay with <tt>erlang(2, λ)</tt>.
 * - @p lambda must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Gamma_distribution">Wikipedia: Gamma distribution</a>
 */
CREATE OR REPLACE FUNCTION gamma(k double precision, lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(k) THEN
    RAISE EXCEPTION 'provsql.gamma: k must be finite (got %)', k;
  END IF;
  IF k <= 0 THEN
    RAISE EXCEPTION 'provsql.gamma: k must be strictly positive (got %)', k;
  END IF;
  IF NOT provsql.is_finite_float8(lambda) THEN
    RAISE EXCEPTION 'provsql.gamma: lambda must be finite (got %)', lambda;
  END IF;
  IF lambda <= 0 THEN
    RAISE EXCEPTION 'provsql.gamma: lambda must be strictly positive (got %)', lambda;
  END IF;
  IF k = floor(k) AND k <= 2147483647 THEN
    RETURN provsql.erlang(k::INTEGER, lambda);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'gamma:' || k || ',' || lambda);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a chi-squared random variable with @p k degrees of
 *        freedom: syntactic sugar for <tt>gamma(k/2, 1/2)</tt>
 *
 * @p k is accepted as @c double @c precision so fractional degrees of
 * freedom work; it must be finite and strictly positive.  Even degrees
 * of freedom route through @c erlang via @c gamma's INTEGER-shape rule.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Chi-squared_distribution">Wikipedia: Chi-squared distribution</a>
 */
CREATE OR REPLACE FUNCTION chi_squared(k double precision)
  RETURNS random_variable AS
$$
BEGIN
  IF NOT provsql.is_finite_float8(k) THEN
    RAISE EXCEPTION 'provsql.chi_squared: k must be finite (got %)', k;
  END IF;
  IF k <= 0 THEN
    RAISE EXCEPTION 'provsql.chi_squared: k must be strictly positive (got %)', k;
  END IF;
  RETURN provsql.gamma(k / 2, 0.5);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a log-normal random variable: @c exp of a
 *        Normal(@p mu, @p sigma), parameterised by the underlying
 *        normal (so its median is <tt>exp(mu)</tt> and its mean
 *        <tt>exp(mu + sigma^2/2)</tt>)
 *
 * The multiplicative counterpart of @c normal: products of independent
 * lognormals fold to a lognormal in the simplifier, and the
 * <tt>exp(normal(...))</tt> / <tt>ln(lognormal(...))</tt> bridges fold
 * in both directions, so log-scale models stay closed-form.
 *
 * Validation mirrors @c normal: both parameters must be finite,
 * @p sigma non-negative; the degenerate @c sigma = 0 case is silently
 * routed through @c as_random (a Dirac at <tt>exp(mu)</tt>).
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Log-normal_distribution">Wikipedia: Log-normal distribution</a>
 */
CREATE OR REPLACE FUNCTION lognormal(mu double precision, sigma double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(mu) OR NOT provsql.is_finite_float8(sigma) THEN
    RAISE EXCEPTION 'provsql.lognormal: parameters must be finite (got mu=%, sigma=%)', mu, sigma;
  END IF;
  IF sigma < 0 THEN
    RAISE EXCEPTION 'provsql.lognormal: sigma must be non-negative (got %)', sigma;
  END IF;
  IF sigma = 0 THEN
    RETURN provsql.as_random(exp(mu));
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'lognormal:' || mu || ',' || sigma);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Weibull random variable with shape @p k and
 *        scale @p lambda
 *
 * @p lambda is the SCALE (the 63.2% quantile), not a rate: @c k = 1 is
 * the exponential with rate <tt>1/lambda</tt>, and that case is
 * silently routed through @c exponential to share its gate.  The shape
 * tunes the hazard: @c k < 1 infant mortality, @c k > 1 wear-out.
 * Quantiles are exact, truncated moments are closed-form (via the
 * regularised incomplete gamma), and the min of i.i.d. Weibulls has a
 * closed-form mean (min-stability).
 *
 * Validation: both parameters must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Weibull_distribution">Wikipedia: Weibull distribution</a>
 */
CREATE OR REPLACE FUNCTION weibull(k double precision, lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(k) OR NOT provsql.is_finite_float8(lambda) THEN
    RAISE EXCEPTION 'provsql.weibull: parameters must be finite (got k=%, lambda=%)', k, lambda;
  END IF;
  IF k <= 0 OR lambda <= 0 THEN
    RAISE EXCEPTION 'provsql.weibull: parameters must be strictly positive (got k=%, lambda=%)', k, lambda;
  END IF;
  IF k = 1 THEN
    RETURN provsql.exponential(1 / lambda);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'weibull:' || k || ',' || lambda);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Pareto random variable with scale (minimum)
 *        @p xm and shape @p alpha
 *
 * The canonical heavy-tailed power law.  Raw moments are @b infinite
 * for <tt>alpha <= k</tt> and reported as <tt>Infinity</tt> (the mean
 * for <tt>alpha <= 1</tt>, the variance for <tt>alpha <= 2</tt>)
 * rather than estimated; quantiles, truncated moments, conditional
 * sampling (self-similarity: <tt>X | X > a</tt> is Pareto(a, alpha)),
 * and Pareto-vs-Pareto comparisons are all exact.
 *
 * Validation: both parameters must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Pareto_distribution">Wikipedia: Pareto distribution</a>
 */
CREATE OR REPLACE FUNCTION pareto(xm double precision, alpha double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(xm) OR NOT provsql.is_finite_float8(alpha) THEN
    RAISE EXCEPTION 'provsql.pareto: parameters must be finite (got xm=%, alpha=%)', xm, alpha;
  END IF;
  IF xm <= 0 OR alpha <= 0 THEN
    RAISE EXCEPTION 'provsql.pareto: parameters must be strictly positive (got xm=%, alpha=%)', xm, alpha;
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'pareto:' || xm || ',' || alpha);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct an inverse-gamma random variable with shape
 *        @p alpha and scale @p beta
 *
 * The distribution of <tt>1/Y</tt> for <tt>Y ~ gamma(alpha, beta)</tt>
 * (the conjugate prior for a Gaussian variance).  Its CDF is the
 * regularised upper incomplete gamma, evaluated in closed form by the
 * analytic passes; raw moments are @b infinite for <tt>alpha <= k</tt>
 * and reported as <tt>Infinity</tt> (the mean for <tt>alpha <= 1</tt>,
 * the variance for <tt>alpha <= 2</tt>) rather than estimated.  Positive
 * scalings rescale @p beta in the simplifier.
 *
 * Validation: both parameters must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Inverse-gamma_distribution">Wikipedia: Inverse-gamma distribution</a>
 */
CREATE OR REPLACE FUNCTION inverse_gamma(alpha double precision, beta double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(alpha) OR NOT provsql.is_finite_float8(beta) THEN
    RAISE EXCEPTION 'provsql.inverse_gamma: parameters must be finite (got alpha=%, beta=%)', alpha, beta;
  END IF;
  IF alpha <= 0 OR beta <= 0 THEN
    RAISE EXCEPTION 'provsql.inverse_gamma: parameters must be strictly positive (got alpha=%, beta=%)', alpha, beta;
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'inverse_gamma:' || alpha || ',' || beta);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct an inverse-Gaussian (Wald) random variable with mean
 *        @p mu and shape @p lambda
 *
 * The first-passage time of Brownian motion with drift: a positive,
 * right-skewed family.  Its CDF has a closed form in the standard normal
 * @c Phi, so comparisons and quantiles are analytic; all raw moments are
 * finite.  Positive scalings map <tt>c·IG(mu, lambda)</tt> to
 * <tt>IG(c·mu, c·lambda)</tt>, and a sum of independent inverse
 * Gaussians sharing the ratio <tt>lambda/mu²</tt> folds to a single
 * inverse Gaussian in the simplifier.  @ref wald is an alias.
 *
 * Validation: both parameters must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Inverse_Gaussian_distribution">Wikipedia: Inverse Gaussian distribution</a>
 */
CREATE OR REPLACE FUNCTION inverse_gaussian(mu double precision, lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(mu) OR NOT provsql.is_finite_float8(lambda) THEN
    RAISE EXCEPTION 'provsql.inverse_gaussian: parameters must be finite (got mu=%, lambda=%)', mu, lambda;
  END IF;
  IF mu <= 0 OR lambda <= 0 THEN
    RAISE EXCEPTION 'provsql.inverse_gaussian: parameters must be strictly positive (got mu=%, lambda=%)', mu, lambda;
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'inverse_gaussian:' || mu || ',' || lambda);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Wald distribution: alias for @ref inverse_gaussian.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Inverse_Gaussian_distribution">Wikipedia: Inverse Gaussian distribution</a>
 */
CREATE OR REPLACE FUNCTION wald(mu double precision, lambda double precision)
  RETURNS random_variable AS
$$
  SELECT provsql.inverse_gaussian(mu, lambda);
$$ LANGUAGE sql VOLATILE PARALLEL SAFE;
 
/**
 * @brief Build a discrete (categorical) random variable from outcomes
 *        and UNNORMALISED log-masses
 *
 * The shared back end of the discrete count constructors (@c poisson,
 * @c binomial, @c geometric, @c hypergeometric,
 * @c negative_binomial), and directly usable for any custom discrete
 * pmf: the log-masses are shifted by their maximum (so only relative
 * magnitudes matter and no @c exp underflows), outcomes whose relative
 * mass is below <tt>1e-15</tt> are dropped, and the rest is
 * renormalised before being handed to @c categorical.  Working in log
 * space keeps arbitrarily large parameters stable (e.g. a
 * <tt>Poisson(1000)</tt> pmf whose linear-space recurrence would
 * underflow at @c exp(-1000)).
 *
 * @param outcomes  outcome values, same length as @p log_pmf
 * @param log_pmf   natural logs of the (unnormalised) masses
 */
CREATE OR REPLACE FUNCTION categorical_from_log_pmf(
  outcomes double precision[], log_pmf double precision[])
  RETURNS random_variable AS
$$
DECLARE
  n INT := array_length(outcomes, 1);
  max_lp double precision := '-Infinity';
  kept_o double precision[] := '{}';
  kept_p double precision[] := '{}';
  total double precision := 0;
  v double precision;
  i INT;
BEGIN
  IF n IS NULL OR n = 0 OR n <> coalesce(array_length(log_pmf, 1), 0) THEN
    RAISE EXCEPTION 'provsql.categorical_from_log_pmf: outcomes and log_pmf must be non-empty arrays of the same length';
  END IF;
  FOR i IN 1..n LOOP
    IF log_pmf[i] > max_lp THEN max_lp := log_pmf[i]; END IF;
  END LOOP;
  IF max_lp = '-Infinity' THEN
    RAISE EXCEPTION 'provsql.categorical_from_log_pmf: all masses are zero';
  END IF;
  FOR i IN 1..n LOOP
    v := exp(log_pmf[i] - max_lp);
    IF v >= 1e-15 THEN
      kept_o := array_append(kept_o, outcomes[i]);
      kept_p := array_append(kept_p, v);
      total := total + v;
    END IF;
  END LOOP;
  FOR i IN 1..array_length(kept_p, 1) LOOP
    kept_p[i] := kept_p[i] / total;
  END LOOP;
  RETURN provsql.categorical(kept_p, kept_o);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Poisson random variable with mean @p lambda, as a
 *        truncated categorical
 *
 * The pmf is enumerated over <tt>[max(0, λ-12√λ), λ+12√λ+30]</tt> (the
 * omitted tails carry ~1e-30 of mass) by the log-space recurrence
 * <tt>ln p(k+1) = ln p(k) + ln λ - ln(k+1)</tt> and handed to
 * @c categorical_from_log_pmf, so moments, quantiles, and (in)equality
 * comparisons are exact over the enumerated support.  @c lambda = 0 is
 * a Dirac at @c 0 (routed through @c as_random); supports up to 10000
 * outcomes (λ up to ~170000), beyond which it raises -- approximate
 * huge means by @c normal(λ, √λ) instead.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Poisson_distribution">Wikipedia: Poisson distribution</a>
 */
CREATE OR REPLACE FUNCTION poisson(lambda double precision)
  RETURNS random_variable AS
$$
DECLARE
  lo INT;
  hi INT;
  outcomes double precision[] := '{}';
  lps double precision[] := '{}';
  lp double precision := 0;
  k INT;
BEGIN
  IF NOT provsql.is_finite_float8(lambda) OR lambda < 0 THEN
    RAISE EXCEPTION 'provsql.poisson: lambda must be finite and non-negative (got %)', lambda;
  END IF;
  IF lambda = 0 THEN
    RETURN provsql.as_random(0);
  END IF;
  lo := greatest(0, floor(lambda - 12 * sqrt(lambda)))::INT;
  hi := ceil(lambda + 12 * sqrt(lambda))::INT + 30;
  IF hi - lo + 1 > 10000 THEN
    RAISE EXCEPTION 'provsql.poisson: support window of % outcomes exceeds 10000; approximate with normal(%, sqrt(%))', hi - lo + 1, lambda, lambda;
  END IF;
  -- ln p(0) = -λ; walk the recurrence, keeping only the window.
  lp := -lambda;
  FOR k IN 1..hi LOOP
    lp := lp + ln(lambda) - ln(k::double precision);
    IF k >= lo THEN
      outcomes := array_append(outcomes, k::double precision);
      lps := array_append(lps, lp);
    END IF;
  END LOOP;
  IF lo = 0 THEN
    outcomes := array_prepend(0::double precision, outcomes);
    lps := array_prepend(-lambda, lps);
  END IF;
  RETURN provsql.categorical_from_log_pmf(outcomes, lps);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Poisson with a LATENT rate: @c poisson(random_variable).
 *
 * A latent (token-valued) rate cannot be enumerated into a categorical at
 * construction, so this builds a parametric @c gate_rv leaf (family
 * @c "poisson") wiring the rate, exactly like the continuous latent
 * constructors.  Only the Monte Carlo sampler resolves the rate (per draw,
 * then draws a Poisson); @c observe weights by the Poisson pmf; the mean is
 * exact (E[Poisson(Λ)] = E[Λ], affine).  Unblocks discrete-likelihood
 * posteriors such as @c "R | (poisson(120*R) = observed_count)".
 */
CREATE OR REPLACE FUNCTION poisson(lambda random_variable)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric1('poisson', ($1)::UUID); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Beta(α, β) random variable on the unit interval
 *
 * The conjugate prior of Bernoulli / binomial success probabilities:
 * closed-form moments, CDF through the regularised incomplete beta,
 * quantiles through the generic CDF bisection over the finite
 * @c [0, 1] support, and closed-form truncated moments (interval
 * conditioning).  <tt>Beta(1, 1)</tt> IS <tt>Uniform(0, 1)</tt> and is
 * silently routed through @c uniform to share its richer closed forms.
 *
 * Validation: both shapes must be finite and strictly positive.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing; see the warning on
 * @ref normal.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Beta_distribution">Wikipedia: Beta distribution</a>
 */
CREATE OR REPLACE FUNCTION beta(alpha double precision, beta double precision)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF NOT provsql.is_finite_float8(alpha) OR NOT provsql.is_finite_float8(beta) THEN
    RAISE EXCEPTION 'provsql.beta: parameters must be finite (got alpha=%, beta=%)', alpha, beta;
  END IF;
  IF alpha <= 0 OR beta <= 0 THEN
    RAISE EXCEPTION 'provsql.beta: parameters must be strictly positive (got alpha=%, beta=%)', alpha, beta;
  END IF;
  IF alpha = 1 AND beta = 1 THEN
    RETURN provsql.uniform(0, 1);
  END IF;
  token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(token, 'rv');
  PERFORM provsql.set_extra(token, 'beta:' || alpha || ',' || beta);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Binomial(n, p) random variable (number of
 *        successes in @p n independent trials), as a categorical
 *
 * Enumerated over <tt>{0..n}</tt> by the log-space recurrence
 * <tt>ln p(k+1) = ln p(k) + ln((n-k)/(k+1)) + ln(p/(1-p))</tt>
 * (outcomes below 1e-15 relative mass are dropped).  @c p = 0 /
 * @c p = 1 are Diracs at @c 0 / @c n; @c n is capped at 10000.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Binomial_distribution">Wikipedia: Binomial distribution</a>
 */
CREATE OR REPLACE FUNCTION binomial(n INTEGER, p double precision)
  RETURNS random_variable AS
$$
DECLARE
  outcomes double precision[] := '{}';
  lps double precision[] := '{}';
  lp double precision;
  k INT;
BEGIN
  IF n IS NULL OR n < 0 THEN
    RAISE EXCEPTION 'provsql.binomial: n must be non-negative (got %)', n;
  END IF;
  IF NOT provsql.is_finite_float8(p) OR p < 0 OR p > 1 THEN
    RAISE EXCEPTION 'provsql.binomial: p must be in [0, 1] (got %)', p;
  END IF;
  IF n > 10000 THEN
    RAISE EXCEPTION 'provsql.binomial: n = % exceeds 10000; approximate with normal(n*p, sqrt(n*p*(1-p)))', n;
  END IF;
  IF n = 0 OR p = 0 THEN
    RETURN provsql.as_random(0);
  END IF;
  IF p = 1 THEN
    RETURN provsql.as_random(n);
  END IF;
  lp := n * ln(1 - p);   -- ln p(0)
  outcomes := array_append(outcomes, 0::double precision);
  lps := array_append(lps, lp);
  FOR k IN 0..(n - 1) LOOP
    lp := lp + ln((n - k)::double precision / (k + 1)) + ln(p / (1 - p));
    outcomes := array_append(outcomes, (k + 1)::double precision);
    lps := array_append(lps, lp);
  END LOOP;
  RETURN provsql.categorical_from_log_pmf(outcomes, lps);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Binomial with a fixed trial count and a LATENT success
 *        probability: @c binomial(INTEGER, random_variable).
 *
 * @p n is a literal trial count; @p p is a latent (token-valued) success
 * probability (e.g. @c "40.0 / N" for a latent population size @c N).
 * Builds a parametric @c gate_rv leaf (family @c "binomial", @c extra
 * @c "binomial:n,$0") the Monte Carlo sampler resolves per draw; @c observe
 * weights by the Binomial pmf.  Unblocks capture-recapture-style posteriors
 * such as @c "N | (binomial(50, 40.0/N) = recaptured_count)".
 */
CREATE OR REPLACE FUNCTION binomial(n INTEGER, p random_variable)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric2('binomial', NULL, $1::double precision,
                                 ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Geometric(p) random variable -- the number of
 *        TRIALS up to and including the first success (support
 *        starting at 1; subtract 1 for the failures convention)
 *
 * <tt>P(X = k) = (1-p)^{k-1} p</tt>, enumerated up to the 1e-15
 * relative-mass tail and renormalised.  @c p = 1 is a Dirac at @c 1.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Geometric_distribution">Wikipedia: Geometric distribution</a>
 */
CREATE OR REPLACE FUNCTION geometric(p double precision)
  RETURNS random_variable AS
$$
DECLARE
  k_max INT;
  outcomes double precision[] := '{}';
  lps double precision[] := '{}';
  k INT;
BEGIN
  IF NOT provsql.is_finite_float8(p) OR p <= 0 OR p > 1 THEN
    RAISE EXCEPTION 'provsql.geometric: p must be in (0, 1] (got %)', p;
  END IF;
  IF p = 1 THEN
    RETURN provsql.as_random(1);
  END IF;
  k_max := 1 + ceil(ln(1e-15) / ln(1 - p))::INT;
  IF k_max > 10000 THEN
    RAISE EXCEPTION 'provsql.geometric: support window of % outcomes exceeds 10000 (p = % is too small); approximate with exponential(%)', k_max, p, p;
  END IF;
  FOR k IN 1..k_max LOOP
    outcomes := array_append(outcomes, k::double precision);
    lps := array_append(lps, (k - 1) * ln(1 - p) + ln(p));
  END LOOP;
  RETURN provsql.categorical_from_log_pmf(outcomes, lps);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Geometric with a LATENT success probability: @c geometric(random_variable).
 *
 * A latent (token-valued) @p p cannot be enumerated at construction, so this
 * builds a parametric @c gate_rv leaf (family @c "geometric") wiring the
 * probability, resolved per draw by the sampler.  @c observe weights by the
 * geometric pmf; unblocks a Beta-Geometric conjugate posterior.
 */
CREATE OR REPLACE FUNCTION geometric(p random_variable)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric1('geometric', ($1)::UUID); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a Hypergeometric(N, K, n) random variable: the
 *        number of marked items among @p n draws WITHOUT replacement
 *        from a population of @p pop_n items of which @p k_marked are
 *        marked
 *
 * The exact finite support <tt>[max(0, n-(N-K)), min(n, K)]</tt> is
 * enumerated by the pmf ratio recurrence (in log space, so large
 * populations cannot overflow) and normalised -- exact "sampling
 * without replacement" probabilities with no combinatorial functions
 * needed.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Hypergeometric_distribution">Wikipedia: Hypergeometric distribution</a>
 */
CREATE OR REPLACE FUNCTION hypergeometric(pop_n INTEGER, k_marked INTEGER, n INTEGER)
  RETURNS random_variable AS
$$
DECLARE
  lo INT;
  hi INT;
  outcomes double precision[] := '{}';
  lps double precision[] := '{}';
  lp double precision := 0;   -- relative log-mass; normalised later
  k INT;
BEGIN
  IF pop_n IS NULL OR k_marked IS NULL OR n IS NULL
     OR pop_n < 0 OR k_marked < 0 OR n < 0
     OR k_marked > pop_n OR n > pop_n THEN
    RAISE EXCEPTION 'provsql.hypergeometric: need 0 <= k_marked, n <= pop_n (got pop_n=%, k_marked=%, n=%)', pop_n, k_marked, n;
  END IF;
  lo := greatest(0, n - (pop_n - k_marked));
  hi := least(n, k_marked);
  IF hi - lo + 1 > 10000 THEN
    RAISE EXCEPTION 'provsql.hypergeometric: support window of % outcomes exceeds 10000', hi - lo + 1;
  END IF;
  outcomes := array_append(outcomes, lo::double precision);
  lps := array_append(lps, lp);
  FOR k IN lo..(hi - 1) LOOP
    -- pmf(k+1)/pmf(k) = (K-k)(n-k) / ((k+1)(N-K-n+k+1))
    lp := lp + ln((k_marked - k)::double precision * (n - k))
             - ln((k + 1)::double precision * (pop_n - k_marked - n + k + 1));
    outcomes := array_append(outcomes, (k + 1)::double precision);
    lps := array_append(lps, lp);
  END LOOP;
  RETURN provsql.categorical_from_log_pmf(outcomes, lps);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Construct a negative-binomial random variable: the number of
 *        FAILURES before the @p r-th success (support starting at 0),
 *        with real @p r > 0 allowed (the Polya / overdispersed-count
 *        parameterisation, the Poisson-Gamma mixture)
 *
 * <tt>P(X = k) = C(k+r-1, k) p^r (1-p)^k</tt>, enumerated by the
 * log-space recurrence
 * <tt>ln p(k+1) = ln p(k) + ln((k+r)/(k+1)) + ln(1-p)</tt> up to the
 * 1e-15 relative-mass tail.  @c p = 1 is a Dirac at @c 0.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Negative_binomial_distribution">Wikipedia: Negative binomial distribution</a>
 */
CREATE OR REPLACE FUNCTION negative_binomial(r double precision, p double precision)
  RETURNS random_variable AS
$$
DECLARE
  outcomes double precision[] := '{}';
  lps double precision[] := '{}';
  lp double precision;
  max_lp double precision;
  mean double precision;
  k INT := 0;
BEGIN
  IF NOT provsql.is_finite_float8(r) OR r <= 0 THEN
    RAISE EXCEPTION 'provsql.negative_binomial: r must be finite and strictly positive (got %)', r;
  END IF;
  IF NOT provsql.is_finite_float8(p) OR p <= 0 OR p > 1 THEN
    RAISE EXCEPTION 'provsql.negative_binomial: p must be in (0, 1] (got %)', p;
  END IF;
  IF p = 1 THEN
    RETURN provsql.as_random(0);
  END IF;
  mean := r * (1 - p) / p;
  lp := r * ln(p);   -- ln p(0)
  max_lp := lp;
  outcomes := array_append(outcomes, 0::double precision);
  lps := array_append(lps, lp);
  LOOP
    lp := lp + ln((k + r) / (k + 1)) + ln(1 - p);
    k := k + 1;
    IF lp > max_lp THEN max_lp := lp; END IF;
    outcomes := array_append(outcomes, k::double precision);
    lps := array_append(lps, lp);
    EXIT WHEN k > mean AND lp < max_lp + ln(1e-15);
    IF k >= 10000 THEN
      RAISE EXCEPTION 'provsql.negative_binomial: support window exceeds 10000 outcomes (r=%, p=%)', r, p;
    END IF;
  END LOOP;
  RETURN provsql.categorical_from_log_pmf(outcomes, lps);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/*
 * NegativeBinomial with a LATENT parameter: the count r (number of successes)
 * stays a plain number while the success probability p is a random_variable,
 * built as a parametric gate_rv leaf (family "negative_binomial") -- the
 * Beta-NegativeBinomial conjugate shape.  The all-random and latent-r forms
 * are provided for uniformity with the continuous constructors.
 */
CREATE OR REPLACE FUNCTION negative_binomial(r double precision, p random_variable)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric2('negative_binomial', NULL, $1, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION negative_binomial(r random_variable, p double precision)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric2('negative_binomial', ($1)::UUID, NULL, NULL, $2); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
CREATE OR REPLACE FUNCTION negative_binomial(r random_variable, p random_variable)
  RETURNS random_variable AS
$$ SELECT provsql.rv_parametric2('negative_binomial', ($1)::UUID, NULL, ($2)::UUID, NULL); $$
  LANGUAGE sql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Catalog of the registered continuous-distribution families.
 *
 * One row per @c gate_rv family known to this build of the extension:
 * @c name is the on-disk token (the part before the colon in the gate's
 * @c extra encoding), @c nparams the parameter count, @c param_names the
 * conventional parameter symbols in @c extra order (e.g.
 * <tt>{μ, σ}</tt>), and @c label a short display glyph (e.g. @c "N",
 * @c "Γ").  UI clients (ProvSQL Studio's circuit inspector) read this to
 * render families they were not hard-coded for, so a newly added family
 * shows up without a client release.
 */
CREATE OR REPLACE FUNCTION rv_families()
  RETURNS TABLE(name TEXT, nparams INT, param_names TEXT[], label TEXT) AS
  'provsql','rv_families' LANGUAGE C STABLE PARALLEL SAFE;
 
/**
 * @brief Construct a probabilistic-mixture random variable.
 *
 * Returns a @c random_variable whose distribution is a Bernoulli
 * mixture of two scalar RV roots: with probability <tt>P(p = true)</tt>
 * the mixture samples @p x, with the complementary probability it
 * samples @p y.  The mixing token @p p is a @c gate_input Bernoulli
 * whose probability has been pinned with @c set_prob, and the same
 * @p p can be shared with other branches of the circuit -- the
 * Monte-Carlo sampler's per-iteration cache couples every reference
 * to the same draw, so users can build joint conditional structures
 * (e.g. <tt>mixture(p, X1, Y1) + mixture(p, X2, Y2)</tt> samples
 * X1 + X2 with prob π and Y1 + Y2 with prob 1-π).
 *
 * @p x and @p y may be any scalar RV root: a base @c gate_rv
 * (@c normal / @c uniform / @c exponential / @c erlang), a
 * @c gate_value Dirac (@c as_random), a @c gate_arith expression, or
 * another @c mixture.  N-ary mixtures are built by composition --
 * <tt>mixture(p1, A, mixture(p2, B, C))</tt> realises a 3-component
 * mixture with effective weights <tt>π1, (1-π1)·π2, (1-π1)·(1-π2)</tt>.
 *
 * Validation:
 * - @p p must point to a Boolean gate (@c input, @c mulinput,
 *   @c update, @c plus, @c times, @c monus, @c project, @c eq,
 *   @c cmp, @c zero, @c one).  Compound Boolean gates derive their
 *   probability from their atoms via the active probability-evaluation
 *   method; a bare @c gate_input's probability is whatever @c set_prob
 *   pinned (@c set_prob is responsible for keeping it in [0, 1]).
 * - @p x and @p y must be scalar RV roots; aggregate / Boolean roots
 *   are rejected at construction.
 *
 * Two calls to @c mixture with the same @c (p, x, y) operands collapse
 * to the same @c gate_mixture node by v5-hash, exactly like
 * @c arith(PLUS, X, Y).  Draw independence is controlled by @p p:
 * sharing @p p couples branch selection across consumers via the
 * sampler's @c bool_cache_; minting independent Bernoullis (e.g. via
 * the @c mixture(p_value, …) overload) decouples them.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Mixture_distribution">Wikipedia: Mixture distribution</a>
 */
CREATE OR REPLACE FUNCTION mixture(
  p UUID, x random_variable, y random_variable)
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
  p_kind provsql.PROVENANCE_GATE;
  x_uuid UUID;
  y_uuid UUID;
  x_kind provsql.PROVENANCE_GATE;
  y_kind provsql.PROVENANCE_GATE;
BEGIN
  p_kind := provsql.get_gate_type(p);
  IF p_kind NOT IN ('input','mulinput','update',
                    'plus','times','monus',
                    'project','eq','cmp',
                    'zero','one') THEN
    RAISE EXCEPTION 'provsql.mixture: p must be a Boolean gate '
                    '(input/mulinput/update/plus/times/monus/project/eq/cmp/zero/one), got %', p_kind;
  END IF;
 
  x_uuid := (x)::UUID;
  y_uuid := (y)::UUID;
  x_kind := provsql.get_gate_type(x_uuid);
  y_kind := provsql.get_gate_type(y_uuid);
  IF x_kind NOT IN ('rv','value','arith','mixture') THEN
    RAISE EXCEPTION 'provsql.mixture: x must be a scalar RV root (rv / value / arith / mixture), got %', x_kind;
  END IF;
  IF y_kind NOT IN ('rv','value','arith','mixture') THEN
    RAISE EXCEPTION 'provsql.mixture: y must be a scalar RV root (rv / value / arith / mixture), got %', y_kind;
  END IF;
 
  token := public.uuid_generate_v5(
    provsql.uuid_ns_provsql(),
    concat('mixture', p, x_uuid, y_uuid));
  PERFORM provsql.create_gate(token, 'mixture', ARRAY[p, x_uuid, y_uuid]);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Ad-hoc mixture constructor that mints a fresh anonymous
 *        @c gate_input Bernoulli with probability @p p_value.
 *
 * Sugar over the @c mixture(UUID, x, y) form: when the caller doesn't
 * care about reusing the Bernoulli token elsewhere in the circuit
 * (which is the common case &ndash; "give me a 0.3 / 0.7 weighted GMM,
 * I don't need to share the coin"), this overload creates the
 * underlying @c gate_input on the fly with a fresh
 * @c uuid_generate_v4() token, pins @p p_value via @c set_prob, and
 * threads everything into the UUID-keyed constructor.
 *
 * Each call mints a NEW Bernoulli, so two calls to
 * <tt>mixture(0.5, X, Y)</tt> are *independent* mixtures whose branch
 * selections are uncorrelated.  When coupling is desired (e.g. two
 * mixtures sharing a coin), use the @c mixture(UUID, x, y) form with a
 * user-managed @c gate_input token.
 *
 * @warning <tt>VOLATILE</tt> is load-bearing for the same reason as
 * @ref normal and the other RV constructors -- folding under
 * @c STABLE / @c IMMUTABLE would collapse two independent draws into
 * one shared gate.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Mixture_distribution">Wikipedia: Mixture distribution</a>
 */
CREATE OR REPLACE FUNCTION mixture(
  p_value double precision,
  x random_variable,
  y random_variable)
  RETURNS random_variable AS
$$
DECLARE
  p_token UUID;
BEGIN
  IF p_value IS NULL OR p_value <> p_value OR p_value < 0 OR p_value > 1 THEN
    RAISE EXCEPTION 'provsql.mixture: probability must be in [0,1] (got %)', p_value;
  END IF;
  p_token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(p_token, 'input');
  PERFORM provsql.set_prob(p_token, p_value);
  RETURN provsql.mixture(p_token, x, y);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Categorical-RV constructor over explicit (probabilities,
 *        values) arrays.
 *
 * Builds a categorical-form @c gate_mixture directly: a fresh
 * @c gate_input "key" anchor and one @c gate_mulinput per outcome with
 * positive mass, all sharing the key.  The wires
 * <tt>[key, mul_1, ..., mul_n]</tt> are what downstream evaluators
 * (@c Expectation, @c MonteCarloSampler, @c AnalyticEvaluator,
 * @c RangeCheck) recognise via @c isCategoricalMixture and treat as a
 * scalar RV with the categorical distribution @p probs over
 * @p outcomes.
 *
 * Validation:
 * - @p probs and @p outcomes must be non-null, same length, length &ge; 1.
 * - Each @c probs[i] must be finite, in <tt>[0, 1]</tt>, and the array
 *   must sum to 1 within @c 1e-9.
 * - Each @c outcomes[i] must be finite.
 *
 * Each call mints a fresh key gate and a fresh set of mulinputs, so
 * two calls to @c categorical with the same arrays are *independent*
 * categorical RVs.  The marking is @c VOLATILE accordingly.
 *
 * Degenerate case: a categorical with exactly one positive-mass
 * outcome reduces to @c as_random(v) at construction (the block would
 * just be a single mulinput, which is operationally a Dirac point
 * mass).  Two such calls share the @c gate_value UUID via the v5
 * convention @c as_random already uses.
 *
 * @sa @c mixture for the Bernoulli-weighted choice constructor.
 * @sa <a href="https://en.wikipedia.org/wiki/Categorical_distribution">Wikipedia: Categorical distribution</a>
 */
CREATE OR REPLACE FUNCTION categorical(
  probs    double precision[],
  outcomes double precision[])
  RETURNS random_variable AS
$$
DECLARE
  n INTEGER;
  p_sum double precision := 0.0;
  i INTEGER;
  key_token UUID;
  mix_token UUID;
  mul_token UUID;
  mul_tokens UUID[] := ARRAY[]::UUID[];
  mix_wires  UUID[];
  pi_i double precision;
  vi_i double precision;
BEGIN
  IF probs IS NULL OR outcomes IS NULL THEN
    RAISE EXCEPTION 'provsql.categorical: probs and outcomes must be non-null';
  END IF;
  n := array_length(probs, 1);
  IF n IS NULL OR n < 1 THEN
    RAISE EXCEPTION 'provsql.categorical: probs must be non-empty';
  END IF;
  IF array_length(outcomes, 1) <> n THEN
    RAISE EXCEPTION 'provsql.categorical: probs and outcomes must have the same length (got % and %)',
      n, array_length(outcomes, 1);
  END IF;
 
  FOR i IN 1..n LOOP
    pi_i := probs[i];
    vi_i := outcomes[i];
    -- PostgreSQL diverges from IEEE 754: NaN = NaN is TRUE there, so
    -- the canonical x <> x NaN test doesn't fire.  Compare against the
    -- literal 'NaN'::float8 instead, and reject ±Infinity for outcomes
    -- explicitly.
    IF pi_i IS NULL OR pi_i = 'NaN'::float8 OR pi_i < 0 OR pi_i > 1 THEN
      RAISE EXCEPTION 'provsql.categorical: probs[%] must be in [0,1] (got %)', i, pi_i;
    END IF;
    IF vi_i IS NULL OR vi_i = 'NaN'::float8
       OR vi_i = 'Infinity'::float8 OR vi_i = '-Infinity'::float8 THEN
      RAISE EXCEPTION 'provsql.categorical: outcomes[%] must be finite (got %)', i, vi_i;
    END IF;
    p_sum := p_sum + pi_i;
  END LOOP;
  IF abs(p_sum - 1.0) > 1e-9 THEN
    RAISE EXCEPTION 'provsql.categorical: probs must sum to 1 within 1e-9 (got %)', p_sum;
  END IF;
 
  -- Degenerate case: exactly one positive-mass outcome (the rest are
  -- zero).  The "categorical" is then a Dirac point mass; skip the
  -- block-allocation entirely and return @c as_random(v), which yields
  -- a shared, v5-keyed gate_value -- exactly what downstream
  -- evaluators (rv_moment, AnalyticEvaluator, rv_support) treat
  -- specially.  Saves a key gate and a mulinput per call, and lets
  -- two calls to @c categorical({1.0}, {v}) collide on the same
  -- gate_value UUID instead of producing distinct anonymous blocks.
  DECLARE
    nb_positive INTEGER := 0;
    only_idx    INTEGER := 0;
  BEGIN
    FOR i IN 1..n LOOP
      IF probs[i] > 0.0 THEN
        nb_positive := nb_positive + 1;
        only_idx := i;
      END IF;
    END LOOP;
    IF nb_positive = 1 THEN
      RETURN provsql.as_random(outcomes[only_idx]);
    END IF;
  END;
 
  -- Mint the block's key anchor.  Probability 1.0 matches the
  -- joint-table convention: the categorical mass lives on the
  -- mulinputs, the key just identifies the block.
  key_token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(key_token, 'input');
  PERFORM provsql.set_prob(key_token, 1.0);
 
  -- One mulinput per positive-probability outcome.  Zero-probability
  -- entries contribute no mass and are skipped: the gate_mixture's
  -- wire vector is otherwise polluted with no-op leaves.
  FOR i IN 1..n LOOP
    pi_i := probs[i];
    IF pi_i <= 0.0 THEN CONTINUE; END IF;
    mul_token := public.uuid_generate_v4();
    PERFORM provsql.create_gate(mul_token, 'mulinput', ARRAY[key_token]);
    PERFORM provsql.set_prob(mul_token, pi_i);
    PERFORM provsql.set_infos(mul_token, (i - 1));
    PERFORM provsql.set_extra(mul_token, outcomes[i]::TEXT);
    mul_tokens := mul_tokens || mul_token;
  END LOOP;
 
  mix_wires := ARRAY[key_token] || mul_tokens;
  mix_token := public.uuid_generate_v4();
  PERFORM provsql.create_gate(mix_token, 'mixture', mix_wires);
  RETURN provsql.random_variable_make(mix_token);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Gaussian-mixture-model (GMM) constructor.
 *
 * Packages the common fitted-density pattern -- a categorical choice
 * among Normal components -- into one call:
 *
 * @code
 * provsql.gmm(weights => ARRAY[0.3, 0.5, 0.2],
 *             means   => ARRAY[120.0, 380.0, 1200.0],
 *             stddevs => ARRAY[40.0, 90.0, 250.0])
 * @endcode
 *
 * No new gate: the mixture decomposes into a stick-breaking cascade of
 * Bernoulli @c gate_mixture nodes over @c gate_rv Normal leaves
 * (component @c i is selected with conditional probability
 * @c w_i / (w_i + ... + w_n), so the joint selection probabilities are
 * exactly @p weights), which every evaluator already handles: moments
 * are closed-form through the mixture recursion, sampling is exact,
 * and comparisons ride the existing mixture machinery.  Zero-weight
 * components are skipped; a single positive-weight component returns
 * its Normal directly (no mixture node).
 *
 * Validation mirrors @c categorical: same-length non-empty arrays,
 * weights finite in <tt>[0, 1]</tt> summing to 1 within @c 1e-9; the
 * component parameters are validated by @c provsql.normal (finite
 * @c mu, non-negative @c sigma; @c sigma @c = @c 0 degenerates to a
 * Dirac component).
 *
 * @sa @c mixture, @c categorical, @c normal
 * @sa <a href="https://en.wikipedia.org/wiki/Mixture_model">Wikipedia: Mixture model</a>
 */
CREATE OR REPLACE FUNCTION gmm(
  weights double precision[],
  means   double precision[],
  stddevs double precision[])
  RETURNS random_variable AS
$$
DECLARE
  n INTEGER;
  w_sum double precision := 0.0;
  i INTEGER;
  acc random_variable := NULL;
  remaining double precision := 0.0;
BEGIN
  IF weights IS NULL OR means IS NULL OR stddevs IS NULL THEN
    RAISE EXCEPTION 'provsql.gmm: weights, means, and stddevs must be non-null';
  END IF;
  n := array_length(weights, 1);
  IF n IS NULL OR n < 1 THEN
    RAISE EXCEPTION 'provsql.gmm: weights must be non-empty';
  END IF;
  IF array_length(means, 1) <> n OR array_length(stddevs, 1) <> n THEN
    RAISE EXCEPTION 'provsql.gmm: weights, means, and stddevs must have the same length (got %, %, %)',
      n, array_length(means, 1), array_length(stddevs, 1);
  END IF;
  FOR i IN 1..n LOOP
    IF weights[i] IS NULL OR weights[i] = 'NaN'::float8
       OR weights[i] < 0 OR weights[i] > 1 THEN
      RAISE EXCEPTION 'provsql.gmm: weights[%] must be in [0,1] (got %)',
        i, weights[i];
    END IF;
    w_sum := w_sum + weights[i];
  END LOOP;
  IF abs(w_sum - 1.0) > 1e-9 THEN
    RAISE EXCEPTION 'provsql.gmm: weights must sum to 1 within 1e-9 (got %)', w_sum;
  END IF;
 
  -- Stick-breaking, built back to front: acc holds the mixture of
  -- components i+1..n, and prepending component i selects it with
  -- conditional probability w_i / (w_i + ... + w_n).
  FOR i IN REVERSE n..1 LOOP
    IF weights[i] <= 0.0 THEN
      CONTINUE;
    END IF;
    IF acc IS NULL THEN
      acc := provsql.normal(means[i], stddevs[i]);
      remaining := weights[i];
    ELSE
      remaining := remaining + weights[i];
      acc := provsql.mixture(least(1.0, weights[i] / remaining),
                             provsql.normal(means[i], stddevs[i]), acc);
    END IF;
  END LOOP;
  RETURN acc;
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Empirical-samples constructor: the ecdf of a sample bundle as
 *        a @c random_variable.
 *
 * Loads a Monte Carlo / MCMC / bootstrap sample array as the discrete
 * distribution putting mass @c 1/n on each draw (duplicates merge, so a
 * value drawn @c k times carries @c k/n) -- the standard empirical
 * distribution.  Reduces entirely to @ref categorical, so the whole
 * exact discrete surface applies: moments are the sample moments,
 * comparisons against constants are decided analytically ("fraction of
 * samples below c"), and quantiles are the exact empirical quantiles.
 *
 * @code
 * -- Bulk load via array_agg over a sample table
 * INSERT INTO model_posteriors
 * SELECT param, provsql.empirical_samples(array_agg(value))
 * FROM mcmc_chain GROUP BY param;
 * @endcode
 *
 * At most 10000 distinct values (the categorical block cap): thin the
 * chain or bin the samples (e.g. with @c width_bucket) beyond that.
 *
 * @sa @ref categorical, @ref empirical_cdf
 * @sa <a href="https://en.wikipedia.org/wiki/Empirical_distribution_function">Wikipedia: Empirical distribution function</a>
 */
CREATE OR REPLACE FUNCTION empirical_samples(samples double precision[])
  RETURNS random_variable AS
$$
DECLARE
  n INTEGER;
  sorted double precision[];
  outcomes double precision[] := '{}';
  probs double precision[] := '{}';
  v double precision;
  prev double precision;
  run INTEGER := 0;
  started BOOLEAN := false;
BEGIN
  n := array_length(samples, 1);
  IF n IS NULL OR n < 1 THEN
    RAISE EXCEPTION 'provsql.empirical_samples: samples must be non-empty';
  END IF;
  sorted := ARRAY(SELECT s FROM unnest(samples) AS s ORDER BY 1);
  FOREACH v IN ARRAY sorted LOOP
    IF v IS NULL OR v = 'NaN'::float8
       OR v = 'Infinity'::float8 OR v = '-Infinity'::float8 THEN
      RAISE EXCEPTION
        'provsql.empirical_samples: samples must be finite (got %)', v;
    END IF;
    IF started AND v = prev THEN
      run := run + 1;
    ELSE
      IF started THEN
        outcomes := outcomes || prev;
        probs := probs || (run::double precision / n);
      END IF;
      prev := v;
      run := 1;
      started := true;
    END IF;
  END LOOP;
  outcomes := outcomes || prev;
  probs := probs || (run::double precision / n);
  IF array_length(outcomes, 1) > 10000 THEN
    RAISE EXCEPTION
      'provsql.empirical_samples: at most 10000 distinct values are '
      'supported (got %); thin the chain or bin the samples (e.g. with '
      'width_bucket)', array_length(outcomes, 1);
  END IF;
  RETURN provsql.categorical(probs, outcomes);
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Empirical-CDF constructor: a piecewise-linear CDF table as a
 *        @c random_variable.
 *
 * Loads a tabulated CDF -- simulation output percentile tables, risk
 * models, expert-elicited forecasts -- as the distribution whose CDF is
 * @c cdf[i] at @c grid[i], linear in between: mass
 * @c cdf[i+1] @c - @c cdf[i] spread uniformly over
 * <tt>(grid[i], grid[i+1])</tt>, plus (when @c cdf[1] @c > @c 0) an
 * atom of mass @c cdf[1] at @c grid[1] for the probability at or below
 * the grid start.  Packaged, like @ref gmm, as a stick-breaking cascade
 * of Bernoulli @ref mixture nodes over @ref uniform components (and the
 * optional @ref as_random atom), so moments and sampling are exact
 * through the existing mixture machinery; comparisons ride Monte Carlo.
 *
 * @code
 * provsql.empirical_cdf(
 *   grid => ARRAY[0.0, 0.5, 1.0, 2.0, 5.0, 10.0, 20.0],
 *   cdf  => ARRAY[0.32, 0.51, 0.67, 0.82, 0.94, 0.99, 1.0])
 * @endcode
 *
 * Validation: same-length arrays of at least two entries, @p grid
 * strictly increasing and finite, @p cdf non-decreasing within
 * <tt>[0, 1]</tt> and ending at @c 1 within @c 1e-9.
 *
 * @sa @ref gmm, @ref empirical_samples
 * @sa <a href="https://en.wikipedia.org/wiki/Cumulative_distribution_function">Wikipedia: Cumulative distribution function</a>
 */
CREATE OR REPLACE FUNCTION empirical_cdf(grid double precision[],
                                         cdf double precision[])
  RETURNS random_variable AS
$$
DECLARE
  n INTEGER;
  i INTEGER;
  acc random_variable := NULL;
  remaining double precision := 0.0;
  w double precision;
  comp random_variable;
BEGIN
  n := array_length(grid, 1);
  IF n IS NULL OR n < 2 THEN
    RAISE EXCEPTION 'provsql.empirical_cdf: grid must have at least two entries';
  END IF;
  IF array_length(cdf, 1) <> n THEN
    RAISE EXCEPTION 'provsql.empirical_cdf: grid and cdf must have the same length (got % and %)',
      n, array_length(cdf, 1);
  END IF;
  IF n > 10000 THEN
    RAISE EXCEPTION 'provsql.empirical_cdf: at most 10000 grid points are supported (got %)', n;
  END IF;
  FOR i IN 1..n LOOP
    IF grid[i] IS NULL OR grid[i] = 'NaN'::float8
       OR grid[i] = 'Infinity'::float8 OR grid[i] = '-Infinity'::float8 THEN
      RAISE EXCEPTION 'provsql.empirical_cdf: grid[%] must be finite (got %)', i, grid[i];
    END IF;
    IF i > 1 AND NOT grid[i] > grid[i-1] THEN
      RAISE EXCEPTION 'provsql.empirical_cdf: grid must be strictly increasing (grid[%] = %, grid[%] = %)',
        i-1, grid[i-1], i, grid[i];
    END IF;
    IF cdf[i] IS NULL OR cdf[i] = 'NaN'::float8 OR cdf[i] < 0 OR cdf[i] > 1 THEN
      RAISE EXCEPTION 'provsql.empirical_cdf: cdf[%] must be in [0,1] (got %)', i, cdf[i];
    END IF;
    IF i > 1 AND cdf[i] < cdf[i-1] THEN
      RAISE EXCEPTION 'provsql.empirical_cdf: cdf must be non-decreasing (cdf[%] = %, cdf[%] = %)',
        i-1, cdf[i-1], i, cdf[i];
    END IF;
  END LOOP;
  IF abs(cdf[n] - 1.0) > 1e-9 THEN
    RAISE EXCEPTION 'provsql.empirical_cdf: cdf must end at 1 within 1e-9 (got %)', cdf[n];
  END IF;
 
  -- Stick-breaking cascade, back to front: component i = 1 is the atom
  -- at the grid start (mass cdf[1]); component i >= 2 is
  -- uniform(grid[i-1], grid[i]) with mass cdf[i] - cdf[i-1].
  FOR i IN REVERSE n..1 LOOP
    w := CASE WHEN i = 1 THEN cdf[1] ELSE cdf[i] - cdf[i-1] END;
    IF w <= 0.0 THEN
      CONTINUE;
    END IF;
    comp := CASE WHEN i = 1 THEN provsql.as_random(grid[1])
                 ELSE provsql.uniform(grid[i-1], grid[i]) END;
    IF acc IS NULL THEN
      acc := comp;
      remaining := w;
    ELSE
      remaining := remaining + w;
      acc := provsql.mixture(least(1.0, w / remaining), comp, acc);
    END IF;
  END LOOP;
  RETURN acc;
END
$$ LANGUAGE plpgsql STRICT VOLATILE PARALLEL SAFE;
 
/**
 * @brief Lift a deterministic constant into a random_variable
 *
 * Creates a <tt>gate_value</tt> carrying the constant's TEXT form so
 * that comparisons against a <tt>random_variable</tt> column produce
 * the same circuit shape regardless of whether the operand is an
 * actual RV or a literal constant.
 *
 * Marked <tt>IMMUTABLE</tt>: the gate UUID is derived deterministically
 * from the constant via the same v5 convention as <tt>provenance_semimod</tt>'s
 * inline value gate (<tt>concat('value', CAST(c AS VARCHAR))</tt>), so
 * <tt>as_random(2)</tt> always resolves to the same gate, and any other
 * code path that already creates a value gate for the same constant
 * (e.g. <tt>provenance_semimod</tt>) shares the UUID.
 * <tt>create_gate</tt> is idempotent on already-mapped tokens, so
 * repeat invocations are harmless.
 *
 * @sa <a href="https://en.wikipedia.org/wiki/Degenerate_distribution">Wikipedia: Degenerate distribution (Dirac point mass)</a>
 */
CREATE OR REPLACE FUNCTION as_random(c double precision)
  RETURNS random_variable AS
$$
DECLARE
  -- Canonicalise -0.0 to +0.0: IEEE 754 defines x + 0.0 = +0.0 for
  -- both signed zeros, and is identity for finite, NaN, and ±Infinity.
  -- Without this, as_random(-0.0) and as_random(+0.0) would produce
  -- different gate UUIDs (their CAST AS VARCHAR TEXT representations
  -- differ: '-0' vs '0') even though they denote the same constant.
  c_canon double precision := c + 0.0;
  c_text varchar := CAST(c_canon AS VARCHAR);
  token UUID := public.uuid_generate_v5(
    provsql.uuid_ns_provsql(), concat('value', c_text));
BEGIN
  PERFORM provsql.create_gate(token, 'value');
  PERFORM provsql.set_extra(token, c_text);
  RETURN provsql.random_variable_make(token);
END
$$ LANGUAGE plpgsql STRICT IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Implicit cast double precision -> random_variable (lifts a
 *        scalar literal to a constant RV).
 *
 * Lets users write <tt>WHERE reading > 2.5::float8</tt> instead of
 * <tt>WHERE reading > provsql.as_random(2.5)</tt>; the planner-hook
 * rewriter then sees a uniform <tt>random_variable</tt> on both sides.
 * Sibling casts below cover @c INTEGER and @c NUMERIC literals so
 * plain <tt>WHERE reading > 2</tt> and <tt>WHERE reading > 2.5</tt>
 * also work; PostgreSQL's operator resolution does not chain casts
 * across more than one step, so each NUMERIC-source type needs its
 * own direct cast.
 */
CREATE CAST (double precision AS random_variable)
  WITH FUNCTION as_random(double precision) AS IMPLICIT;
 
/** @brief @c as_random for @c INTEGER (delegates to the @c float8 form). */
CREATE OR REPLACE FUNCTION as_random(c INTEGER)
  RETURNS random_variable AS
$$ SELECT provsql.as_random(c::double precision); $$
LANGUAGE sql STRICT IMMUTABLE PARALLEL SAFE;
 
/** @brief @c as_random for @c NUMERIC (delegates to the @c float8 form). */
CREATE OR REPLACE FUNCTION as_random(c NUMERIC)
  RETURNS random_variable AS
$$ SELECT provsql.as_random(c::double precision); $$
LANGUAGE sql STRICT IMMUTABLE PARALLEL SAFE;
 
/** @brief Implicit cast INTEGER -> random_variable. */
CREATE CAST (INTEGER AS random_variable)
  WITH FUNCTION as_random(INTEGER) AS IMPLICIT;
 
/** @brief Implicit cast NUMERIC -> random_variable. */
CREATE CAST (NUMERIC AS random_variable)
  WITH FUNCTION as_random(NUMERIC) AS IMPLICIT;
 
/**
 * @name Arithmetic and comparison on random_variable
 *
 * Each binary operator below is declared on @c (random_variable,
 * random_variable) only; mixed shapes such as <tt>rv + 2</tt> or
 * <tt>2.5 > rv</tt> resolve through the implicit casts from
 * @c INTEGER / @c NUMERIC / @c double @c precision to
 * @c random_variable declared above.  This avoids the resolution
 * ambiguity that would arise if both <tt>(rv, NUMERIC)</tt> and
 * <tt>(rv, rv)</tt> overloads were declared while implicit casts also
 * existed.
 *
 * Arithmetic operators build a @c gate_arith via @c provenance_arith
 * and return a new @c random_variable wrapping its UUID.
 *
 * Comparison operators are placeholders that return @c BOOLEAN and
 * raise if executed -- the @c BOOLEAN return type is required so that
 * PostgreSQL accepts <tt>WHERE rv > 2</tt> at parse-analyze.  The
 * planner hook intercepts every such @c OpExpr (matched by
 * @c opfuncid against @c constants_t::OID_FUNCTION_RV_CMP) and rewrites
 * it into a @c provenance_cmp call whose UUID is conjoined into the
 * tuple's @c provsql column via @c provenance_times.  Code that needs
 * a @c gate_cmp UUID directly (without going through the planner hook)
 * uses the @c rv_cmp_* family below, which call @c provenance_cmp
 * with the matching float8-comparator OID.
 *
 * @{
 */
 
/** @brief @c random_variable + @c random_variable (gate_arith PLUS). */
CREATE OR REPLACE FUNCTION random_variable_plus(
  a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      0,  -- PROVSQL_ARITH_PLUS
      ARRAY[(a)::UUID,
            (b)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief @c random_variable - @c random_variable (gate_arith MINUS). */
CREATE OR REPLACE FUNCTION random_variable_minus(
  a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      2,  -- PROVSQL_ARITH_MINUS
      ARRAY[(a)::UUID,
            (b)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief @c random_variable * @c random_variable (gate_arith TIMES). */
CREATE OR REPLACE FUNCTION random_variable_times(
  a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      1,  -- PROVSQL_ARITH_TIMES
      ARRAY[(a)::UUID,
            (b)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief @c random_variable / @c random_variable (gate_arith DIV). */
CREATE OR REPLACE FUNCTION random_variable_div(
  a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      3,  -- PROVSQL_ARITH_DIV
      ARRAY[(a)::UUID,
            (b)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Unary @c -random_variable (gate_arith NEG). */
CREATE OR REPLACE FUNCTION random_variable_neg(a random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      4,  -- PROVSQL_ARITH_NEG
      ARRAY[(a)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief @c random_variable ^ @c random_variable (gate_arith POW).
 *
 * Real-valued branch only: evaluation raises if a negative base is
 * drawn together with a non-INTEGER exponent (write
 * <tt>pow(greatest(x, 0), p)</tt> for the non-negative branch).
 */
CREATE OR REPLACE FUNCTION random_variable_pow(
  a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      7,  -- PROVSQL_ARITH_POW
      ARRAY[(a)::UUID,
            (b)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Natural logarithm of a @c random_variable (gate_arith LN).
 *
 * Defined on @c [0, +Infinity): evaluation raises if a negative value
 * is drawn (restrict the argument's support); a draw of exactly @c 0
 * yields @c -Infinity.
 */
CREATE OR REPLACE FUNCTION ln(a random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      8,  -- PROVSQL_ARITH_LN
      ARRAY[(a)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief @c e^x for a @c random_variable (gate_arith EXP).  Total. */
CREATE OR REPLACE FUNCTION exp(a random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(
    provsql.provenance_arith(
      9,  -- PROVSQL_ARITH_EXP
      ARRAY[(a)::UUID]));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief @c pow / @c power spellings of the @c ^ operator, mirroring
 *        PostgreSQL's NUMERIC surface.  Scalar exponents resolve
 *        through the implicit NUMERIC-to-rv casts:
 *        <tt>pow(x, 0.5)</tt> is <tt>x ^ 0.5</tt>.
 */
CREATE OR REPLACE FUNCTION pow(a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_pow(a, b);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION power(a random_variable, b random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_pow(a, b);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Square root of a @c random_variable: sugar for
 *        <tt>x ^ 0.5</tt> (no gate or opcode of its own).  Evaluation
 *        raises on a negative draw, like any non-INTEGER exponent.
 */
CREATE OR REPLACE FUNCTION sqrt(a random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.random_variable_pow(a, provsql.as_random(0.5));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Internal helper: float8-comparator OID for a given symbol.
 *
 * Wraps the @c '&lt;sym&gt;(double precision,double precision)'::regoperator
 * lookup so the per-comparator functions read uniformly.  Marked
 * @c IMMUTABLE because the resolved OID is fixed at catalog level
 * (the float8 comparators are core PG and never re-installed).
 */
CREATE OR REPLACE FUNCTION random_variable_cmp_oid(sym TEXT)
  RETURNS oid AS
$$
  SELECT (sym || '(double precision,double precision)')::regoperator::oid;
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/* The six @c random_variable_{lt,le,eq,ne,ge,gt} functions below are
 * BOOLEAN placeholders -- they exist only so the @c (rv, rv) operators
 * can be declared at all (PostgreSQL needs a procedure to bind to the
 * operator definition, and a procedure returning anything but @c BOOLEAN
 * would be rejected by parse-analyze in a WHERE position).  They MUST
 * NOT be invoked directly: the planner hook in @c src/provsql.c
 * intercepts every @c OpExpr whose @c opfuncid matches one of these and
 * rewrites it into a @c provenance_cmp() call against the row's
 * provenance.  If the executor ever reaches one of these, it means the
 * planner hook was bypassed (e.g. @c provsql.active was off), in which
 * case raising is the right behaviour. */
 
/** @brief Placeholder body shared by every <tt>random_variable_*</tt>
 *  comparison procedure.  Raises with a uniform message. */
CREATE OR REPLACE FUNCTION random_variable_cmp_placeholder(
  a random_variable, b random_variable)
  RETURNS BOOLEAN AS
$$
BEGIN
  RAISE EXCEPTION 'random_variable comparison must be rewritten by the '
                  'ProvSQL planner hook (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_lt(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_le(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_eq(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_ne(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_ge(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION random_variable_gt(
  a random_variable, b random_variable) RETURNS BOOLEAN AS
$$ SELECT provsql.random_variable_cmp_placeholder(a, b); $$
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/* Direct UUID constructors -- used by tests and any caller that wants
 * a @c gate_cmp without going through the planner hook (e.g. building
 * a circuit fragment in a SELECT list).  Each delegates to
 * @c provenance_cmp with the matching float8-comparator OID. */
 
/** @brief Build a @c gate_cmp for <tt>a &lt; b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_lt(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('<'),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Build a @c gate_cmp for <tt>a &le; b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_le(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('<='),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Build a @c gate_cmp for <tt>a = b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_eq(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('='),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Build a @c gate_cmp for <tt>a &lt;&gt; b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_ne(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('<>'),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Build a @c gate_cmp for <tt>a &ge; b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_ge(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('>='),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Build a @c gate_cmp for <tt>a &gt; b</tt> and return its UUID. */
CREATE OR REPLACE FUNCTION rv_cmp_gt(
  a random_variable, b random_variable) RETURNS UUID AS
$$
  SELECT provsql.provenance_cmp(
    (a)::UUID,
    provsql.random_variable_cmp_oid('>'),
    (b)::UUID);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR + (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_plus,
  COMMUTATOR = +
);
 
CREATE OPERATOR - (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_minus
);
 
CREATE OPERATOR * (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_times,
  COMMUTATOR = *
);
 
CREATE OPERATOR / (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_div
);
 
/** @brief Prefix unary minus on @c random_variable. */
CREATE OPERATOR - (
  RIGHTARG  = random_variable,
  PROCEDURE = random_variable_neg
);
 
CREATE OPERATOR ^ (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_pow
);
 
CREATE OPERATOR < (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_lt,
  COMMUTATOR = >,
  NEGATOR    = >=
);
 
CREATE OPERATOR <= (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_le,
  COMMUTATOR = >=,
  NEGATOR    = >
);
 
CREATE OPERATOR = (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_eq,
  COMMUTATOR = =,
  NEGATOR    = <>
);
 
CREATE OPERATOR <> (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_ne,
  COMMUTATOR = <>,
  NEGATOR    = =
);
 
CREATE OPERATOR >= (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_ge,
  COMMUTATOR = <=,
  NEGATOR    = <
);
 
CREATE OPERATOR > (
  LEFTARG    = random_variable,
  RIGHTARG   = random_variable,
  PROCEDURE  = random_variable_gt,
  COMMUTATOR = <,
  NEGATOR    = <=
);
 
/**
 * @brief btree comparison support for @c random_variable -- always an error.
 *
 * A @c random_variable is a distribution, not a scalar, so it has no total
 * order: sorting (@c ORDER @c BY), de-duplicating (@c DISTINCT), grouping, and
 * the built-in @c GREATEST / @c LEAST all reduce to this btree comparison
 * proc, which raises a clear diagnostic rather than a placeholder message.
 *
 * The proc exists only so a DEFAULT btree operator class can be declared for
 * @c random_variable -- which is what lets PostgreSQL's @c GREATEST / @c LEAST
 * grammar parse over random variables so the planner hook can lift it into a
 * @c gate_arith @c MAX / @c MIN order statistic.  When the hook is active the
 * @c GREATEST / @c LEAST node is rewritten before it ever calls this proc.
 */
CREATE OR REPLACE FUNCTION random_variable_btree_cmp(
  a random_variable, b random_variable) RETURNS INTEGER AS
$$
BEGIN
  RAISE EXCEPTION 'comparison or ordering of random_variable values is '
                  'meaningless: a random_variable is a distribution, not a scalar'
    USING HINT =
      'Compare them as a probabilistic event -- in a WHERE / JOIN clause or '
      'with probability(x > y); take order statistics with provsql.greatest / '
      'provsql.least (or the min / max aggregates); summarise numerically with '
      'expected / variance / support.';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
-- DEFAULT btree operator class over the (planner-hook-lifted) comparison
-- operators.  Its only purpose is to make GREATEST / LEAST over random_variable
-- parse; every actual comparison it would drive (ORDER BY, DISTINCT, an
-- un-rewritten GREATEST) funnels through random_variable_btree_cmp above and
-- raises the "meaningless" diagnostic.
CREATE OPERATOR CLASS random_variable_ops
  DEFAULT FOR TYPE random_variable USING btree AS
    OPERATOR 1 <,
    OPERATOR 2 <=,
    OPERATOR 3 =,
    OPERATOR 4 >=,
    OPERATOR 5 >,
    FUNCTION 1 random_variable_btree_cmp(random_variable, random_variable);
 
/**
 * @brief Condition a random variable on an event: @c "X | C".
 *
 * Returns a conditioned distribution that flows onward like any other
 * @c random_variable: it can be stored, re-conditioned, and queried with
 * @c expected / @c variance / @c moment / @c support, which then report the
 * conditional distribution.  @p cond is a Boolean-event provenance token,
 * typically a comparison over the variable itself (@c "X | rv_cmp_gt(X,
 * as_random(3))" -- a truncation) or any external event.
 *
 * Unlike the UUID carrier's terminal @c cond, the random-variable form is a
 * composable two-child @c gate_conditioned @c [target, condition]: the moment
 * / support dispatchers unpack it and route through the existing conditional
 * evaluator (@c rv_moment over the joint of the target and the condition).
 * Nested conditioning folds: @c "(X|A)|B = X|(A∧B)".
 */
CREATE OR REPLACE FUNCTION random_variable_cond(rv random_variable, cond UUID)
  RETURNS random_variable AS
$$
DECLARE
  tgt UUID;
  ev  UUID;
  result UUID;
  ch UUID[];
BEGIN
  IF cond IS NULL OR cond = gate_one() THEN
    RETURN rv;
  END IF;
 
  -- A point-equality "Y = c" on a bare random-variable leaf is an
  -- OBSERVATION, not a truncation: rewrite it to the internal likelihood-
  -- weighting evidence (its density / mass at c).  This is what lets
  -- "X | (normal(mu,1) = 8)" (a continuous point event, measure-zero as a
  -- Boolean selection) condition as the disintegration rather than fold to
  -- an infeasible event.
  cond := provsql.evidence_as_observation(cond);
 
  tgt := (rv)::UUID;
  IF get_gate_type(tgt) = 'conditioned'
     AND array_length(get_children(tgt), 1) = 2 THEN
    -- Fold (X|A)|B = X|(A∧B): the rv-carrier conditioned gate is the
    -- two-child [target, condition] shape; accumulate the new event.
    ch  := get_children(tgt);
    tgt := ch[1];
    ev  := provenance_times(ch[2], cond);
  ELSE
    ev := cond;
  END IF;
 
  result := public.uuid_generate_v5(uuid_ns_provsql(),
                                    concat('conditioned', tgt, ev));
  PERFORM create_gate(result, 'conditioned', ARRAY[tgt, ev]);
  RETURN (result)::random_variable;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public
   SECURITY DEFINER PARALLEL SAFE;
 
CREATE OPERATOR | (
  LEFTARG   = random_variable,
  RIGHTARG  = UUID,
  PROCEDURE = random_variable_cond
);
 
/**
 * @brief Placeholder for @c "X | (predicate)" -- conditioning a random
 *        variable on a Boolean comparison written naturally.
 *
 * Lets one write @c "X | (X > 3)" instead of
 * @c "X | rv_cmp_gt(X, as_random(3))".  Never executes: the ProvSQL planner
 * hook rewrites the Boolean operand (a combination of random_variable
 * comparisons) into the corresponding condition gate and emits
 * @c random_variable_cond.  Reaching it at runtime means the rewriter was
 * inactive or the predicate was not a random_variable comparison.
 */
CREATE OR REPLACE FUNCTION random_variable_cond_predicate(
  rv random_variable, predicate BOOLEAN) RETURNS random_variable AS
$$
BEGIN
  RAISE EXCEPTION 'random_variable | (predicate) must be rewritten by the '
    'ProvSQL planner hook: the right operand must be a Boolean combination '
    'of random_variable comparisons (is provsql.active off?)';
END
$$ LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
 
CREATE OPERATOR | (
  LEFTARG   = random_variable,
  RIGHTARG  = BOOLEAN,
  PROCEDURE = random_variable_cond_predicate
);
 
/**
 * @brief Unpack the target of a random-variable conditioning gate.
 *
 * For a two-child @c gate_conditioned @c [target, condition] (the @c "X | C"
 * shape) returns @p target; for any other token returns it unchanged.  Used
 * by the moment / support dispatchers to route a conditioned distribution
 * through the existing conditional evaluator.
 */
CREATE OR REPLACE FUNCTION rv_conditioned_target(token UUID) RETURNS UUID AS
$$
  SELECT CASE
    WHEN provsql.get_gate_type(token) = 'conditioned'
         AND array_length(provsql.get_children(token), 1) = 2
    THEN (provsql.get_children(token))[1]
    ELSE token
  END;
$$ LANGUAGE sql STABLE PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Combine a conditioning gate's event with an explicit @p prov.
 *
 * For a two-child @c gate_conditioned @c [target, condition] returns
 * @c "condition ∧ prov"; otherwise returns @p prov unchanged.  Lets a stored
 * @c "X | C" be queried as @c expected(X|C) (prov defaulting to one) or have
 * an extra condition conjoined as @c expected(X|C, extra_prov).
 */
CREATE OR REPLACE FUNCTION rv_conditioned_prov(token UUID, prov UUID)
  RETURNS UUID AS
$$
  SELECT CASE
    WHEN provsql.get_gate_type(token) = 'conditioned'
         AND array_length(provsql.get_children(token), 1) = 2
    THEN provsql.provenance_times((provsql.get_children(token))[2], prov)
    ELSE prov
  END;
$$ LANGUAGE sql STABLE PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/*
 * Latent-variable posterior inference.
 *
 * Likelihood weighting (self-normalised importance sampling): bind an
 * observed datum to a latent-dependent random-variable leaf with observe,
 * conjoin the per-observation evidence with and_agg into a single evidence
 * token, and pass it as the prov conditioning argument of any moment /
 * quantile / sample readout.  Latents are drawn from the prior (the
 * existing forward recursion) and each draw is weighted by the observed
 * leaves' densities at the data; the readouts then report the posterior.
 * It is the continuous generalisation of the rejection-based conditioning:
 * a Boolean event in the evidence contributes a 0/1 weight, an observe
 * contributes a pdf weight -- same evidence conjunction, same
 * "P(query AND evidence)/P(evidence)" normaliser, now weighted.
 */
 
/**
 * @brief Internal: rewrite a point-equality conditioning event into an
 *        observation.  If @p ev is a @c gate_cmp with the @c "=" operator,
 *        one side a bare @c gate_rv leaf and the other a constant, return
 *        @c observe(leaf, const); otherwise return @p ev unchanged.
 *
 * This is the bridge that makes the natural equality form the surface for
 * likelihood-weighting conditioning: @c "X | (Y = c)" and @c "given(Y = c)"
 * both produce a @c gate_cmp, which this turns into density evidence.  A
 * point event on a bare leaf is only meaningful as an observation (a
 * continuous @c "Y = c" is measure-zero as a Boolean selection), so the
 * rewrite is unambiguous.  Non-equality / non-leaf events pass through as
 * ordinary Boolean conditioning.
 */
CREATE OR REPLACE FUNCTION evidence_as_observation(ev UUID) RETURNS UUID AS
$$
DECLARE
  ch UUID[];
  i1 INTEGER;
  leaf UUID;
  datum_gate UUID;
BEGIN
  IF ev IS NULL OR provsql.get_gate_type(ev) <> 'cmp' THEN
    RETURN ev;
  END IF;
  ch := provsql.get_children(ev);
  IF array_length(ch, 1) <> 2 THEN
    RETURN ev;
  END IF;
  -- The cmp stores the comparison OPERATOR's OID in info1; match on its name
  -- '=' the same way the C-side cmpOpFromOid does (get_opname), rather than a
  -- fixed operator OID (which varies per install / carrier type).
  SELECT info1 INTO i1 FROM provsql.get_infos(ev);
  IF (SELECT oprname FROM pg_catalog.pg_operator WHERE oid = i1) IS DISTINCT FROM '=' THEN
    RETURN ev;   -- not an equality
  END IF;
  IF provsql.get_gate_type(ch[1]) = 'rv'
     AND provsql.get_gate_type(ch[2]) = 'value' THEN
    leaf := ch[1]; datum_gate := ch[2];
  ELSIF provsql.get_gate_type(ch[2]) = 'rv'
        AND provsql.get_gate_type(ch[1]) = 'value' THEN
    leaf := ch[2]; datum_gate := ch[1];
  ELSE
    RETURN ev;   -- not a bare-leaf-vs-constant point event
  END IF;
  RETURN provsql.observe((leaf)::random_variable,
                         provsql.get_extra(datum_gate)::double precision);
END
$$ LANGUAGE plpgsql VOLATILE
   SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Internal: bind an observed datum to a random-variable leaf --
 *        the likelihood-weighting evidence behind @c "X | (Y = d)".
 *
 * @p x MUST be a bare @c gate_rv leaf (typically a latent-parameterised
 * one, e.g. @c normal(mu, 1) sharing a latent @c mu across rows).
 * Returns an @b evidence UUID -- a @c gate_observe wrapping the leaf with
 * the datum in @c extra -- that composes with other evidence through
 * @c and_agg (a @c gate_times conjunction) and is consumed by the
 * importance-sampling weight walk, contributing the factor @c f_X(d).
 *
 * Internal: the user-facing surface is the equality form @c "X | (Y = d)"
 * (single conditioning) and @c "given(Y = d)" (per-row evidence for
 * @c and_agg), both of which route here through @c evidence_as_observation.
 *
 * A fresh gate is minted per call (each observation is a distinct
 * evidence atom, so a repeated @c (leaf, datum) contributes its density
 * factor once per row -- and each is a separate Shapley atom).  Observing
 * a derived quantity (@c observe(X+Y, d)) is out of scope: it needs a
 * change-of-variables density; a non-leaf argument is refused.
 */
CREATE OR REPLACE FUNCTION observe(x random_variable, datum double precision)
  RETURNS UUID AS
$$
DECLARE
  leaf UUID := (x)::UUID;
  result UUID;
BEGIN
  IF provsql.get_gate_type(leaf) <> 'rv' THEN
    RAISE EXCEPTION 'provsql.observe: the argument must be a bare '
      'random-variable leaf (a gate_rv), got a % gate', provsql.get_gate_type(leaf)
      USING HINT = 'observe binds a datum to a single distribution leaf; '
        'observing a derived quantity (a sum, product, or comparison) needs '
        'a change-of-variables density and is out of scope.';
  END IF;
  IF NOT provsql.is_finite_float8(datum) THEN
    RAISE EXCEPTION 'provsql.observe: datum must be finite (got %)', datum;
  END IF;
  result := public.uuid_generate_v4();
  PERFORM provsql.create_gate(result, 'observe', ARRAY[leaf]);
  PERFORM provsql.set_extra(result, datum::TEXT);
  RETURN result;
END
$$ LANGUAGE plpgsql VOLATILE
   SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE;
 
/**
 * @brief Conjunction state function for @c and_agg (evidence @c gate_times).
 *
 * Not @c STRICT: @c provenance_times maps a @c NULL operand to the times
 * neutral, so an empty group leaves the state @c NULL (no evidence) and a
 * first row seeds it with that row's evidence.
 */
CREATE OR REPLACE FUNCTION and_agg_sfunc(state UUID, ev UUID)
  RETURNS UUID AS
$$
  SELECT provsql.provenance_times(state, ev);
$$ LANGUAGE sql PARALLEL SAFE;
 
/**
 * @brief Conjoin per-row evidence tokens into one evidence circuit.
 *
 * The evidence-conjunction counterpart used to fold one @c observe (or any
 * Boolean conditioning event) per row into a single @c gate_times root, to
 * be passed as the @c prov argument of the moment / quantile / sample
 * readouts.  An empty group yields @c NULL (no evidence).
 */
CREATE AGGREGATE and_agg(UUID) (
  SFUNC = and_agg_sfunc,
  STYPE = UUID
);
 
/**
 * @brief Marginal likelihood @c P(data) of an evidence circuit.
 *
 * The mean raw importance weight over @c provsql.rv_mc_samples prior draws
 * -- the same quantity rejection conditioning computes as @c P(C), now the
 * product of the observations' densities.  @p evidence is an @c and_agg
 * conjunction of @c observe tokens (and/or Boolean events).
 */
CREATE OR REPLACE FUNCTION evidence(evidence UUID)
  RETURNS double precision
  AS 'provsql','rv_evidence' LANGUAGE C STRICT PARALLEL SAFE;
 
/**
 * @brief The @c observe atoms of an evidence circuit.
 *
 * Collects every @c gate_observe leaf reachable through the @c gate_times
 * conjunction spine (the shape @c and_agg builds -- a possibly left-nested
 * tree, since @c provenance_times does not flatten).  Used by
 * @c shapley_observe to recover the flat observation set regardless of the
 * conjunction's nesting.
 */
CREATE OR REPLACE FUNCTION observe_atoms(evidence UUID)
  RETURNS UUID[] AS
$$
  WITH RECURSIVE walk(tok) AS (
    SELECT evidence
    UNION
    SELECT c
    FROM walk, LATERAL unnest(provsql.get_children(walk.tok)) AS c
    WHERE provsql.get_gate_type(walk.tok) = 'times'
  )
  SELECT array_agg(tok ORDER BY tok)
  FROM walk
  WHERE provsql.get_gate_type(tok) = 'observe';
$$ LANGUAGE sql STABLE PARALLEL SAFE SET search_path=provsql,pg_temp,public;
 
/**
 * @brief Shapley attribution of each observation to a posterior moment.
 *
 * "Which observation most shifted my posterior?"  Because the importance
 * weight is a product of per-observation density factors, dropping an
 * observation is dropping one factor: the classical Shapley value of each
 * @c gate_observe atom over the coalitional value function
 * @c "v(S) = payoff(target | observations in S)" is the attribution, a
 * byproduct of the same likelihood-weighting machinery (see the
 * explainable-inference angle in the continuous-distributions notes).
 *
 * @p target is the latent (its @c UUID); @p evidence is the @c and_agg
 * conjunction of @c observe atoms; @p payoff is @c 'expected' or
 * @c 'variance'.  Returns each observation atom with its Shapley value; the
 * values sum to @c "payoff(target | all data) - payoff(target)" (Shapley
 * efficiency: the total shift from prior to posterior).
 *
 * Exact enumeration over the @c 2^n observation subsets, so it is capped at
 * @c n = 12 observations (sampling-based attribution for larger sets is
 * future work); pin @c provsql.monte_carlo_seed so the coalitional value
 * functions share common random numbers (lower-variance differences).
 */
CREATE OR REPLACE FUNCTION shapley_observe(
    target UUID, evidence UUID, payoff TEXT DEFAULT 'expected')
  RETURNS TABLE(observation UUID, value double precision) AS
$$
DECLARE
  atoms UUID[];
  n INT;
  nmasks INT;
  pv double precision[];
  popc INT[];
  fact double precision[];
  mask INT;
  i INT;
  j INT;
  cnt INT;
  subset UUID[];
  ev_s UUID;
  sh double precision;
  bit INT;
  s_size INT;
BEGIN
  IF payoff NOT IN ('expected', 'variance') THEN
    RAISE EXCEPTION 'provsql.shapley_observe: payoff must be ''expected'' or '
      '''variance'' (got %)', payoff;
  END IF;
  atoms := provsql.observe_atoms(evidence);
  n := coalesce(array_length(atoms, 1), 0);
  IF n = 0 THEN
    RAISE EXCEPTION 'provsql.shapley_observe: evidence contains no observe() '
      'atoms (got a % gate)', provsql.get_gate_type(evidence);
  END IF;
  IF n > 12 THEN
    RAISE EXCEPTION 'provsql.shapley_observe: exact attribution over % '
      'observations is exponential; capped at 12 (sampling-based '
      'attribution is future work)', n;
  END IF;
 
  -- factorials 0!..n!  (fact[k+1] = k!)
  fact := ARRAY[1::double precision];
  FOR i IN 1..n LOOP fact := fact || (fact[i] * i); END LOOP;
 
  nmasks := (1 << n);
  pv   := array_fill(NULL::double precision, ARRAY[nmasks]);
  popc := array_fill(0, ARRAY[nmasks]);
 
  -- Payoff value function for every subset of observations.
  FOR mask IN 0 .. nmasks - 1 LOOP
    subset := ARRAY[]::UUID[];
    cnt := 0;
    FOR i IN 0 .. n - 1 LOOP
      IF (mask >> i) & 1 = 1 THEN
        subset := subset || atoms[i + 1];
        cnt := cnt + 1;
      END IF;
    END LOOP;
    popc[mask + 1] := cnt;
    IF cnt = 0 THEN
      ev_s := provsql.gate_one();              -- prior (no evidence)
    ELSE
      ev_s := provsql.provenance_times(VARIADIC subset);
    END IF;
    IF payoff = 'expected' THEN
      pv[mask + 1] := provsql.rv_moment(target, 1, false, ev_s);
    ELSE
      pv[mask + 1] := provsql.rv_moment(target, 2, true, ev_s);
    END IF;
  END LOOP;
 
  -- Shapley value of each observation atom.
  FOR i IN 0 .. n - 1 LOOP
    sh := 0;
    bit := (1 << i);
    FOR mask IN 0 .. nmasks - 1 LOOP
      IF (mask >> i) & 1 = 0 THEN               -- subsets S not containing i
        s_size := popc[mask + 1];
        -- weight |S|! (n-|S|-1)! / n!
        sh := sh + (fact[s_size + 1] * fact[n - s_size] / fact[n + 1])
                 * (pv[(mask | bit) + 1] - pv[mask + 1]);
      END IF;
    END LOOP;
    observation := atoms[i + 1];
    value := sh;
    RETURN NEXT;
  END LOOP;
END
$$ LANGUAGE plpgsql VOLATILE
   SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/**
 * @name Order statistics over random_variable
 *
 * Same-row @c greatest / @c least over @c random_variable arguments: the
 * order-statistic counterpart of the element-wise @c "+ - * /" operators.
 * They lower to a single @c gate_arith with the @c MAX / @c MIN opcode over
 * the argument circuits, the same n-ary shape the @c max / @c min aggregates
 * build.  Evaluation is Monte-Carlo-correct out of the box (@c std::max /
 * @c std::min over the jointly-sampled children, so shared base RVs stay
 * coupled); closed forms for i.i.d. families come from the analytic
 * order-statistic pass.
 *
 * PostgreSQL's built-in @c GREATEST / @c LEAST are dedicated syntax (a
 * @c MinMaxExpr requiring a btree comparison), not overloadable functions, so
 * the surface is the schema-qualified @c provsql.greatest(...) /
 * @c provsql.least(...).  @c NULL arguments are ignored, matching the built-in
 * (an all-@c NULL / empty call returns @c NULL).
 * @{
 */
 
-- "greatest" / "least" are col_name keywords, so the CREATE FUNCTION name
-- must be quoted; callers reach them qualified as provsql.greatest(...).
-- Idempotence: max / min ignore repeats, so identical children (same gate)
-- are de-duplicated -- greatest(x, x, y) == greatest(x, y) -- and a single
-- surviving child collapses to itself -- greatest(x) == x.  DISTINCT also sorts
-- the children, so the argument order does not matter for gate sharing.  (Two
-- independent draws of the same distribution are distinct gates and are NOT
-- de-duplicated.)
CREATE OR REPLACE FUNCTION "greatest"(VARIADIC args random_variable[])
  RETURNS random_variable AS
$$
DECLARE
  children UUID[];
BEGIN
  IF args IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT array_agg(DISTINCT (a)::UUID) INTO children
    FROM unnest(args) a WHERE a IS NOT NULL;
  IF children IS NULL OR array_length(children, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(children, 1) = 1 THEN
    RETURN provsql.random_variable_make(children[1]);
  END IF;
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(5, children));  -- 5 = PROVSQL_ARITH_MAX
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION "least"(VARIADIC args random_variable[])
  RETURNS random_variable AS
$$
DECLARE
  children UUID[];
BEGIN
  IF args IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT array_agg(DISTINCT (a)::UUID) INTO children
    FROM unnest(args) a WHERE a IS NOT NULL;
  IF children IS NULL OR array_length(children, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(children, 1) = 1 THEN
    RETURN provsql.random_variable_make(children[1]);
  END IF;
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(6, children));  -- 6 = PROVSQL_ARITH_MIN
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Build a @c random_variable from a guarded-selection @c gate_case.
 *
 * Thin @c random_variable wrapper over @c provenance_case (defined with the
 * other gate builders, since it is UUID-only), the target of the planner-hook
 * @c CASE-over-RV rewrite: the hook flattens the branches into
 * @c [guard_1, value_1, ..., default] and emits this call so an RV-typed
 * @c CASE surfaces as a first-class @c random_variable.
 */
CREATE OR REPLACE FUNCTION rv_case(
  children UUID[]
)
RETURNS random_variable AS
$$
  SELECT provsql.random_variable_make(provsql.provenance_case(children));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Build an @c AGG_TOKEN from a guarded-selection @c gate_case.
 *
 * The aggregate-carrier analogue of @c rv_case: a thin @c AGG_TOKEN wrapper
 * over the carrier-agnostic @c provenance_case, the target of the planner-hook
 * lowering of a searched @c CASE whose guards are aggregate comparisons and
 * whose branches are aggregates. The branches (and default) are already
 * flattened into @c [guard_1, value_1, ..., default] UUIDs.  The display cell
 * carries the actual-world CASE value -- the branch selected on the actual
 * data, resolved by @c agg_gate_value, exactly as a bare aggregate's cell
 * carries its actual-world value.  The probabilistic result is produced by
 * the measure evaluators (``expected`` / ``probability`` / possible-worlds /
 * Monte Carlo) from the gate, not the token's cell.
 */
CREATE OR REPLACE FUNCTION agg_case(
  children UUID[]
)
RETURNS AGG_TOKEN AS
$$
  SELECT provsql.agg_token_make(t, coalesce(provsql.agg_gate_value(t), 0))
  FROM (SELECT provsql.provenance_case(children) AS t) AS s;
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @} */
 
/**
 * @name Aggregates over random_variable
 *
 * An overload of the standard
 * @c sum aggregate that takes a @c random_variable per row and returns
 * the @c random_variable representing the (provenance-weighted) sum.
 * Lives in the @c provsql schema so a @c sum(random_variable) call
 * resolves to it without colliding with the built-in NUMERIC @c sum
 * overloads in @c pg_catalog.
 *
 * Direct calls outside a provenance-tracked query treat each row's
 * contribution unconditionally (no per-row Boolean selector).  When
 * the planner hook sees a @c provsql.sum @c Aggref over a
 * provenance-tracked query, it wraps the per-row argument @c x in
 * <tt>provsql.mixture(prov_token, x, provsql.as_random(0))</tt> so the
 * aggregate's effective semantics become
 * @f$\mathrm{SUM}(x) = \sum_i \mathbf{1}\{\varphi_i\} \cdot X_i@f$,
 * the natural extension of semimodule-provenance to RV-valued M.
 *
 * The internal state is the array of UUIDs of the per-row mixtures.
 * The final function builds a single @c gate_arith @c PLUS over them
 * (or returns @c as_random(0) for an empty group, the additive
 * identity).  Sharing on @c provenance_arith's v5 hash means two
 * @c sum invocations over the same set of rows collide on the same
 * gate.
 *
 * @{
 */
 
/**
 * @brief Per-row helper: wrap an RV in @c mixture(prov, rv, as_random(0)).
 *
 * Internal helper used by the planner-hook rewriter to lift a
 * @c sum(random_variable) argument into its provenance-aware form.
 * Encodes one row's contribution to the SUM as a Bernoulli mixture
 * over the row's provenance: with probability @c P(prov) the mixture
 * samples @c rv, otherwise it samples the additive identity
 * @c as_random(0).  Exposed as a regular SQL function so the planner
 * can construct a @c FuncExpr by name without needing to disambiguate
 * @c mixture / @c as_random overloads at OID-lookup time.
 */
CREATE OR REPLACE FUNCTION rv_aggregate_semimod(
  prov UUID, rv random_variable)
  RETURNS random_variable AS
$$
  SELECT provsql.mixture(prov, rv, provsql.as_random(0::double precision));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Identity-parameterised per-row wrap for an RV-returning aggregate.
 *
 * Generalises the two-argument @ref rv_aggregate_semimod.  The else-branch
 * (a row's contribution when its provenance is false) is
 * @c as_random(@p identity) instead of the additive @c as_random(0).  The
 * planner-hook rewrite bakes each aggregate's own identity element into the
 * wrap -- @c 1 for @c product, @f$-\infty@f$ / @f$+\infty@f$ for @c max /
 * @c min -- so the aggregate's final function is a plain fold over the
 * per-row mixtures with no gate inspection.  @c sum keeps the two-argument
 * form (@c identity @c = @c 0).
 */
CREATE OR REPLACE FUNCTION rv_aggregate_semimod(
  prov UUID, rv random_variable, identity double precision)
  RETURNS random_variable AS
$$
  SELECT provsql.mixture(prov, rv, provsql.as_random(identity));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Per-row denominator wrap for @c avg(random_variable): the
 *        provenance indicator @f$\mathbf{1}\{\varphi\}@f$.
 *
 * The row contributes @c 1 to the running count when present and @c 0 when
 * absent, so @c sum over these wraps is the provenance-weighted count
 * @f$\sum_i \mathbf{1}\{\varphi_i\}@f$.  The planner-hook rewrites
 * @c avg(x) into @c rv_sum_or_null(rv_aggregate_semimod(prov, x)) @c /
 * @c sum(rv_aggregate_indicator(prov)) -- the "@c AVG @c = @c SUM @c /
 * @c COUNT" identity lifted into the @c random_variable algebra -- so
 * @c avg rides entirely on @c sum's fold and never inspects a gate.
 */
CREATE OR REPLACE FUNCTION rv_aggregate_indicator(prov UUID)
  RETURNS random_variable AS
$$
  SELECT provsql.rv_aggregate_semimod(prov, provsql.as_random(1::double precision));
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Value-aware presence indicator: NULL when the row's aggregated
 *        value is NULL.
 *
 * SQL aggregates skip NULL inputs, so a NULL @c random_variable cell must
 * not count in @c avg's denominator: the wrap yields NULL (which the
 * @c sum fold skips) exactly when the value is NULL, and the plain
 * one-argument indicator otherwise.  The planner-hook @c avg rewrite
 * emits this form; the one-argument indicator remains for the internal
 * public-form defaults.
 */
CREATE OR REPLACE FUNCTION rv_aggregate_indicator(prov UUID, rv random_variable)
  RETURNS random_variable AS
$$
  SELECT CASE WHEN rv IS NULL THEN NULL
              ELSE provsql.rv_aggregate_indicator(prov) END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief State-transition function for @c sum(random_variable).
 *
 * Appends the input RV's UUID to the running array.  NULL inputs are
 * skipped (matching standard SUM semantics).  The aggregate's INITCOND
 * is @c '{}' so the FINALFUNC always runs and can tell an empty group
 * (state @c '{}') apart from a group whose every input was NULL -- both
 * of which SQL reports as @c NULL.
 */
CREATE OR REPLACE FUNCTION sum_rv_sfunc(
  state UUID[], rv random_variable)
  RETURNS UUID[] AS
$$
  SELECT CASE
    WHEN rv IS NULL THEN state
    ELSE array_append(state, (rv)::UUID)
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Final function for @c sum(random_variable): build a
 *        @c gate_arith PLUS root.
 *
 * Empty group (@c state = @c '{}'): return @c NULL, as SQL's @c sum
 * does over zero rows -- the same answer the @c AGG_TOKEN path gives
 * for @c sum over an empty aggregation.
 *
 * Singleton group: return the single child directly without minting a
 * useless single-child @c gate_arith.
 *
 * Otherwise: build @c gate_arith(PLUS, state) via @c provenance_arith.
 */
CREATE OR REPLACE FUNCTION sum_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  arith_token UUID;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(state, 1) = 1 THEN
    RETURN provsql.random_variable_make(state[1]);
  END IF;
  arith_token := provsql.provenance_arith(0, state);  -- 0 = PROVSQL_ARITH_PLUS
  RETURN provsql.random_variable_make(arith_token);
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE sum(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = sum_rv_ffunc
);
 
/**
 * @brief Numerator final function for the @c avg rewrite: @c sum,
 *        @c NULL on an empty group.
 *
 * Behaviourally identical to @ref sum_rv_ffunc; kept as a separate
 * catalog entry because the @c avg rewrite names it explicitly.  The
 * planner-hook @c avg rewrite emits
 * @c rv_sum_or_null(rv_aggregate_semimod(prov, x)) @c /
 * @c sum(rv_aggregate_indicator(prov)); @c random_variable_div is
 * @c STRICT, so an empty group propagates the numerator's @c NULL and
 * @c avg is @c NULL -- the standard SQL @c AVG convention -- while a
 * non-empty group behaves exactly like @c sum.
 */
CREATE OR REPLACE FUNCTION rv_sum_or_null_ffunc(state UUID[])
  RETURNS random_variable AS
$$
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(state, 1) = 1 THEN
    RETURN provsql.random_variable_make(state[1]);
  END IF;
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(0, state));  -- 0 = PROVSQL_ARITH_PLUS
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE rv_sum_or_null(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = rv_sum_or_null_ffunc
);
 
/**
 * @brief Final function for @c avg(random_variable).
 *
 * @c avg lifts the "@c AVG @c = @c SUM @c / @c COUNT" identity into the
 * @c random_variable algebra:
 * @f[
 *   \mathrm{AVG}(x) \;=\; \frac{\sum_i \mathbf{1}\{\varphi_i\} \cdot X_i}
 *                                {\sum_i \mathbf{1}\{\varphi_i\}}.
 * @f]
 * In a provenance-tracked query the planner-hook rewrites @c avg(x) into
 * @c rv_sum_or_null(rv_aggregate_semimod(prov, x)) @c /
 * @c sum(rv_aggregate_indicator(prov)) (see
 * @c make_rv_aggregate_expression), so both the numerator and the
 * provenance-weighted count denominator are built by @c sum's fold and no
 * gate is inspected.  This FFUNC is therefore reached only on an
 * @em untracked call, where every row is unconditionally present: the
 * numerator is @c sum over the raw per-row RVs and the denominator is the
 * plain row count @c n (each row contributing @c as_random(1)).
 *
 * Empty group: returns @c NULL, matching standard SQL @c AVG (and unlike
 * @c sum, whose empty group is the additive identity @c as_random(0)):
 * the caller cannot otherwise disambiguate "0 rows" from "rows summing
 * to 0".
 */
CREATE OR REPLACE FUNCTION avg_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  n INTEGER;
  i INTEGER;
  num_token UUID;
  denom_token UUID;
  denom_state UUID[] := '{}';
  one_uuid UUID;
BEGIN
  IF state IS NULL THEN
    RETURN NULL;
  END IF;
  n := array_length(state, 1);
  IF n IS NULL THEN
    RETURN NULL;
  END IF;
 
  one_uuid := (provsql.as_random(1::double precision))::UUID;
  FOR i IN 1..n LOOP
    denom_state := array_append(denom_state, one_uuid);
  END LOOP;
 
  IF n = 1 THEN
    num_token := state[1];
    denom_token := denom_state[1];
  ELSE
    num_token := provsql.provenance_arith(0, state);          -- 0 = PLUS
    denom_token := provsql.provenance_arith(0, denom_state);  -- 0 = PLUS
  END IF;
 
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(
      3,  -- 3 = PROVSQL_ARITH_DIV
      ARRAY[num_token, denom_token]));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE avg(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = avg_rv_ffunc
);
 
/**
 * @brief Final function for @c product(random_variable): fold a
 *        @c gate_arith TIMES root over the per-row contributions.
 *
 * Multiplicative analogue of @c sum(random_variable):
 * @f[
 *   \mathrm{PRODUCT}(x) \;=\; \prod_i \big(\mathbf{1}\{\varphi_i\} \cdot X_i
 *                                          + \mathbf{1}\{\neg\varphi_i\} \cdot 1\big)
 *                       \;=\; \prod_{i : \varphi_i} X_i.
 * @f]
 * Each per-row contribution already carries the multiplicative identity
 * as its absent-row value: a provenance-tracked query wraps the argument
 * as @c mixture(prov_i, X_i, as_random(1)) (identity baked in by the
 * three-argument @ref rv_aggregate_semimod), and an untracked call passes
 * the raw RV through.  So the FFUNC is a plain fold with no gate
 * inspection: @c gate_arith(TIMES, state).
 *
 * Reuses @c sum_rv_sfunc as the state-transition function.  Empty group:
 * @c NULL, by symmetry with @c sum / @c avg / @c min / @c max, which take
 * it from their standard-SQL counterparts.  The multiplicative identity
 * @c as_random(1) stays the absent-row value inside the fold, where it
 * belongs; a group containing no row has no product to report.
 * Singleton group: the single child directly, without a one-child TIMES
 * root.
 */
CREATE OR REPLACE FUNCTION product_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(state, 1) = 1 THEN
    RETURN provsql.random_variable_make(state[1]);
  END IF;
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(1, state));  -- 1 = PROVSQL_ARITH_TIMES
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE product(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = product_rv_ffunc
);
 
/**
 * @brief Final function for @c max(random_variable) / @c min(random_variable):
 *        fold a @c gate_arith MAX / MIN root over the per-row contributions.
 *
 * The order-statistic analogues of @c sum / @c product:
 * @f[
 *   \mathrm{MAX}(x) = \max_{i : \varphi_i} X_i, \qquad
 *   \mathrm{MIN}(x) = \min_{i : \varphi_i} X_i.
 * @f]
 * A row absent in a world (its provenance @f$\varphi_i@f$ false) must not
 * perturb the extremum, so it contributes the order-statistic identity
 * @f$\mp\infty@f$.  That identity is baked into each per-row contribution
 * upstream: a provenance-tracked query wraps the argument as
 * @c mixture(prov_i, X_i, as_random(∓∞)) (via the three-argument
 * @ref rv_aggregate_semimod), and an untracked call passes the raw RV
 * through.  So the FFUNC is a plain fold with no gate inspection:
 * @c gate_arith(@p op, state).
 *
 * Empty group: @c NULL, as SQL's @c min / @c max report over zero rows.
 * @p identity belongs to the catalog signature and describes the per-row
 * absent contribution baked in upstream; the empty group does not consult
 * it, since @f$\mp\infty@f$ is an artefact of the fold rather than a value
 * the group actually contains.
 * Singleton group: the single child directly.
 */
CREATE OR REPLACE FUNCTION extremum_rv_ffunc(
  state UUID[], op INTEGER, identity double precision)
  RETURNS random_variable AS
$$
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  IF array_length(state, 1) = 1 THEN
    RETURN provsql.random_variable_make(state[1]);
  END IF;
  RETURN provsql.random_variable_make(
    provsql.provenance_arith(op, state));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION max_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
  -- 5 = PROVSQL_ARITH_MAX; empty-group / row-absent identity -inf.
  SELECT provsql.extremum_rv_ffunc(state, 5, '-Infinity'::double precision);
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
CREATE OR REPLACE FUNCTION min_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
  -- 6 = PROVSQL_ARITH_MIN; empty-group / row-absent identity +inf.
  SELECT provsql.extremum_rv_ffunc(state, 6, 'Infinity'::double precision);
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE max(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = max_rv_ffunc
);
 
CREATE AGGREGATE min(random_variable) (
  SFUNC     = sum_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = min_rv_ffunc
);
 
-- ---------------------------------------------------------------------
-- SQL-standard statistic aggregates over random_variable rows:
-- covar_pop / covar_samp / corr (two-argument), stddev_pop / stddev_samp
-- (one-argument), and the ordered-set percentile_cont.
--
-- Row presence is carried by a per-row 0/1 indicator RV: the public
-- aggregates use the certain indicator as_random(1) (every row present),
-- and a provenance-tracked query is rewritten by the planner hook
-- (make_rv_aggregate_expression) to the rv_*_impl aggregates whose extra
-- leading argument is rv_aggregate_indicator(prov), so a row absent in a
-- world drops out of every sum, the count, and the percentile member set.
-- The moment statistics are built from indicator-weighted power sums with
-- existing gate_arith opcodes (e.g. covar_pop = SXY/N - (SX/N)(SY/N)); a
-- world where the statistic is undefined (N = 0, or N = 1 for the sample
-- forms) evaluates to NaN, the established undefined-world convention the
-- moment estimators skip.  percentile_cont is the one gate the arithmetic
-- cannot express: it mints the PROVSQL_ARITH_PERCENTILE gate_arith
-- (interleaved [ind_1, x_1, ...] wires, fraction in extra) that the Monte
-- Carlo sampler evaluates by sorting each draw's present values and
-- interpolating.
-- ---------------------------------------------------------------------
 
/** @brief State transition for the one-argument RV statistic aggregates
 *  (@c stddev_pop / @c stddev_samp): append the certain indicator and the
 *  row's RV as a pair.  NULL rows are skipped (standard SQL). */
CREATE OR REPLACE FUNCTION rv_stat1_sfunc(state UUID[], x random_variable)
  RETURNS UUID[] AS
$$
  SELECT CASE
    WHEN x IS NULL THEN state
    ELSE state || ARRAY[(provsql.as_random(1::double precision))::UUID,
                        (x)::UUID]
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/** @brief State transition for the two-argument RV statistic aggregates
 *  (@c covar_pop / @c covar_samp / @c corr): append the certain indicator
 *  and the row's RV pair as a triple.  Rows with either side NULL are
 *  skipped (standard SQL covariance semantics). */
CREATE OR REPLACE FUNCTION rv_stat2_sfunc(
  state UUID[], x random_variable, y random_variable)
  RETURNS UUID[] AS
$$
  SELECT CASE
    WHEN x IS NULL OR y IS NULL THEN state
    ELSE state || ARRAY[(provsql.as_random(1::double precision))::UUID,
                        (x)::UUID, (y)::UUID]
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/** @brief Indicator-carrying state transition for the one-argument
 *  @c rv_*_impl statistic aggregates: the planner-hook rewrite passes the
 *  row's provenance indicator @c rv_aggregate_indicator(prov) as @p ind. */
CREATE OR REPLACE FUNCTION rv_stat1_impl_sfunc(
  state UUID[], ind random_variable, x random_variable)
  RETURNS UUID[] AS
$$
  SELECT CASE
    WHEN x IS NULL THEN state
    ELSE state || ARRAY[coalesce((ind)::UUID,
                          (provsql.as_random(1::double precision))::UUID),
                        (x)::UUID]
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/** @brief Indicator-carrying state transition for the two-argument
 *  @c rv_*_impl statistic aggregates. */
CREATE OR REPLACE FUNCTION rv_stat2_impl_sfunc(
  state UUID[], ind random_variable, x random_variable, y random_variable)
  RETURNS UUID[] AS
$$
  SELECT CASE
    WHEN x IS NULL OR y IS NULL THEN state
    ELSE state || ARRAY[coalesce((ind)::UUID,
                          (provsql.as_random(1::double precision))::UUID),
                        (x)::UUID, (y)::UUID]
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Mint the indicator-weighted power-sum gates shared by the
 *        covariance / stddev final functions.
 *
 * @p state is the flat interleaved aggregate state -- pairs
 * @c [ind, x, ...] (@p stride 2) or triples @c [ind, x, y, ...]
 * (@p stride 3).  Emits @c gate_arith tokens for
 * @f$N = \sum_i \mathbf{1}_i@f$, @f$SX = \sum_i \mathbf{1}_i x_i@f$,
 * @f$SXX = \sum_i \mathbf{1}_i x_i^2@f$ and, at stride 3, @f$SY@f$,
 * @f$SXY@f$, @f$SYY@f$.  The per-row indicator gate is shared between
 * @f$N@f$ and every product it weighs, so the Monte Carlo per-iteration
 * cache keeps the row's presence coupled across all the sums (and a
 * repeated child @c [ind, x, x] reuses the same draw of @c x, giving
 * @f$x^2@f$, not two independent draws).
 */
CREATE OR REPLACE FUNCTION rv_stat_sum_tokens(
  state UUID[], stride INTEGER,
  OUT n_tok UUID, OUT sx_tok UUID, OUT sxx_tok UUID,
  OUT sy_tok UUID, OUT sxy_tok UUID, OUT syy_tok UUID)
AS
$$
DECLARE
  nrows INTEGER := coalesce(array_length(state, 1), 0) / stride;
  inds UUID[] := '{}';
  xs   UUID[] := '{}';
  xxs  UUID[] := '{}';
  ys   UUID[] := '{}';
  xys  UUID[] := '{}';
  yys  UUID[] := '{}';
  ind UUID;
  x UUID;
  y UUID;
BEGIN
  FOR i IN 1..nrows LOOP
    ind := state[(i-1) * stride + 1];
    x   := state[(i-1) * stride + 2];
    inds := array_append(inds, ind);
    xs   := array_append(xs,  provenance_arith(1, ARRAY[ind, x]));
    xxs  := array_append(xxs, provenance_arith(1, ARRAY[ind, x, x]));
    IF stride = 3 THEN
      y := state[(i-1) * stride + 3];
      ys  := array_append(ys,  provenance_arith(1, ARRAY[ind, y]));
      xys := array_append(xys, provenance_arith(1, ARRAY[ind, x, y]));
      yys := array_append(yys, provenance_arith(1, ARRAY[ind, y, y]));
    END IF;
  END LOOP;
  n_tok   := provenance_arith(0, inds);
  sx_tok  := provenance_arith(0, xs);
  sxx_tok := provenance_arith(0, xxs);
  IF stride = 3 THEN
    sy_tok  := provenance_arith(0, ys);
    sxy_tok := provenance_arith(0, xys);
    syy_tok := provenance_arith(0, yys);
  END IF;
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Population-variance gate @f$SXX/N - (SX/N)^2@f$ from the
 *  power-sum tokens. */
CREATE OR REPLACE FUNCTION rv_stat_var_pop_token(
  n_tok UUID, s_tok UUID, ss_tok UUID)
  RETURNS UUID AS
$$
  SELECT provsql.provenance_arith(2, ARRAY[
    provsql.provenance_arith(3, ARRAY[ss_tok, n_tok]),
    provsql.provenance_arith(1, ARRAY[
      provsql.provenance_arith(3, ARRAY[s_tok, n_tok]),
      provsql.provenance_arith(3, ARRAY[s_tok, n_tok])])]);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Sample-variance gate @f$(SXX - SX^2/N) / (N - 1)@f$ from the
 *  power-sum tokens (NaN in a world with @f$N \le 1@f$, the undefined-world
 *  convention). */
CREATE OR REPLACE FUNCTION rv_stat_var_samp_token(
  n_tok UUID, s_tok UUID, ss_tok UUID)
  RETURNS UUID AS
$$
  SELECT provsql.provenance_arith(3, ARRAY[
    provsql.provenance_arith(2, ARRAY[
      ss_tok,
      provsql.provenance_arith(3, ARRAY[
        provsql.provenance_arith(1, ARRAY[s_tok, s_tok]), n_tok])]),
    provsql.provenance_arith(2, ARRAY[
      n_tok, (provsql.as_random(1::double precision))::UUID])]);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief @f$\sqrt{\max(v, 0)}@f$ gate over a variance token: the max-clamp
 *  removes the tiny negative values float error can produce (variance is
 *  mathematically non-negative), so the POW domain guard never fires. */
CREATE OR REPLACE FUNCTION rv_stat_sqrt_token(v_tok UUID)
  RETURNS UUID AS
$$
  SELECT provsql.provenance_arith(7, ARRAY[
    provsql.provenance_arith(5, ARRAY[
      v_tok, (provsql.as_random(0::double precision))::UUID]),
    (provsql.as_random(0.5::double precision))::UUID]);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Population-covariance gate @f$SXY/N - (SX/N)(SY/N)@f$ from the
 *  power-sum tokens. */
CREATE OR REPLACE FUNCTION rv_stat_covar_pop_token(
  n_tok UUID, sx_tok UUID, sy_tok UUID, sxy_tok UUID)
  RETURNS UUID AS
$$
  SELECT provsql.provenance_arith(2, ARRAY[
    provsql.provenance_arith(3, ARRAY[sxy_tok, n_tok]),
    provsql.provenance_arith(1, ARRAY[
      provsql.provenance_arith(3, ARRAY[sx_tok, n_tok]),
      provsql.provenance_arith(3, ARRAY[sy_tok, n_tok])])]);
$$ LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Final function for @c covar_pop(random_variable, random_variable). */
CREATE OR REPLACE FUNCTION covar_pop_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  t RECORD;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT * INTO t FROM rv_stat_sum_tokens(state, 3);
  RETURN random_variable_make(
    rv_stat_covar_pop_token(t.n_tok, t.sx_tok, t.sy_tok, t.sxy_tok));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Final function for @c covar_samp(random_variable, random_variable):
 *  @f$(SXY - SX\,SY/N) / (N-1)@f$. */
CREATE OR REPLACE FUNCTION covar_samp_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  t RECORD;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT * INTO t FROM rv_stat_sum_tokens(state, 3);
  RETURN random_variable_make(
    provenance_arith(3, ARRAY[
      provenance_arith(2, ARRAY[
        t.sxy_tok,
        provenance_arith(3, ARRAY[
          provenance_arith(1, ARRAY[t.sx_tok, t.sy_tok]), t.n_tok])]),
      provenance_arith(2, ARRAY[
        t.n_tok, (as_random(1::double precision))::UUID])]));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Final function for @c corr(random_variable, random_variable):
 *  @f$\mathrm{covar\_pop} / \sqrt{\max(v_x v_y, 0)}@f$ (a zero-variance
 *  world divides to @f$\pm\infty@f$ / NaN, the undefined-world convention,
 *  matching SQL's NULL for a zero-stddev input). */
CREATE OR REPLACE FUNCTION corr_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  t RECORD;
  vx UUID;
  vy UUID;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT * INTO t FROM rv_stat_sum_tokens(state, 3);
  vx := rv_stat_var_pop_token(t.n_tok, t.sx_tok, t.sxx_tok);
  vy := rv_stat_var_pop_token(t.n_tok, t.sy_tok, t.syy_tok);
  RETURN random_variable_make(
    provenance_arith(3, ARRAY[
      rv_stat_covar_pop_token(t.n_tok, t.sx_tok, t.sy_tok, t.sxy_tok),
      rv_stat_sqrt_token(provenance_arith(1, ARRAY[vx, vy]))]));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Final function for @c stddev_pop(random_variable). */
CREATE OR REPLACE FUNCTION stddev_pop_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  t RECORD;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT * INTO t FROM rv_stat_sum_tokens(state, 2);
  RETURN random_variable_make(
    rv_stat_sqrt_token(
      rv_stat_var_pop_token(t.n_tok, t.sx_tok, t.sxx_tok)));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief Final function for @c stddev_samp(random_variable). */
CREATE OR REPLACE FUNCTION stddev_samp_rv_ffunc(state UUID[])
  RETURNS random_variable AS
$$
DECLARE
  t RECORD;
BEGIN
  IF state IS NULL OR array_length(state, 1) IS NULL THEN
    RETURN NULL;
  END IF;
  SELECT * INTO t FROM rv_stat_sum_tokens(state, 2);
  RETURN random_variable_make(
    rv_stat_sqrt_token(
      rv_stat_var_samp_token(t.n_tok, t.sx_tok, t.sxx_tok)));
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
CREATE AGGREGATE covar_pop(random_variable, random_variable) (
  SFUNC     = rv_stat2_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = covar_pop_rv_ffunc
);
 
CREATE AGGREGATE covar_samp(random_variable, random_variable) (
  SFUNC     = rv_stat2_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = covar_samp_rv_ffunc
);
 
CREATE AGGREGATE corr(random_variable, random_variable) (
  SFUNC     = rv_stat2_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = corr_rv_ffunc
);
 
CREATE AGGREGATE stddev_pop(random_variable) (
  SFUNC     = rv_stat1_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = stddev_pop_rv_ffunc
);
 
CREATE AGGREGATE stddev_samp(random_variable) (
  SFUNC     = rv_stat1_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = stddev_samp_rv_ffunc
);
 
-- The indicator-carrying rewrite targets (planner hook only; never called
-- directly by users).
 
CREATE AGGREGATE rv_covar_pop_impl(
  random_variable, random_variable, random_variable) (
  SFUNC     = rv_stat2_impl_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = covar_pop_rv_ffunc
);
 
CREATE AGGREGATE rv_covar_samp_impl(
  random_variable, random_variable, random_variable) (
  SFUNC     = rv_stat2_impl_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = covar_samp_rv_ffunc
);
 
CREATE AGGREGATE rv_corr_impl(
  random_variable, random_variable, random_variable) (
  SFUNC     = rv_stat2_impl_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = corr_rv_ffunc
);
 
CREATE AGGREGATE rv_stddev_pop_impl(random_variable, random_variable) (
  SFUNC     = rv_stat1_impl_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = stddev_pop_rv_ffunc
);
 
CREATE AGGREGATE rv_stddev_samp_impl(random_variable, random_variable) (
  SFUNC     = rv_stat1_impl_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = stddev_samp_rv_ffunc
);
 
/**
 * @brief Mint the @c PROVSQL_ARITH_PERCENTILE gate: the continuous
 *        percentile (SQL @c percentile_cont) over a group of RV rows.
 *
 * @p pairs is the interleaved wire list @c [ind_1, x_1, ..., ind_n, x_n]
 * (each @p ind_i a 0/1 presence-indicator RV).  The @p fraction is
 * TEXT-encoded in the gate's @c extra and participates in the token UUID
 * (two percentiles of the same group at different fractions are distinct
 * gates).  Per Monte Carlo draw, the sampler collects the values whose
 * indicator draws 1, sorts them, and linearly interpolates at the
 * fraction; a draw with no present row is NaN (undefined world).
 */
CREATE OR REPLACE FUNCTION rv_percentile_make(fraction double precision,
                                              pairs UUID[])
  RETURNS random_variable AS
$$
DECLARE
  token UUID;
BEGIN
  IF fraction IS NULL THEN
    RETURN NULL;
  END IF;
  IF fraction < 0 OR fraction > 1 THEN
    RAISE EXCEPTION
      'percentile_cont: fraction must be between 0 and 1 (got %)', fraction;
  END IF;
  token := public.uuid_generate_v5(
    uuid_ns_provsql(),
    concat('arith', '10', pairs::TEXT, fraction::TEXT));
  PERFORM create_gate(token, 'arith', pairs);
  PERFORM set_infos(token, 10);  -- 10 = PROVSQL_ARITH_PERCENTILE
  PERFORM set_extra(token, fraction::TEXT);
  RETURN random_variable_make(token);
END
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/** @brief State transition for the public ordered-set
 *  @c percentile_cont(float8) WITHIN GROUP (ORDER BY random_variable):
 *  append the certain indicator and the row's RV.  Only reachable on
 *  untracked input (a provenance-tracked query is rewritten to
 *  @c rv_percentile_impl before planning), where the sort over
 *  @c random_variable raises the ordering-is-meaningless diagnostic
 *  first -- so in practice this runs only for empty input. */
CREATE OR REPLACE FUNCTION percentile_cont_rv_sfunc(
  state UUID[], x random_variable)
  RETURNS UUID[] AS
$$
  SELECT provsql.rv_stat1_sfunc(state, x);
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/** @brief Final function for the public ordered-set @c percentile_cont:
 *  receives the direct @p fraction argument after the state. */
CREATE OR REPLACE FUNCTION percentile_cont_rv_ffunc(
  state UUID[], fraction double precision)
  RETURNS random_variable AS
$$
  SELECT CASE
    WHEN state IS NULL OR array_length(state, 1) IS NULL THEN NULL
    ELSE provsql.rv_percentile_make(fraction, state)
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE percentile_cont(double precision ORDER BY random_variable) (
  SFUNC     = percentile_cont_rv_sfunc,
  STYPE     = UUID[],
  INITCOND  = '{}',
  FINALFUNC = percentile_cont_rv_ffunc
);
 
/** @brief Transition state for @c rv_percentile_impl: the fraction (from
 *  the first row) plus the interleaved indicator/value token pairs. */
CREATE TYPE rv_percentile_state AS (
  fraction double precision,
  tokens UUID[]
);
 
/** @brief State transition for @c rv_percentile_impl, the planner-hook
 *  rewrite target of a provenance-tracked @c percentile_cont: stashes the
 *  (group-constant) fraction and appends the indicator/value pair. */
CREATE OR REPLACE FUNCTION rv_percentile_impl_sfunc(
  state rv_percentile_state, fraction double precision,
  ind random_variable, x random_variable)
  RETURNS rv_percentile_state AS
$$
  SELECT ROW(
    coalesce((state).fraction, fraction),
    CASE
      WHEN x IS NULL THEN (state).tokens
      ELSE (state).tokens ||
           ARRAY[coalesce((ind)::UUID,
                   (provsql.as_random(1::double precision))::UUID),
                 (x)::UUID]
    END)::provsql.rv_percentile_state;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/** @brief Final function for @c rv_percentile_impl. */
CREATE OR REPLACE FUNCTION rv_percentile_impl_ffunc(state rv_percentile_state)
  RETURNS random_variable AS
$$
  SELECT CASE
    WHEN state IS NULL OR array_length((state).tokens, 1) IS NULL THEN NULL
    ELSE provsql.rv_percentile_make((state).fraction, (state).tokens)
  END;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
CREATE AGGREGATE rv_percentile_impl(
  double precision, random_variable, random_variable) (
  SFUNC     = rv_percentile_impl_sfunc,
  STYPE     = rv_percentile_state,
  INITCOND  = '(,"{}")',
  FINALFUNC = rv_percentile_impl_ffunc
);
 
/** @} */
 
/** @} */
 
/** @} */
 
/** @defgroup aggregate_provenance Aggregate provenance
 *  Functions for building and evaluating aggregate (GROUP BY) provenance,
 *  including the δ-semiring operator and semimodule multiplication.
 *  @{
 */
 
/**
 * @brief Create a δ-semiring gate wrapping a provenance token
 *
 * Used internally for aggregate provenance. Returns the token unchanged
 * if it is gate_zero() or gate_one(), and gate_one() if the token is NULL.
 */
CREATE OR REPLACE FUNCTION provenance_delta
  (token UUID)
  RETURNS UUID AS
$$
DECLARE
  delta_token UUID;
BEGIN
  -- NULL token ≡ 1 (untracked source), and δ(1) = 1.  Tested first: the
  -- equality comparisons below are not NULL-safe.
  IF token IS NULL THEN
    return gate_one();
  END IF;
 
  IF token = gate_zero() OR token = gate_one() THEN
    return token;
  END IF;
 
  delta_token:=uuid_generate_v5(uuid_ns_provsql(),concat('delta',token));
 
  PERFORM create_gate(delta_token,'delta',ARRAY[token::UUID]);
 
  RETURN delta_token;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SECURITY DEFINER PARALLEL SAFE IMMUTABLE;
 
/**
 * @brief Build an aggregate provenance gate from grouped tokens
 *
 * Called internally by the query rewriter for GROUP BY queries.
 * Creates an agg gate linking all contributing tokens and records
 * the aggregate function OID and the computed scalar value.
 *
 * @param aggfnoid OID of the SQL aggregate function
 * @param aggtype OID of the aggregate result type
 * @param val computed aggregate value
 * @param tokens array of provenance tokens being aggregated
 * @param is_scalar true for a scalar (no GROUP BY) aggregation, whose
 *        output row exists even when no tuple is present; stored in the
 *        high bit of info2
 */
CREATE OR REPLACE FUNCTION provenance_aggregate(
    aggfnoid INTEGER,
    aggtype INTEGER,
    val ANYELEMENT,
    tokens UUID[],
    is_scalar BOOLEAN DEFAULT false)
  RETURNS AGG_TOKEN AS
$$
DECLARE
  c INTEGER;
  agg_tok UUID;
  agg_val varchar;
BEGIN
  -- Drop the NULL placeholders array_agg keeps for rows that did not produce a
  -- semimod gate (provenance_semimod returns NULL for a NULL aggregated value),
  -- so a NULL input never participates in the aggregate.
  tokens := array_remove(tokens, NULL);
  c:=COALESCE(array_length(tokens, 1), 0);
 
  agg_val = CAST(val as VARCHAR);
 
  IF c = 0 THEN
    agg_tok := gate_zero();
  ELSE
    -- aggfnoid must be part of the UUID: SUM(id) and AVG(id) over the
    -- same children would otherwise collapse to a single gate, and
    -- their concurrent set_infos calls would overwrite each other's
    -- aggregation operator (resulting in the wrong agg_kind being
    -- read by provsql_having under cross-backend contention).  The
    -- scalar-aggregation flag must likewise be hashed: a scalar and a
    -- grouped aggregate over identical children carry different info2 and
    -- must stay distinct gates, else the concurrent set_infos calls would
    -- clobber the flag.  The flag is stored in the high bit of info2 (the
    -- low 31 bits keep the result-type OID); aggtype itself is passed clean
    -- so the AGG_TOKEN->scalar cast still finds a valid type.
    agg_tok := uuid_generate_v5(
      uuid_ns_provsql(),
      concat('agg',aggfnoid,tokens,CASE WHEN is_scalar THEN 'S' ELSE '' END));
    PERFORM create_gate(agg_tok, 'agg', tokens);
    PERFORM set_infos(agg_tok, aggfnoid,
                      CASE WHEN is_scalar THEN aggtype | (-2147483648) ELSE aggtype END);
    PERFORM set_extra(agg_tok, agg_val);
  END IF;
 
  RETURN '( '||agg_tok||' , '||agg_val||' )';
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql,pg_temp,public SECURITY DEFINER IMMUTABLE;
 
/**
 * @brief Create a semimodule scalar multiplication gate
 *
 * Pairs a scalar value with a provenance token, used internally by
 * the query rewriter for aggregate provenance.
 *
 * @param val the scalar value
 * @param token the provenance token to multiply
 */
CREATE OR REPLACE FUNCTION provenance_semimod(val ANYELEMENT, token UUID)
  RETURNS UUID AS
$$
DECLARE
  semimod_token UUID;
  value_token UUID;
BEGIN
  -- A NULL value means this row does not participate in the aggregate (SQL
  -- aggregates ignore NULL inputs; only count(*) counts rows unconditionally,
  -- and it passes a constant 1 here).  Produce no semimod gate so the row is
  -- skipped when provenance_aggregate builds the agg gate.
  IF val IS NULL THEN
    RETURN NULL;
  END IF;
 
  SELECT uuid_generate_v5(uuid_ns_provsql(),concat('value',CAST(val AS VARCHAR)))
    INTO value_token;
  SELECT uuid_generate_v5(uuid_ns_provsql(),concat('semimod',value_token,token))
    INTO semimod_token;
 
  --create value gates
  PERFORM create_gate(value_token,'value');
  PERFORM set_extra(value_token, CAST(val AS VARCHAR));
 
  --create semimod gate
  PERFORM create_gate(semimod_token,'semimod',ARRAY[token::UUID,value_token]);
 
  RETURN semimod_token;
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql,pg_temp,public SECURITY DEFINER IMMUTABLE;
 
/** @} */
 
/** @defgroup probability Probability and Shapley values
 *  Functions for computing probabilities, expected values, and
 *  game-theoretic contribution measures (Shapley/Banzhaf values)
 *  from provenance circuits.
 *  @{
 */
 
/**
 * @brief Compute the probability of a provenance token
 *
 * Compiles the provenance circuit to d-DNNF and evaluates the
 * probability. The compilation method can be selected explicitly.
 *
 * @ref probability() "probability" is a shorter alias bound to the same C symbol, so
 * @c probability(token) is exactly @c probability_evaluate(token); it is
 * usually preferable, and additionally carries a @c (BOOLEAN) predicate
 * overload (e.g. @c probability(x @c > @c y)).
 *
 * @param token provenance token to evaluate
 * @param method knowledge compilation method (NULL for default)
 * @param arguments additional arguments for the method
 */
CREATE OR REPLACE FUNCTION probability_evaluate(
  token UUID,
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS
  'provsql','probability_evaluate' LANGUAGE C STABLE;
 
/**
 * @brief Short alias of @ref probability_evaluate.
 *
 * Bound to the same C symbol as @ref probability_evaluate, so
 * @c probability(token) is exactly @c probability_evaluate(token).
 * Provided to match the concise polymorphic surface of @ref expected,
 * @ref variance, and @ref support "support": callers are not forced to
 * spell out @c probability_evaluate.
 *
 * @param token provenance token to evaluate
 * @param method knowledge compilation method (NULL for default)
 * @param arguments additional arguments for the method
 */
CREATE OR REPLACE FUNCTION probability(
  token UUID,
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS
  'provsql','probability_evaluate' LANGUAGE C STABLE;
 
/**
 * @brief Probability of a Boolean event over random variables.
 *
 * The @c (BOOLEAN) overload of @c probability lets a query ask for the
 * probability of an event with the natural infix grammar, e.g.
 * @c probability(x @c > @c y @c AND @c x @c < @c z).  When the argument
 * carries a probabilistic (random_variable / aggregate) comparison, the
 * planner hook intercepts the call and rewrites it into
 * @c probability_evaluate over the argument's event token (a @c gate_cmp /
 * Boolean combination); the body below is then never reached.
 *
 * When the argument is a purely deterministic Boolean (no probabilistic
 * comparison) the hook leaves the call alone and the body runs, so the
 * probability of a definite event is simply @c 1 when it holds and @c 0 when
 * it does not (@c NULL propagates).  This makes @c probability total over
 * Booleans -- @c probability(1 @c > @c 0) is @c 1, @c probability(region @c =
 * @c 'north') is a per-row @c 0/1 -- and it works even with
 * @c provsql.active off.  @c NOT strict so a default-NULL @c method does not
 * short-circuit the cast.
 *
 * The predicate surface deliberately lives only on the short @c probability
 * name, not on @c probability_evaluate: a Boolean overload of the latter
 * would make @c probability_evaluate('<UUID-as-TEXT>') ambiguous (an unknown
 * literal matches both the @c UUID and the @c BOOLEAN overload), breaking
 * existing string-literal callers.  @c probability is new, so it carries the
 * predicate overload without that hazard.
 */
CREATE OR REPLACE FUNCTION probability(
  predicate BOOLEAN,
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS
$$
  SELECT predicate::INTEGER::double precision;
$$ LANGUAGE sql IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Cheap certified probability interval of a DNF-shaped circuit.
 *
 * Returns @c [lower,upper] with @c lower <= probability_evaluate(token) <=
 * @c upper, computed without compiling the circuit (the Olteanu-Huang d-tree
 * leaf bound).  Errors when @p token is not a monotone DNF over input leaves.
 */
CREATE OR REPLACE FUNCTION probability_bounds(
  token UUID,
  OUT lower DOUBLE PRECISION,
  OUT upper DOUBLE PRECISION) AS
  'provsql','probability_bounds' LANGUAGE C STABLE;
 
/**
 * @brief Compute the expected value of a probabilistic scalar
 *
 * Computes E[input | prov] for either an @c AGG_TOKEN (discrete
 * SUM/MIN/MAX aggregation over Boolean-input gate_agg circuits, with
 * @c prov as the Boolean conditioning event) or a @c random_variable
 * (continuous distribution, traversed by the analytical / MC
 * evaluator from @c Expectation.cpp).
 *
 * Implementation: thin wrapper over @c moment(input, 1, prov, method,
 * arguments).  Both branches converge on the same machinery; the
 * AGG_TOKEN side computes E[X] as the @f$k=1@f$ instance of the
 * @f$n^k@f$-tuple enumeration in @c agg_raw_moment, the
 * random_variable side calls @c compute_expectation through
 * @c rv_moment.
 *
 * @param input aggregate expression or random variable to compute E[·] of
 * @param prov provenance condition (defaults to gate_one(), i.e., unconditional)
 * @param method knowledge compilation method (AGG_TOKEN path only)
 * @param arguments additional arguments for the method (AGG_TOKEN path only)
 */
CREATE OR REPLACE FUNCTION expected(
  input ANYELEMENT,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
  SELECT moment(input, 1, prov, method, arguments);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Internal: shared C entry point for variance / moment / central_moment.
 *
 * The @c expected() SQL function reaches the Expectation evaluator
 * through @c provenance_evaluate_compiled(..., 'expectation', ...).
 * The variance / raw-moment / central-moment SQL functions need an
 * extra @p k INTEGER argument that does not fit that dispatcher's
 * signature, so they go through this dedicated entry point.  Returns
 * E[X^k] when @p central is FALSE, or E[(X - E[X])^k] when TRUE.
 */
CREATE OR REPLACE FUNCTION rv_moment(
  token UUID, k INTEGER, central BOOLEAN,
  prov UUID DEFAULT gate_one())
  RETURNS double precision
  AS 'provsql','rv_moment' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Exact E[AVG^k | COUNT >= 1] over independent rows (the joint
 *  (sum, count) fold); NULL when the shape is out of scope (shared
 *  leaves, compound contributors), signalling @c agg_raw_moment's avg
 *  arm to fall back to the Monte-Carlo scalar path. */
CREATE OR REPLACE FUNCTION agg_avg_moment_exact(token UUID, k INTEGER)
  RETURNS double precision
  AS 'provsql','agg_avg_moment_exact' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Collapsed (Rao-Blackwellised) raw moment E[C^k] of a correlated
 *  COUNT / SUM whose per-row selection events are coupled through a single
 *  shared continuous latent: 1-D quadrature over the latent, closed-form
 *  per-row CDF given it (O(G·n), exact up to the grid).  NULL when the
 *  circuit does not match the shared-latent pattern (caller falls back to
 *  the exact n^k enumeration).  k in {1, 2}. */
CREATE OR REPLACE FUNCTION agg_collapsed_moment(token UUID, k INTEGER)
  RETURNS double precision
  AS 'provsql','agg_collapsed_moment' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/** @brief Both collapsed raw moments {E[C], E[C^2]} of a correlated COUNT / SUM
 *  from a single circuit load and plan build; NULL when the shared-latent
 *  pattern does not match.  @c variance() uses this so a mean+variance readout
 *  traverses the circuit once rather than calling @c agg_collapsed_moment twice
 *  (the load and O(n) plan build dominate once the grid loop is arithmetic). */
CREATE OR REPLACE FUNCTION agg_collapsed_moments(token UUID)
  RETURNS double precision[]
  AS 'provsql','agg_collapsed_moments' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Boolean event "this aggregate-carrying gate's value is defined
 *        (non-NULL) in the world".
 *
 * Backs the conditional-on-defined convention of the aggregate moment
 * readouts: @c sum / @c count (and constants) have a value in every
 * world -- the empty group is the real value @c 0 -- so their defined
 * event is @c gate_one(); @c min / @c max / @c avg (and any other
 * aggregate) are @c NULL on an empty group, so their defined event is
 * "some contributing row is present", the OR of the semimod children's
 * row tokens; a @c case gate's value is defined iff its first-match
 * selected branch's value is (the same region walk as the moment
 * evaluator, conjoined per branch).  Anything else (an @c arith
 * composite, whose AGG_TOKEN running value is total) counts as always
 * defined.
 */
CREATE OR REPLACE FUNCTION agg_defined_event(token UUID)
  RETURNS UUID AS $$
DECLARE
  gt PROVENANCE_GATE := get_gate_type(token);
  fname varchar;
  toks UUID[];
  wires UUID[];
  nw INTEGER;
  m INTEGER;
  i INTEGER;
  running_neg UUID := gate_one();
  parts UUID[] := '{}';
BEGIN
  IF gt = 'agg' THEN
    SELECT proname INTO fname
      FROM pg_proc WHERE oid = (get_infos(token)).info1;
    IF fname IN ('sum', 'count') THEN
      RETURN gate_one();
    END IF;
    SELECT array_agg((get_children(c))[1]) INTO toks
      FROM unnest(get_children(token)) AS c;
    IF toks IS NULL THEN
      RETURN gate_zero();   -- structurally empty aggregate: never defined
    END IF;
    RETURN provenance_plus(toks);
  ELSIF gt = 'case' THEN
    wires := get_children(token);
    nw := array_length(wires, 1);
    m := (nw - 1) / 2;
    FOR i IN 1..m LOOP
      parts := parts || provenance_times(
        running_neg, wires[2 * i - 1],
        agg_defined_event(wires[2 * i]));
      running_neg := provenance_times(running_neg,
                                      provenance_not(wires[2 * i - 1]));
    END LOOP;
    parts := parts || provenance_times(running_neg,
                                       agg_defined_event(wires[nw]));
    RETURN provenance_plus(parts);
  END IF;
  -- value / arith / anything else: a value exists in every world.
  RETURN gate_one();
END
$$ LANGUAGE plpgsql STABLE STRICT PARALLEL SAFE
  SET search_path=provsql,pg_temp,public SECURITY DEFINER;
 
/**
 * @brief Compute the raw moment E[X^k | prov] of an AGG_TOKEN aggregate
 *
 * Sister of @c expected() for the AGG_TOKEN side of the polymorphic
 * @c moment / @c variance / @c central_moment dispatch.  Supports the
 * same aggregation functions as @c expected: SUM (which COUNT
 * normalises to at the gate level via @c Aggregation.cpp:322), MIN,
 * MAX, and AVG (exact over independent / laminar rows via the joint
 * (sum, count) distribution, Monte-Carlo scalar fallback otherwise).
 * MIN / MAX / AVG are NULL on an empty group, so their moments are
 * CONDITIONAL on the aggregate being defined -- NULL only when it never
 * is; SUM / COUNT treat the empty world as the real value 0.
 *
 * Strategy:
 * - <b>SUM</b>: with X = Σᵢ Iᵢ·vᵢ (Iᵢ the per-row inclusion indicator,
 *   vᵢ the row's value), expanding X^k and taking expectation gives
 *   @f$E[X^k] = \sum_{(i_1,\ldots,i_k) \in \{1..n\}^k} v_{i_1}\cdots v_{i_k}
 *               \cdot P(\bigwedge_{i \in \TEXT{distinct}(i_1..i_k)} I_i)@f$.
 *   We enumerate the @f$n^k@f$ tuples, conjoin the distinct inclusion
 *   tokens (and @p prov when conditioning), and evaluate the
 *   probability via @c probability_evaluate.
 * - <b>MIN / MAX</b>: replace @c v with @c v^k in the rank-based
 *   enumeration that @c expected already uses; @c MAX is handled by
 *   sign-flipping per the existing trick (negate vs.  rerank), with
 *   the outer multiplier becoming @f$(-1)^k@f$ instead of just @f$-1@f$.
 *
 * Cost: SUM is @f$O(n^k)@f$ probability evaluations -- tractable for
 * small @p k or small @p n; for larger sizes, prefer reaching for the
 * sampler.  MIN / MAX stay linear in @p n.
 */
CREATE OR REPLACE FUNCTION agg_raw_moment(
  token AGG_TOKEN,
  k INTEGER,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
DECLARE
  aggregation_function VARCHAR;
  child_pairs UUID[];
  pair_children UUID[];
  n INTEGER;
  i INTEGER;
  j INTEGER;
  vals float8[];
  toks UUID[];
  total float8;
  total_probability float8;
  tup INTEGER[];
  d INTEGER;
  prod_v float8;
  distinct_tok UUID[];
  conj_token UUID;
  prob float8;
  sign_max float8;
BEGIN
  IF token IS NULL OR k IS NULL THEN
    RETURN NULL;
  END IF;
  IF k < 0 THEN
    RAISE EXCEPTION 'agg_raw_moment(): k must be non-negative (got %)', k;
  END IF;
 
  -- Aggregate-carrier CASE (a gate_case over aggregate branches): a first-match
  -- guarded selection.  The moment is CONDITIONAL on the CASE's value being
  -- defined (NULL only when it never is, mirroring the MIN/MAX convention):
  --   E[pick^k | defined ∧ prov]
  --     = Σ_i P(region_i ∧ def_i) · E[value_i^k | region_i ∧ def_i]
  --       / Σ_i P(region_i ∧ def_i),
  -- where region_i = (¬g_1 ∧ … ∧ ¬g_{i-1}) ∧ g_i ∧ prov is the world set that
  -- selects branch i (the default's region is "all guards false") and def_i is
  -- the branch's defined event (agg_defined_event: gate_one for sum / count /
  -- constants, "some row present" for min / max / avg, recursive for a nested
  -- CASE).  Both factors are exact: probability() over the region ∧ def event,
  -- and the conditional aggregate moment (a recursive agg_raw_moment on the
  -- branch aggregate, which conditions on its own definedness within the
  -- region, so the two factors weigh the same worlds).  The regions are
  -- mutually exclusive, so the terms sum with no inclusion-exclusion, and
  -- correlation between a guard and its branch (shared input tuples) is
  -- carried by the conditioning, exactly as HAVING carries it.  When every
  -- branch is defined everywhere, the defined mass equals P(prov) and the
  -- formula reduces to the plain region-weighted sum.
  IF get_gate_type(token) = 'case' THEN
    IF k = 0 THEN
      RETURN 1;
    END IF;
    DECLARE
      wires UUID[] := get_children(token);
      nw INTEGER := array_length(get_children(token), 1);
      m INTEGER := (array_length(get_children(token), 1) - 1) / 2;
      running_neg UUID := gate_one();
      region_full UUID;
      prov_p float8;
      p float8;
      total float8 := 0;
      def_mass float8 := 0;
      ci INTEGER;
      vuid UUID;
      bm float8;
    BEGIN
      prov_p := probability(prov);
      IF prov_p IS NULL OR prov_p <= 0 THEN
        RETURN NULL;   -- impossible conditioning event
      END IF;
      -- Branches 1..m are the guarded WHENs; branch m+1 is the ELSE default,
      -- whose region is "all guards false".
      FOR ci IN 1 .. m + 1 LOOP
        IF ci <= m THEN
          region_full := provenance_times(running_neg, wires[2 * ci - 1], prov);
          vuid := wires[2 * ci];
          running_neg :=
            provenance_times(running_neg, provenance_not(wires[2 * ci - 1]));
        ELSE
          region_full := provenance_times(running_neg, prov);
          vuid := wires[nw];
        END IF;
        p := probability(provenance_times(region_full,
                                          agg_defined_event(vuid)));
        IF p > 0 THEN
          -- E[value_i^k | region_i ∧ def_i]: a constant branch is a Dirac
          -- (c^k, exact); a single aggregate or nested CASE is exact via
          -- agg_raw_moment (whose MIN/MAX/CASE arms condition on their own
          -- definedness within the region); an arithmetic / composite branch
          -- takes the Monte-Carlo scalar path (which composes with the
          -- aggregate leaves).
          IF get_gate_type(vuid) = 'value' THEN
            bm := power(CAST(get_extra(vuid) AS float8), k);
          ELSIF get_gate_type(vuid) IN ('agg', 'case') THEN
            bm := agg_raw_moment(agg_token_make(vuid, 0), k, region_full,
                                 method, arguments);
          ELSE
            bm := rv_moment(vuid, k, false, region_full);
          END IF;
          total := total + p * bm;
          def_mass := def_mass + p;
        END IF;
      END LOOP;
      IF def_mass <= epsilon() THEN
        RETURN NULL;   -- the CASE's value is never defined under prov
      END IF;
      RETURN total / def_mass;
    END;
  END IF;
 
  IF get_gate_type(token) <> 'agg' THEN
    IF get_gate_type(token) IN ('arith', 'conditioned') THEN
      RAISE EXCEPTION 'expected / variance / moment over an arithmetic '
        'combination of aggregates (e.g. SUM(x) + SUM(y) or SUM(x) + 5), or a '
        'conditioning of one, is not yet supported: a moment can be taken only '
        'over a single aggregate (SUM / COUNT / MIN / MAX), optionally '
        'conditioned (SUM(x) | C)'
        USING HINT = 'Take the moment of each aggregate separately, or condition '
          'the bare aggregate.';
    ELSE
      RAISE EXCEPTION USING MESSAGE='Wrong gate type for agg_raw_moment computation';
    END IF;
  END IF;
  IF k = 0 THEN
    RETURN 1;
  END IF;
 
  SELECT pp.proname::varchar FROM pg_proc pp
    WHERE oid=(get_infos(token)).info1
    INTO aggregation_function;
 
  child_pairs := get_children(token);
  n := COALESCE(array_length(child_pairs, 1), 0);
 
  IF aggregation_function = 'sum' OR aggregation_function = 'count' THEN
    -- count(col) keeps the COUNT identity at the gate level but its value is a
    -- SUM of per-row 0/1 indicators, so its moments are computed exactly like
    -- SUM (and its empty group is the real value 0, like SUM).  count(*)
    -- arrives here as 'sum' (it normalises to F_SUM_INT4); count(col) as 'count'.
    -- Trivial empty aggregation: SUM = 0, so SUM^k = 0 for k >= 1.
    -- Note: AGG_TOKEN semantics treat the "no row included" world as
    -- SUM = 0, so this stays consistent with k = 1 (= expected()).
    IF n = 0 THEN
      RETURN 0;
    END IF;
 
    -- Collapsed fast path: a correlated COUNT / SUM whose per-row selection
    -- events share a single continuous latent has an O(G·n) 1-D quadrature,
    -- vastly cheaper than the O(n^k) tuple enumeration below (which is the
    -- O(n^2) pair-probability bottleneck for the variance).  Only fires
    -- unconditionally (prov = one) and for k in {1, 2}; agg_collapsed_moment
    -- returns NULL when the shared-latent pattern does not match, and we
    -- fall through to the exact enumeration.
    IF prov = gate_one() AND k <= 2 THEN
      total := agg_collapsed_moment((token)::UUID, k);
      IF total IS NOT NULL THEN
        RETURN total;
      END IF;
    END IF;
 
    -- Extract per-child token + value arrays.
    vals := ARRAY[]::float8[];
    toks := ARRAY[]::UUID[];
    FOR i IN 1..n LOOP
      pair_children := get_children(child_pairs[i]);
      toks := toks || pair_children[1];
      vals := vals || CAST(get_extra(pair_children[2]) AS float8);
    END LOOP;
 
    -- Enumerate all k-tuples (i_1, ..., i_k) in {1..n}^k.  tup is the
    -- current tuple; we step through them in lexicographic order.
    total := 0;
    tup := array_fill(1, ARRAY[k]);
    LOOP
      prod_v := 1;
      FOR j IN 1..k LOOP
        prod_v := prod_v * vals[tup[j]];
      END LOOP;
 
      SELECT array_agg(DISTINCT toks[idx]) INTO distinct_tok
        FROM unnest(tup) AS idx;
 
      IF prov <> gate_one() THEN
        distinct_tok := distinct_tok || prov;
      END IF;
      conj_token := provenance_times(VARIADIC distinct_tok);
      prob := probability_evaluate(conj_token, method, arguments);
 
      total := total + prod_v * prob;
 
      d := k;
      WHILE d >= 1 AND tup[d] = n LOOP
        tup[d] := 1;
        d := d - 1;
      END LOOP;
      EXIT WHEN d = 0;
      tup[d] := tup[d] + 1;
    END LOOP;
  ELSIF aggregation_function = 'min' OR aggregation_function = 'max' THEN
    -- Rank enumeration: per distinct value v, P(MIN = v) is the
    -- probability that some t_i with v_i=v is true and all t_j with
    -- smaller v are false.  For MAX we negate values so the same
    -- "smaller-than" rank logic computes MIN-of-negated, then flip.
    -- The outer multiplier picks up the right sign for the k-th moment
    -- of MAX: E[MAX^k] = (-1)^k * E[MIN(-v)^k], so sign_max = (-1)^k.
    sign_max := CASE
                  WHEN aggregation_function = 'max'
                  THEN power(-1::float8, k)
                  ELSE 1
                END;
 
    -- MIN/MAX over the empty input world are NULL (no elements), not ±Infinity:
    -- SQL returns one row with a NULL value.  The moment is therefore CONDITIONAL
    -- on the aggregate being defined (non-empty) -- the empty world is excluded
    -- and the result renormalised by P(prov AND non-empty).  (count, whose empty
    -- value 0 is a real value, keeps the empty world; sum keeps it too, as 0.)
    IF n = 0 THEN
      RETURN NULL;  -- structurally empty: MIN/MAX undefined
    END IF;
 
    -- Numerator E[MIN^k . 1{prov AND non-empty}] (the rank sum naturally omits
    -- the empty world, since every term requires a present token).
    WITH tok_value AS (
      SELECT (get_children(c))[1] AS tok,
             (CASE WHEN aggregation_function='max' THEN -1 ELSE 1 END)
               * CAST(get_extra((get_children(c))[2]) AS DOUBLE PRECISION) AS v
      FROM UNNEST(child_pairs) AS c
    ) SELECT sign_max * COALESCE(SUM(p * power(v, k)), 0) FROM (
        SELECT t1.v AS v,
          probability_evaluate(
            CASE WHEN prov = gate_one()
                 THEN provenance_monus(provenance_plus(ARRAY_AGG(t1.tok)),
                                       provenance_plus(ARRAY_AGG(t2.tok)))
                 ELSE provenance_times(prov,
                        provenance_monus(provenance_plus(ARRAY_AGG(t1.tok)),
                                         provenance_plus(ARRAY_AGG(t2.tok)))) END,
            method, arguments) AS p
        FROM tok_value t1 LEFT OUTER JOIN tok_value t2 ON t1.v > t2.v
        GROUP BY t1.v) tmp
      INTO total;
 
    -- Denominator P(prov AND non-empty) = P(prov (x) (+) tokens).
    SELECT probability_evaluate(
             CASE WHEN prov = gate_one()
                  THEN provenance_plus(ARRAY_AGG(tok))
                  ELSE provenance_times(prov, provenance_plus(ARRAY_AGG(tok))) END,
             method, arguments)
      FROM (SELECT (get_children(c))[1] AS tok FROM UNNEST(child_pairs) AS c) s
      INTO total_probability;
 
    IF total_probability <= epsilon() THEN
      RETURN NULL;  -- never defined under prov: MIN/MAX undefined
    END IF;
    RETURN total / total_probability;  -- already conditional; skip generic norm
  ELSIF aggregation_function = 'avg' THEN
    -- AVG = SUM/COUNT is a ratio of two correlated world-dependent
    -- quantities, so the k-tuple expansion above does not apply.  Like
    -- MIN/MAX, AVG over the empty world is NULL, so its moment conditions
    -- on the aggregate being defined (COUNT >= 1), NULL when it never is.
    -- Two routes:
    --  * EXACT (independent rows, unconditional): the joint (sum, count)
    --    PMF folded in C by agg_avg_moment_exact --
    --    E[AVG^k | COUNT>=1] = Σ_{(s,c), c>=1} (s/c)^k pmf(s,c) / P(c>=1).
    --  * Monte-Carlo scalar fallback otherwise (an outer conditioning
    --    event, shared leaves, compound contributors): rv_moment samples
    --    the agg gate per world; its NaN-skip on empty draws implements
    --    the same conditional-on-defined convention, at the
    --    provsql.rv_mc_samples budget (0 raises, per convention).
    IF n = 0 THEN
      RETURN NULL;  -- structurally empty: AVG undefined
    END IF;
    IF prov = gate_one() THEN
      total := agg_avg_moment_exact((token)::UUID, k);
      IF total IS NOT NULL THEN
        RETURN total;
      END IF;
    END IF;
    RETURN rv_moment((token)::UUID, k, false, prov);
  ELSE
    RAISE EXCEPTION USING MESSAGE=
      'Cannot compute moment for aggregation function ' || aggregation_function;
  END IF;
 
  -- Conditional normalisation: E[X^k · 1_A] / P(A) = E[X^k | A].
  IF prov <> gate_one()
     AND total <> 0
     AND total <> 'Infinity'::float8
     AND total <> '-Infinity'::float8 THEN
    total := total / probability_evaluate(prov, method, arguments);
  END IF;
 
  RETURN total;
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Compute the variance Var[X | prov] of a probabilistic scalar
 *
 * Polymorphic dispatcher that mirrors @c expected: @c random_variable
 * inputs go through the analytical / MC evaluator
 * (@c rv_moment(UUID, 2, true)); @c AGG_TOKEN inputs go through the
 * @c agg_raw_moment helper, computing
 * @f$\mathrm{Var}[X|A] = E[X^2|A] - E[X|A]^2@f$.  Conditioning on
 * @c prov is supported for @c AGG_TOKEN (matching @c expected) but
 * not yet for @c random_variable.
 */
CREATE OR REPLACE FUNCTION variance(
  input ANYELEMENT,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
DECLARE
  m1 float8;
  m2 float8;
BEGIN
  IF pg_typeof(input) = 'random_variable'::REGTYPE THEN
    IF input IS NULL THEN
      RETURN NULL;
    END IF;
    -- Conditioning on prov is handled inside rv_moment: when prov
    -- resolves to gate_one() (the default, or load-time
    -- simplification of any always-true sub-circuit) the
    -- unconditional analytical path runs unchanged; otherwise the
    -- joint-circuit loader unifies shared gate_rv leaves between
    -- input and prov, and the conditional path runs either
    -- truncated-distribution closed form or MC rejection.
    RETURN provsql.rv_moment(
      rv_conditioned_target((input::random_variable)::UUID), 2, true,
      rv_conditioned_prov((input::random_variable)::UUID, prov));
  END IF;
 
  IF pg_typeof(input) = 'AGG_TOKEN'::REGTYPE THEN
    IF input IS NULL THEN
      RETURN NULL;
    END IF;
    -- Collapsed fast path: E[C] and E[C^2] from a single circuit load and plan
    -- build, instead of two agg_raw_moment() calls that each reload.  Mirrors
    -- the guard in agg_raw_moment (unconditional only, prov = one); on any
    -- mismatch agg_collapsed_moments returns NULL and we fall through to the
    -- generic per-order path (which handles conditioning, SUM enumeration, ...).
    IF rv_conditioned_prov(input::UUID, prov) = gate_one() THEN
      DECLARE ms float8[];
      BEGIN
        ms := agg_collapsed_moments(
                (agg_conditioned_target(input::AGG_TOKEN))::UUID);
        IF ms IS NOT NULL THEN
          RETURN ms[2] - ms[1] * ms[1];
        END IF;
      END;
    END IF;
    m1 := agg_raw_moment(agg_conditioned_target(input::AGG_TOKEN), 1,
                         rv_conditioned_prov(input::UUID, prov), method, arguments);
    m2 := agg_raw_moment(agg_conditioned_target(input::AGG_TOKEN), 2,
                         rv_conditioned_prov(input::UUID, prov), method, arguments);
    IF m1 IS NULL OR m2 IS NULL THEN
      RETURN NULL;
    END IF;
    RETURN m2 - m1 * m1;
  END IF;
 
  -- Bernoulli event token (see moment()): Var[X] = p(1 - p).
  IF pg_typeof(input) = 'UUID'::REGTYPE THEN
    IF input IS NULL THEN
      RETURN NULL;
    END IF;
    m1 := provsql.probability_evaluate(provsql.cond(input::UUID, prov),
                                       method, arguments);
    RETURN m1 * (1 - m1);
  END IF;
 
  RAISE EXCEPTION 'variance() is not yet supported for input type %', pg_typeof(input);
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Compute the raw moment E[X^k | prov] of a probabilistic scalar
 *
 * @c k must be a non-negative INTEGER.  @c k = 0 returns 1; @c k = 1
 * is equivalent to @c expected(input).  Polymorphic dispatcher: routes
 * @c random_variable through @c rv_moment (analytical / MC) and
 * @c AGG_TOKEN through @c agg_raw_moment (SUM via tuple enumeration,
 * MIN / MAX via rank enumeration, AVG via the joint (sum, count)
 * distribution over independent / laminar rows with a Monte-Carlo
 * fallback).
 */
CREATE OR REPLACE FUNCTION moment(
  input ANYELEMENT,
  k INTEGER,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
BEGIN
  IF pg_typeof(input) = 'random_variable'::REGTYPE THEN
    IF input IS NULL OR k IS NULL THEN
      RETURN NULL;
    END IF;
    -- See variance() above: rv_moment handles the conditional/unconditional
    -- dispatch internally based on the resolved prov gate type.
    RETURN provsql.rv_moment(
      rv_conditioned_target((input::random_variable)::UUID), k, false,
      rv_conditioned_prov((input::random_variable)::UUID, prov));
  END IF;
 
  IF pg_typeof(input) = 'AGG_TOKEN'::REGTYPE THEN
    RETURN agg_raw_moment(agg_conditioned_target(input::AGG_TOKEN), k,
                          rv_conditioned_prov(input::UUID, prov), method, arguments);
  END IF;
 
  -- A bare provenance event token (a gate_cmp lifted from an RV comparison,
  -- e.g. expected(x <= c)) is a Bernoulli indicator: X in {0,1}, so every raw
  -- moment E[X^k] with k >= 1 equals P(event), and E[X^0] = 1.  cond() applies
  -- the optional conditioning prov (a no-op for the default gate_one()).
  IF pg_typeof(input) = 'UUID'::REGTYPE THEN
    IF input IS NULL OR k IS NULL THEN
      RETURN NULL;
    END IF;
    IF k = 0 THEN
      RETURN 1;
    END IF;
    RETURN provsql.probability_evaluate(provsql.cond(input::UUID, prov),
                                        method, arguments);
  END IF;
 
  RAISE EXCEPTION 'moment() is not yet supported for input type %', pg_typeof(input);
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Internal: rv-side quantile computation.
 *
 * C entry point behind the polymorphic @c quantile dispatcher.
 * Closed-form inverse CDF where the family has one (Normal via
 * Beasley-Springer-Moro polished by Newton steps, Uniform and
 * Exponential by algebraic inversion), generic monotone-CDF bisection
 * otherwise (Erlang, Gamma), exact generalised inverse for categorical
 * mixtures, and the empirical Monte Carlo quantile for compound scalar
 * circuits.  A non-trivial @p prov conditions (truncates) the
 * distribution first, in closed form when the event reduces to an
 * interval on a bare @c gate_rv.
 */
CREATE OR REPLACE FUNCTION rv_quantile(
  token UUID, p double precision,
  prov UUID DEFAULT gate_one())
  RETURNS double precision
  AS 'provsql','rv_quantile' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Compute the p-quantile (inverse CDF) of a probabilistic scalar
 *
 * @f$F^{-1}(p) = \min\{x : P(X \le x) \ge p\}@f$ for @f$p \in [0,1]@f$:
 * medians (@c p = 0.5), percentiles, Value-at-Risk, and credible
 * intervals.  @c p = 0 / @c p = 1 return the (possibly infinite)
 * support edges.  Polymorphic dispatcher mirroring @c expected /
 * @c moment: @c random_variable routes through @c rv_quantile
 * (analytical inverse CDF / MC), plain numerics are their own quantile
 * (a Dirac's inverse CDF is constant), and the optional @p prov
 * argument conditions on a provenance event, e.g.
 * <tt>quantile(x | (x > 0), 0.5)</tt> for the median of a truncated
 * distribution.
 */
CREATE OR REPLACE FUNCTION quantile(
  input ANYELEMENT,
  p double precision,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
BEGIN
  IF p IS NULL THEN
    RETURN NULL;
  END IF;
  IF p <> p OR p < 0 OR p > 1 THEN
    RAISE EXCEPTION 'quantile: p must be in [0, 1] (got %)', p;
  END IF;
 
  IF pg_typeof(input) = 'random_variable'::REGTYPE THEN
    IF input IS NULL THEN
      RETURN NULL;
    END IF;
    -- See variance(): rv_quantile handles the conditional/unconditional
    -- dispatch internally based on the resolved prov gate type.
    RETURN provsql.rv_quantile(
      rv_conditioned_target((input::random_variable)::UUID), p,
      rv_conditioned_prov((input::random_variable)::UUID, prov));
  END IF;
 
  IF pg_typeof(input) IN ('smallint'::REGTYPE, 'INTEGER'::REGTYPE,
                          'bigint'::REGTYPE, 'NUMERIC'::REGTYPE,
                          'real'::REGTYPE, 'double precision'::REGTYPE) THEN
    -- A deterministic scalar is a Dirac: every quantile is the value.
    RETURN input::double precision;
  END IF;
 
  RAISE EXCEPTION 'quantile() is not yet supported for input type %', pg_typeof(input);
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Internal: rv-side support computation
 *
 * Lifts @c provsql.compute_support out of @c RangeCheck.cpp -- the
 * same interval-arithmetic propagation @c runRangeCheck uses to
 * decide @c gate_cmps.  Returns @c [-Infinity, +Infinity] when the
 * tightest bound is the conservative all-real interval (e.g. for a
 * normal RV, or any sub-circuit that mixes a normal in).
 */
CREATE OR REPLACE FUNCTION rv_support(
  token UUID, prov UUID DEFAULT gate_one(),
  OUT lo float8, OUT hi float8)
  AS 'provsql','rv_support' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Compute the support interval @c [lo, hi] of a probabilistic
 *        (or deterministic) scalar
 *
 * Polymorphic dispatcher mirroring @c expected / @c variance /
 * @c moment / @c central_moment, with two extra "free" branches:
 *
 * - <b>Plain NUMERIC</b> (@c smallint / @c INTEGER / @c bigint /
 *   @c NUMERIC / @c real / @c double @c precision): degenerate
 *   point support @f$[c, c]@f$.  Lets callers ask for the support
 *   of a literal without round-tripping through @c as_random.
 * - <b>@c random_variable / bare @c UUID</b> (any provenance gate
 *   token; the @c random_variable branch reinterprets the value via
 *   the binary-coercible @c random_variable @c -> @c UUID cast):
 *   routes to @c rv_support, which propagates distribution
 *   supports (uniform exact, exponential @c [0,+∞), normal
 *   @c (-∞,+∞)) through @c gate_arith via interval arithmetic.
 *   @c gate_value gives the same @f$[c, c]@f$ point support as the
 *   NUMERIC branch; any non-scalar gate (Boolean gates, aggregates,
 *   ...) safely falls back to the conservative all-real interval
 *   without raising.  Conditioning on @c prov is not yet supported.
 *
 * - @c AGG_TOKEN: closed-form per aggregation function:
 *   - @c SUM : @f$[\sum_i \min(0,v_i), \sum_i \max(0,v_i)]@f$
 *     (every row is independently in or out of the included set; the
 *     extreme SUMs are reached by including only positive or only
 *     negative-valued rows).
 *   - @c MIN : @f$[\min_i v_i, \max_i v_i]@f$ in the non-empty
 *     subsets, plus @c +Infinity if the empty subset has positive
 *     probability under @c prov.
 *   - @c MAX : symmetric -- @c -Infinity if empty has positive
 *     probability under @c prov, otherwise @c min_i v_i; @c hi is
 *     always @c max_i v_i.
 *
 * Other aggregation functions raise.
 *
 * Returns the composite RECORD @c (lo, hi) via the function's
 * @c OUT parameters, with @c -Infinity / @c +Infinity marking
 * unbounded ends.
 */
CREATE OR REPLACE FUNCTION support(
  input ANYELEMENT,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL,
  OUT lo float8,
  OUT hi float8)
  AS $$
DECLARE
  aggregation_function VARCHAR;
  child_pairs UUID[];
  values_arr float8[];
  total_probability float8;
BEGIN
  IF input IS NULL THEN
    lo := NULL; hi := NULL; RETURN;
  END IF;
 
  -- Plain NUMERIC: degenerate point support.  Lets `support(2.5)` /
  -- `support(42)` / etc.  return (2.5, 2.5) without making the user
  -- wrap in `as_random`.
  IF pg_typeof(input) IN (
       'smallint'::REGTYPE, 'INTEGER'::REGTYPE, 'bigint'::REGTYPE,
       'NUMERIC'::REGTYPE, 'real'::REGTYPE, 'double precision'::REGTYPE) THEN
    lo := input::double precision;
    hi := input::double precision;
    RETURN;
  END IF;
 
  -- random_variable is binary-coercible to UUID (explicit cast
  -- below), so a single rv_support call covers both shapes.
  -- rv_support handles
  -- gate_value (point), gate_rv (distribution), gate_arith
  -- (propagated), and falls back to the conservative all-real
  -- interval for any other gate kind.  Conditioning on prov is not
  -- supported (would require restricting the underlying joint
  -- distribution by the indicator of prov, which has no closed form
  -- for the basic distributions we ship).
  IF pg_typeof(input) IN ('random_variable'::REGTYPE, 'UUID'::REGTYPE) THEN
    -- Conditional support: rv_support folds the AND-conjunct interval
    -- constraints from prov into the unconditional support.  When
    -- prov is gate_one() the unconditional support is returned
    -- unchanged.
    SELECT r.lo, r.hi INTO lo, hi
      FROM provsql.rv_support(
             rv_conditioned_target(input::UUID),
             rv_conditioned_prov(input::UUID, prov)) r;
    RETURN;
  END IF;
 
  IF pg_typeof(input) = 'AGG_TOKEN'::REGTYPE THEN
    -- A conditioned aggregate SUM(x)|C: the value-range support is that of
    -- the target aggregate (conditioning can only tighten it; the
    -- conservative range stays valid), so unpack to the target gate.
    DECLARE
      atok AGG_TOKEN := agg_conditioned_target(input::AGG_TOKEN);
    BEGIN
    IF get_gate_type(atok) <> 'agg' THEN
      RAISE EXCEPTION USING MESSAGE='Wrong gate type for support computation';
    END IF;
    SELECT pp.proname::varchar FROM pg_proc pp
      WHERE oid=(get_infos(atok)).info1
      INTO aggregation_function;
    child_pairs := get_children(atok);
 
    IF aggregation_function = 'sum' OR aggregation_function = 'count' THEN
      -- count(col) is a SUM of per-row 0/1 indicators (empty group = 0), so its
      -- support is computed like SUM; count(*) arrives as 'sum'.
      -- Empty AGG_TOKEN: SUM is identically 0.
      IF COALESCE(array_length(child_pairs, 1), 0) = 0 THEN
        lo := 0; hi := 0; RETURN;
      END IF;
      SELECT sum(LEAST(v, 0::float8)), sum(GREATEST(v, 0::float8))
        INTO lo, hi
        FROM (SELECT CAST(get_extra((get_children(c))[2]) AS float8) AS v
              FROM unnest(child_pairs) AS c) sub;
    ELSIF aggregation_function = 'min' OR aggregation_function = 'max' THEN
      -- MIN/MAX over the empty input world are NULL, not ±Infinity (matching the
      -- moment surface): the empty world carries no value, so the support is just
      -- the range of the per-row values [min(v), max(v)].  A structurally empty
      -- aggregate has no defined value at all -> NULL support.
      IF COALESCE(array_length(child_pairs, 1), 0) = 0 THEN
        lo := NULL; hi := NULL; RETURN;
      END IF;
 
      SELECT min(v), max(v)
        INTO lo, hi
        FROM (SELECT CAST(get_extra((get_children(c))[2]) AS float8) AS v
              FROM UNNEST(child_pairs) AS c) sub;
    ELSE
      RAISE EXCEPTION USING MESSAGE=
        'Cannot compute support for aggregation function ' || aggregation_function;
    END IF;
    RETURN;
    END;
  END IF;
 
  RAISE EXCEPTION 'support() is not yet supported for input type %', pg_typeof(input);
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Compute the central moment E[(X - E[X|prov])^k | prov]
 *
 * @c k = 0 returns 1; @c k = 1 returns 0; @c k = 2 is equivalent to
 * @c variance(input, prov, ...).  Polymorphic dispatcher: routes
 * @c random_variable through @c rv_moment, and @c AGG_TOKEN through
 * the binomial expansion
 * @f$E[(X-\mu)^k|A] = \sum_{i=0}^{k} \binom{k}{i} (-\mu)^{k-i} E[X^i|A]@f$
 * with @f$\mu = E[X|A]@f$, where each @f$E[X^i|A]@f$ comes from
 * @c agg_raw_moment.
 */
CREATE OR REPLACE FUNCTION central_moment(
  input ANYELEMENT,
  k INTEGER,
  prov UUID = gate_one(),
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS $$
DECLARE
  mu float8;
  total float8;
  i INTEGER;
  raw_i float8;
  binom float8;
  -- iterative binomial coefficient C(k, i)
  k_double float8;
BEGIN
  IF pg_typeof(input) = 'random_variable'::REGTYPE THEN
    IF input IS NULL OR k IS NULL THEN
      RETURN NULL;
    END IF;
    -- See variance() above: rv_moment handles the conditional/unconditional
    -- dispatch internally based on the resolved prov gate type.
    RETURN provsql.rv_moment(
      rv_conditioned_target((input::random_variable)::UUID), k, true,
      rv_conditioned_prov((input::random_variable)::UUID, prov));
  END IF;
 
  IF pg_typeof(input) = 'AGG_TOKEN'::REGTYPE THEN
    IF input IS NULL OR k IS NULL THEN
      RETURN NULL;
    END IF;
    IF k < 0 THEN
      RAISE EXCEPTION 'central_moment(): k must be non-negative (got %)', k;
    END IF;
    IF k = 0 THEN RETURN 1; END IF;
    IF k = 1 THEN RETURN 0; END IF;
 
    mu := agg_raw_moment(agg_conditioned_target(input::AGG_TOKEN), 1,
                         rv_conditioned_prov(input::UUID, prov), method, arguments);
    IF mu IS NULL THEN RETURN NULL; END IF;
    -- mu may be ±Infinity for empty MIN / MAX with positive empty
    -- probability; central_moment is undefined in that case.
    IF mu = 'Infinity'::float8 OR mu = '-Infinity'::float8 THEN
      RETURN mu;
    END IF;
 
    total := 0;
    binom := 1;  -- C(k, 0)
    k_double := k;
    FOR i IN 0..k LOOP
      raw_i := agg_raw_moment(agg_conditioned_target(input::AGG_TOKEN), i,
                              rv_conditioned_prov(input::UUID, prov), method, arguments);
      IF raw_i IS NULL THEN RETURN NULL; END IF;
      total := total + binom * power(-mu, k - i) * raw_i;
      -- C(k, i+1) = C(k, i) * (k - i) / (i + 1)
      IF i < k THEN
        binom := binom * (k_double - i) / (i + 1);
      END IF;
    END LOOP;
    RETURN total;
  END IF;
 
  -- Bernoulli event token (see moment()): with p = P(event),
  -- E[(X-p)^k] = (1-p)(-p)^k + p(1-p)^k; k = 0 -> 1, k = 1 -> 0.
  IF pg_typeof(input) = 'UUID'::REGTYPE THEN
    IF input IS NULL OR k IS NULL THEN
      RETURN NULL;
    END IF;
    IF k < 0 THEN
      RAISE EXCEPTION 'central_moment(): k must be non-negative (got %)', k;
    END IF;
    IF k = 0 THEN RETURN 1; END IF;
    IF k = 1 THEN RETURN 0; END IF;
    mu := provsql.probability_evaluate(provsql.cond(input::UUID, prov),
                                       method, arguments);
    RETURN (1 - mu) * power(-mu, k) + mu * power(1 - mu, k);
  END IF;
 
  RAISE EXCEPTION 'central_moment() is not yet supported for input type %', pg_typeof(input);
END
$$ LANGUAGE plpgsql PARALLEL SAFE SET search_path=provsql SECURITY DEFINER;
 
/** @brief C entry point behind @ref covariance (UUID-level binding). */
CREATE OR REPLACE FUNCTION rv_covariance(x UUID, y UUID, prov UUID)
  RETURNS double precision
  AS 'provsql','rv_covariance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Covariance Cov(X, Y) = E[XY] − E[X]·E[Y] of two random variables.
 *
 * The bivariate readout complementing the univariate moment surface
 * (@ref expected / @ref variance / @ref moment / @ref central_moment).
 * Exact tiers: an exact @c 0 when the two arguments' stochastic-leaf
 * footprints are structurally independent (given @p prov), a variance
 * readout when the two arguments coincide, and the closed-form
 * @c E[XY] − E[X]·E[Y] whenever every factor decomposes analytically.
 * When some factor has no closed form, a SINGLE coupled Monte-Carlo pass
 * over the joint circuit draws @c (x, y) pairs (shared leaves produce one
 * draw both observe) and returns the sample covariance -- the estimator's
 * noise then scales with the covariance signal itself, not with the
 * product of the means as the naive three-run E[XY] − E[X]·E[Y]
 * subtraction would.
 *
 * @param x    the first random variable.
 * @param y    the second random variable.
 * @param prov optional conditioning event (a provenance @c UUID); the
 *   default @c gate_one() is the unconditional covariance.  Conditioning
 *   is applied jointly: the Monte-Carlo pass rejection-samples the pair on
 *   @p prov, giving @c Cov(X, Y | prov).
 */
CREATE OR REPLACE FUNCTION covariance(
  x random_variable, y random_variable, prov UUID DEFAULT gate_one())
  RETURNS double precision AS $$
  SELECT provsql.rv_covariance((x)::UUID, (y)::UUID, prov);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Standard deviation σ(X) = √Var(X) of a random variable.
 *
 * A thin NUMERIC readout over @ref variance.  The square root is taken on
 * the scalar @c double result, so no RV-level @c sqrt is involved and this
 * carries no dependency on RV function application (@c pow / @c sqrt).
 * @c NULL propagates from a @c NULL input; the order-2 central moment is
 * non-negative by construction, so the root is always real.
 *
 * @param x    the random variable.
 * @param prov optional conditioning event; default @c gate_one()
 *   (unconditional).
 */
CREATE OR REPLACE FUNCTION stddev(
  x random_variable, prov UUID DEFAULT gate_one())
  RETURNS double precision AS $$
  SELECT sqrt(provsql.variance(x, prov));
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/** @brief C entry point behind @ref correlation (UUID-level binding). */
CREATE OR REPLACE FUNCTION rv_correlation(x UUID, y UUID, prov UUID)
  RETURNS double precision
  AS 'provsql','rv_correlation' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Pearson correlation ρ(X, Y) = Cov(X, Y) / (σ(X)·σ(Y)).
 *
 * Same exact tiers as @ref covariance; on the Monte-Carlo path the
 * covariance and BOTH standard deviations are read off the same coupled
 * pass, instead of stacking five independent estimates (three for the
 * covariance, one per standard deviation).  Returns @c NULL when either
 * standard deviation is @c 0 (a degenerate / constant variable, for which
 * correlation is undefined) rather than raising a division-by-zero.
 *
 * @param x    the first random variable.
 * @param y    the second random variable.
 * @param prov optional conditioning event; default @c gate_one()
 *   (unconditional).
 */
CREATE OR REPLACE FUNCTION correlation(
  x random_variable, y random_variable, prov UUID DEFAULT gate_one())
  RETURNS double precision AS $$
  SELECT provsql.rv_correlation((x)::UUID, (y)::UUID, prov);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/** @brief C entry point behind @ref entropy (UUID-level binding). */
CREATE OR REPLACE FUNCTION rv_entropy(token UUID, prov UUID)
  RETURNS double precision
  AS 'provsql','rv_entropy' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Entropy H(X) of a random variable, in nats.
 *
 * Shannon entropy for a discrete distribution (a categorical / discrete
 * count / constant -- a point mass has entropy @c 0), differential
 * entropy for a continuous one (quadrature of @c -f ln f over the
 * family's integration range; also exact through independent-arm
 * Bernoulli mixture trees such as @ref gmm's).  Shapes with no
 * closed density (arithmetic composites) and the conditional form fall
 * back to a Monte Carlo histogram plug-in estimate at the
 * @c provsql.rv_mc_samples budget.
 *
 * @param x    the random variable.
 * @param prov optional conditioning event; default @c gate_one()
 *   (unconditional).
 */
CREATE OR REPLACE FUNCTION entropy(
  x random_variable, prov UUID DEFAULT gate_one())
  RETURNS double precision AS $$
  SELECT provsql.rv_entropy((x)::UUID, prov);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/** @brief C entry point behind @ref kl (UUID-level binding). */
CREATE OR REPLACE FUNCTION rv_kl(p UUID, q UUID)
  RETURNS double precision
  AS 'provsql','rv_kl' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Kullback-Leibler divergence KL(P || Q), in nats.
 *
 * Exact: the defining sum for two discrete distributions (matching
 * outcomes by value) and the defining integral (quadrature over P's
 * integration window) for two continuous ones, including
 * independent-arm mixture trees.  Returns @c Infinity when P is not
 * absolutely continuous with respect to Q -- an outcome of P that Q
 * gives zero mass, mismatched kinds (discrete vs continuous), or a
 * region of P's support where Q's density (under)flows to zero.  Both
 * arguments must resolve to closed-form densities; arithmetic
 * composites and conditioned variables raise.
 */
CREATE OR REPLACE FUNCTION kl(p random_variable, q random_variable)
  RETURNS double precision AS $$
  SELECT provsql.rv_kl((p)::UUID, (q)::UUID);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/** @brief C entry point behind @ref mutual_information (UUID-level
 *  binding). */
CREATE OR REPLACE FUNCTION rv_mutual_information(x UUID, y UUID)
  RETURNS double precision
  AS 'provsql','rv_mutual_information' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
/**
 * @brief Mutual information I(X; Y), in nats.
 *
 * Exactly @c 0 for structurally independent variables (disjoint
 * stochastic-leaf footprints, the same test the moment evaluators use);
 * @c H(X) for a discrete variable paired with itself and @c Infinity
 * for a continuous one (I(X;X) diverges).  A genuinely correlated pair
 * (shared leaves) is estimated by a 2-D histogram plug-in over coupled
 * joint Monte Carlo draws -- both roots evaluated against the same
 * per-iteration cache, so shared leaves keep their joint law -- at the
 * @c provsql.rv_mc_samples budget.
 */
CREATE OR REPLACE FUNCTION mutual_information(
  x random_variable, y random_variable)
  RETURNS double precision AS $$
  SELECT provsql.rv_mutual_information((x)::UUID, (y)::UUID);
$$ LANGUAGE sql PARALLEL SAFE STABLE SET search_path=provsql SECURITY DEFINER;
 
/**
 * @brief Compute the Shapley value of an input variable
 *
 * Measures the contribution of a specific input variable to the
 * truth of a provenance expression, using game-theoretic Shapley values.
 *
 * @param token provenance token to evaluate
 * @param variable UUID of the input variable
 * @param method knowledge compilation method
 * @param arguments additional arguments for the method
 * @param banzhaf if true, compute the Banzhaf value instead
 */
CREATE OR REPLACE FUNCTION shapley(
  token UUID,
  variable UUID,
  method TEXT = NULL,
  arguments TEXT = NULL,
  banzhaf BOOLEAN = 'f')
  RETURNS DOUBLE PRECISION AS
  'provsql','shapley' LANGUAGE C STABLE;
 
/** @brief Compute Shapley values for all input variables at once */
CREATE OR REPLACE FUNCTION shapley_all_vars(
  IN token UUID,
  IN method TEXT = NULL,
  IN arguments TEXT = NULL,
  IN banzhaf BOOLEAN = 'f',
  OUT variable UUID,
  OUT value DOUBLE PRECISION)
  RETURNS SETOF RECORD AS
  'provsql', 'shapley_all_vars'
  LANGUAGE C STABLE;
 
/** @brief Compute the Banzhaf power index of an input variable */
CREATE OR REPLACE FUNCTION banzhaf(
  token UUID,
  variable UUID,
  method TEXT = NULL,
  arguments TEXT = NULL)
  RETURNS DOUBLE PRECISION AS
  $$ SELECT provsql.shapley(token, variable, method, arguments, 't') $$
  LANGUAGE SQL;
 
/** @brief Compute Banzhaf power indices for all input variables at once */
CREATE OR REPLACE FUNCTION banzhaf_all_vars(
  IN token UUID,
  IN method TEXT = NULL,
  IN arguments TEXT = NULL,
  OUT variable UUID,
  OUT value DOUBLE PRECISION)
  RETURNS SETOF RECORD AS
  $$ SELECT * FROM provsql.shapley_all_vars(token, method, arguments, 't') $$
  LANGUAGE SQL;
 
/**
 * @brief Exact reachability probability over bounded-treewidth data
 * (columnar form)
 *
 * Computes the probability that @p target is reachable from @p source in
 * the probabilistic graph given by the parallel edge arrays
 * (two-terminal network reliability).  Unlike
 * @c probability_evaluate(), which compiles the provenance circuit
 * built along the relational query plan, this compiles the query
 * along a tree decomposition of the *data* graph (in the spirit of the
 * provenance refinement of Courcelle's theorem), producing a d-DNNF
 * whose size is linear in the number of edges for data of bounded
 * treewidth.  Exact, and linear-time, on cyclic data as well -- where
 * the recursive-query fixpoint cannot terminate structurally.
 *
 * Edges are independent events.  Two array positions may share a token
 * only if they are mutual reverses (the natural encoding of an
 * undirected edge in a directed edge relation); they are then treated
 * as a single bidirectional edge.  This is an internal/testing surface:
 * the user-facing route is a plain @c WITH @c RECURSIVE reachability
 * query under the 'absorptive' (or 'BOOLEAN') provenance class, which
 * the query rewriter compiles through @c eval_reachability() /
 * @c reachability_materialize().
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param source the vertex reachability starts from
 * @param target the vertex whose reachability is evaluated
 * @param directed if false, each edge can be traversed both ways
 */
CREATE OR REPLACE FUNCTION reachability_evaluate(
  sources INT[],
  destinations INT[],
  tokens UUID[],
  probabilities DOUBLE PRECISION[],
  source INT,
  target INT,
  directed BOOLEAN)
  RETURNS DOUBLE PRECISION AS
  'provsql','reachability_evaluate' LANGUAGE C IMMUTABLE PARALLEL SAFE;
 
/**
 * @brief Reachability probability plus compilation statistics
 * (columnar form)
 *
 * Same compilation as @c reachability_evaluate(), returning the
 * probability together with the structural statistics that
 * substantiate the bounded-treewidth guarantee: the treewidth of the
 * min-fill decomposition of the data graph, its number of bags, the
 * maximum number of dynamic-programming states at any decomposition
 * node, and the size of the emitted d-DNNF (linear in the number of
 * edges for fixed data treewidth).
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param source the vertex reachability starts from
 * @param target the vertex whose reachability is evaluated
 * @param directed if false, each edge can be traversed both ways
 * @param[out] probability the reachability probability
 * @param[out] data_treewidth treewidth of the min-fill decomposition of the
 *        data graph
 * @param[out] nb_bags number of bags in the decomposition
 * @param[out] max_states maximum number of dynamic-programming states at any
 *        decomposition node
 * @param[out] nb_gates number of gates in the emitted d-DNNF
 * @param[out] nb_variables number of variables in the emitted d-DNNF
 */
CREATE OR REPLACE FUNCTION reachability_compile_stats(
  IN sources INT[],
  IN destinations INT[],
  IN tokens UUID[],
  IN probabilities DOUBLE PRECISION[],
  IN source INT,
  IN target INT,
  IN directed BOOLEAN,
  OUT probability DOUBLE PRECISION,
  OUT data_treewidth INT,
  OUT nb_bags BIGINT,
  OUT max_states BIGINT,
  OUT nb_gates BIGINT,
  OUT nb_variables BIGINT)
  AS 'provsql','reachability_compile_stats'
  LANGUAGE C IMMUTABLE PARALLEL SAFE;
 
 
 
/**
 * @brief Boolean UCQ probability plus compilation statistics
 * (columnar form, internal)
 *
 * Same compilation as @c ucq_joint_compile_stats(query jsonb, ...),
 * returning the probability together with the three width columns that
 * substantiate thesis Prop. 4.2.11 empirically -- the adversarial family
 * has small data and circuit widths but large joint width -- and the
 * structural statistics.
 *
 * @param disjunct_nvars number of query variables of each disjunct
 * @param atom_disjunct disjunct index of each atom (parallel to @p atom_rel)
 * @param atom_rel relation id of each atom
 * @param atom_vars query-variable indices of all atom columns, concatenated
 * @param atom_arity number of columns of each atom (slices @p atom_vars)
 * @param fact_rel relation id of each fact
 * @param fact_elems element ids of all fact columns, concatenated
 * @param fact_arity number of columns of each fact (slices @p fact_elems)
 * @param fact_tokens provenance token of each fact
 * @param fact_probs probability of each fact
 * @param[out] probability the exact UCQ probability
 * @param[out] joint_treewidth width of the min-fill decomposition found
 * @param[out] data_treewidth_lb degeneracy lower bound of the data-only graph
 * @param[out] circuit_treewidth_lb degeneracy lower bound of the slice-only graph
 * @param[out] n_bags number of bags in the decomposition
 * @param[out] max_states peak number of DP states at any node
 * @param[out] dd_size number of gates in the emitted d-D
 * @param[out] n_enumerating maximum number of essential (enumerating) query
 *        variables over the disjuncts -- the @c e of the @f$2^{O(k^e)}@f$
 *        bound, with variables functionally determined by others (via FDs
 *        mined from the data) removed
 */
CREATE OR REPLACE FUNCTION ucq_joint_compile_stats(
  IN disjunct_nvars INT[],
  IN atom_disjunct INT[],
  IN atom_rel INT[],
  IN atom_vars INT[],
  IN atom_arity INT[],
  IN fact_rel INT[],
  IN fact_elems INT[],
  IN fact_arity INT[],
  IN fact_tokens UUID[],
  IN fact_probs DOUBLE PRECISION[],
  OUT probability DOUBLE PRECISION,
  OUT joint_treewidth INT,
  OUT data_treewidth_lb INT,
  OUT circuit_treewidth_lb INT,
  OUT n_bags BIGINT,
  OUT max_states BIGINT,
  OUT dd_size BIGINT,
  OUT n_enumerating INT)
  AS 'provsql','ucq_joint_compile_stats'
  LANGUAGE C IMMUTABLE PARALLEL SAFE;
 
 
/**
 * @brief Boolean UCQ probability plus statistics from a JSON specification
 *
 * JSON-spec wrapper over the columnar @c ucq_joint_compile_stats()
 * (see @c ucq_joint_evaluate(query jsonb, ...) for the JSON format).
 */
CREATE OR REPLACE FUNCTION ucq_joint_compile_stats(
  IN query JSONB,
  IN fact_rel INT[],
  IN fact_elems INT[],
  IN fact_arity INT[],
  IN fact_tokens UUID[],
  IN fact_probs DOUBLE PRECISION[],
  OUT probability DOUBLE PRECISION,
  OUT joint_treewidth INT,
  OUT data_treewidth_lb INT,
  OUT circuit_treewidth_lb INT,
  OUT n_bags BIGINT,
  OUT max_states BIGINT,
  OUT dd_size BIGINT,
  OUT n_enumerating INT)
  AS $$
DECLARE
  dnv INT[] := '{}'; adisj INT[] := '{}'; arel INT[] := '{}';
  avars INT[] := '{}'; aarity INT[] := '{}';
  d JSONB; a JSONB; v TEXT; didx INT := 0;
BEGIN
  FOR d IN SELECT * FROM jsonb_array_elements(query->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx;
      arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
  SELECT s.probability, s.joint_treewidth, s.data_treewidth_lb,
         s.circuit_treewidth_lb, s.n_bags, s.max_states, s.dd_size,
         s.n_enumerating
    INTO probability, joint_treewidth, data_treewidth_lb,
         circuit_treewidth_lb, n_bags, max_states, dd_size, n_enumerating
    FROM ucq_joint_compile_stats(dnv, adisj, arel, avars, aarity,
      fact_rel, fact_elems, fact_arity, fact_tokens, fact_probs) s;
END;
$$ LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE;
 
 
 
 
 
 
 
 
 
 
 
/**
 * @brief Correlated Boolean UCQ probability plus compilation statistics
 * (columnar form, internal)
 *
 * Same compilation as @c ucq_joint_evaluate_tracked(); the three width
 * columns substantiate thesis Prop. 4.2.11 on real correlated data (the
 * data-only and circuit-only degeneracy bounds can be small while the
 * joint width is large).
 */
CREATE OR REPLACE FUNCTION ucq_joint_compile_stats_tracked(
  IN disjunct_nvars INT[],
  IN atom_disjunct INT[],
  IN atom_rel INT[],
  IN atom_vars INT[],
  IN atom_arity INT[],
  IN fact_rel INT[],
  IN fact_elems INT[],
  IN fact_arity INT[],
  IN fact_tokens UUID[],
  OUT probability DOUBLE PRECISION,
  OUT joint_treewidth INT,
  OUT data_treewidth_lb INT,
  OUT circuit_treewidth_lb INT,
  OUT n_bags BIGINT,
  OUT max_states BIGINT,
  OUT dd_size BIGINT,
  OUT n_enumerating INT)
  AS 'provsql','ucq_joint_compile_stats_tracked'
  LANGUAGE C STABLE PARALLEL SAFE;
 
 
/**
 * @brief Correlated Boolean UCQ probability plus statistics from a JSON spec
 */
CREATE OR REPLACE FUNCTION ucq_joint_compile_stats_tracked(
  IN query JSONB,
  IN fact_rel INT[],
  IN fact_elems INT[],
  IN fact_arity INT[],
  IN fact_tokens UUID[],
  OUT probability DOUBLE PRECISION,
  OUT joint_treewidth INT,
  OUT data_treewidth_lb INT,
  OUT circuit_treewidth_lb INT,
  OUT n_bags BIGINT,
  OUT max_states BIGINT,
  OUT dd_size BIGINT,
  OUT n_enumerating INT)
  AS $$
DECLARE
  dnv INT[] := '{}'; adisj INT[] := '{}'; arel INT[] := '{}';
  avars INT[] := '{}'; aarity INT[] := '{}';
  d JSONB; a JSONB; v TEXT; didx INT := 0;
BEGIN
  FOR d IN SELECT * FROM jsonb_array_elements(query->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx;
      arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
  SELECT s.probability, s.joint_treewidth, s.data_treewidth_lb,
         s.circuit_treewidth_lb, s.n_bags, s.max_states, s.dd_size,
         s.n_enumerating
    INTO probability, joint_treewidth, data_treewidth_lb,
         circuit_treewidth_lb, n_bags, max_states, dd_size, n_enumerating
    FROM ucq_joint_compile_stats_tracked(dnv, adisj, arel, avars, aarity,
      fact_rel, fact_elems, fact_arity, fact_tokens) s;
END;
$$ LANGUAGE plpgsql STABLE PARALLEL SAFE;
 
/**
 * @brief Compile a correlated UCQ and materialise its certified d-D,
 * returning the root provenance token (columnar form, internal)
 *
 * The architecturally-primary route: the compiler builds the
 * deterministic, decomposable circuit and materialises it as ordinary
 * @c plus / @c times / @c monus provenance gates (carrying the d-DNNF
 * certificate); the answer is then obtained through the standard entry
 * points on the returned token -- @c probability_evaluate(token),
 * @c shapley(token, ...), expectation -- so the joint-width path shares
 * the one evaluation pipeline.  The token is the exact Boolean
 * provenance of the UCQ (no @c 'absorptive' marker).
 */
CREATE OR REPLACE FUNCTION ucq_joint_materialize_tracked(
  disjunct_nvars INT[],
  atom_disjunct INT[],
  atom_rel INT[],
  atom_vars INT[],
  atom_arity INT[],
  fact_rel INT[],
  fact_elems INT[],
  fact_arity INT[],
  fact_tokens UUID[])
  RETURNS UUID AS
  'provsql','ucq_joint_materialize_tracked' LANGUAGE C VOLATILE;
 
/**
 * @brief Compile a correlated UCQ and materialise its certified d-D
 * from a JSON spec, returning the root provenance token
 *
 * JSON-spec wrapper over @c ucq_joint_materialize_tracked().  Evaluate
 * the answer with the standard surface, e.g.
 * @c probability_evaluate(ucq_joint_materialize_tracked(query, ...)).
 */
CREATE OR REPLACE FUNCTION ucq_joint_materialize_tracked(
  query JSONB,
  fact_rel INT[],
  fact_elems INT[],
  fact_arity INT[],
  fact_tokens UUID[])
  RETURNS UUID AS $$
DECLARE
  dnv INT[] := '{}'; adisj INT[] := '{}'; arel INT[] := '{}';
  avars INT[] := '{}'; aarity INT[] := '{}';
  d JSONB; a JSONB; v TEXT; didx INT := 0;
BEGIN
  FOR d IN SELECT * FROM jsonb_array_elements(query->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx;
      arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
  RETURN ucq_joint_materialize_tracked(dnv, adisj, arel, avars, aarity,
    fact_rel, fact_elems, fact_arity, fact_tokens);
END;
$$ LANGUAGE plpgsql VOLATILE;
 
/**
 * @brief Compile a UCQ over named relations into a materialised certified
 * d-D, gathering the facts from the store -- the descriptor-driven engine
 *
 * The query-surface bridge for the joint-width compiler: instead of
 * hand-built columnar arrays, a JSON @p descriptor names the relations
 * and how their columns map to query variables, and this function
 * gathers the facts itself (the provenance rewriting is disabled around
 * the gather), builds the value-based element dictionary shared across
 * the relations (so equal join values get the same dense id), compiles
 * and materialises the certified d-D, and returns its provenance token.
 * The answer is then any standard evaluation on that token --
 * @c probability_evaluate(ucq_joint_provenance(...)),
 * @c shapley(...), expectation.  This is also the engine the planner-time
 * query recogniser drives once it builds the descriptor from a query's
 * abstract syntax.
 *
 * Descriptor shape:
 * @verbatim
 * { "disjuncts": [ { "n_vars": k,
 *                    "atoms": [ {"rel": <relidx>, "vars": [..]}, ... ] }, ... ],
 *   "relations": [ "schema.r", "schema.s", ... ],   -- relidx -> relation
 *   "elem_cols": [ ["x"], ["x","y"], ... ] }         -- per relation: the
 *                                                       element columns, in
 *                                                       the atom's var order
 * @endverbatim
 *
 * @param descriptor the UCQ + the relations and their element columns
 * @param fallback token returned if the joint-width compiler declines
 * @return the materialised joint-width provenance token (NULL UUID-free
 *         exact Boolean provenance of the UCQ)
 */
CREATE OR REPLACE FUNCTION ucq_joint_provenance(
  descriptor JSONB, fallback UUID DEFAULT NULL)
RETURNS UUID AS $$
DECLARE
  legs TEXT; sql TEXT; saved TEXT;
  fact_rel INT[]; fact_elems INT[]; fact_arity INT[]; fact_tokens UUID[];
  dnv INT[]:='{}'; adisj INT[]:='{}'; arel INT[]:='{}';
  avars INT[]:='{}'; aarity INT[]:='{}';
  d jsonb; a jsonb; v TEXT; didx INT:=0;
BEGIN
  -- Parse the UCQ structure into the columnar query arrays.
  FOR d IN SELECT * FROM jsonb_array_elements(descriptor->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx; arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
 
  -- One UNION ALL leg per relation: (relation index, TEXT element array,
  -- provenance token).  No temp tables: a single gather query, with the
  -- value-based dense element dictionary built inline.
  SELECT string_agg(
           format('SELECT %s, ARRAY[%s]::TEXT[], provsql FROM %s%s',
             rn - 1,
             (SELECT string_agg(format('(%I)::TEXT', c), ',')
                FROM jsonb_array_elements_text(descriptor->'elem_cols'->(rn-1)::INT) c),
             rel,
             -- the lifted single-relation selection (a pre-filter), already
             -- deparsed to SQL by the recogniser; '' / absent = unfiltered.
             CASE WHEN coalesce(descriptor->'rel_where'->>(rn-1)::INT,'') <> ''
                  THEN ' WHERE '||(descriptor->'rel_where'->>(rn-1)::INT)
                  ELSE '' END),
           ' UNION ALL ')
    INTO legs
    FROM jsonb_array_elements_text(descriptor->'relations') WITH ORDINALITY t(rel, rn);
 
  sql := format($q$
    WITH facts(rel,elems,tok) AS (%s),
         ord AS (SELECT row_number() OVER () AS ord, rel, elems, tok FROM facts),
         dict AS (SELECT val, (dense_rank() OVER (ORDER BY val))-1 AS id
                    FROM (SELECT DISTINCT unnest(elems) AS val FROM facts) u)
    SELECT (SELECT array_agg(rel ORDER BY ord) FROM ord),
           (SELECT array_agg(cardinality(elems) ORDER BY ord) FROM ord),
           (SELECT array_agg(tok ORDER BY ord) FROM ord),
           (SELECT array_agg(dd.id ORDER BY o.ord, e.k)
              FROM ord o, LATERAL unnest(o.elems) WITH ORDINALITY e(val,k)
              JOIN dict dd ON dd.val = e.val)
  $q$, legs);
 
  -- Read the raw rows with provenance rewriting disabled (we only read
  -- the existing provsql column; this internal gather is not tracked).
  saved := current_setting('provsql.active', true);
  PERFORM set_config('provsql.active','off', true);
  EXECUTE sql INTO fact_rel, fact_arity, fact_tokens, fact_elems;
  PERFORM set_config('provsql.active', saved, true);
 
  RETURN ucq_joint_materialize_tracked(dnv,adisj,arel,avars,aarity,
    fact_rel,fact_elems,fact_arity,fact_tokens);
EXCEPTION WHEN OTHERS THEN
  -- The joint-width compiler declined (unsupported gate type, joint
  -- width too large, ...): fall back to the normal provenance so the
  -- query never fails.  Both give the same probability.
  RETURN fallback;
END;
$$ LANGUAGE plpgsql VOLATILE;
 
-- ===========================================================================
-- Safe-UCQ Möbius-inversion route (mobius_evaluate.cpp).
--
-- The last missing exact route of the Dalvi-Suciu dichotomy: UCQs that are
-- safe only because the \#P-hard terms of their inclusion-exclusion expansion
-- carry a zero Möbius value on the CNF lattice and cancel (canonical witness:
-- QW / q9).  Same TID gather as ucq_joint, then the lattice-walking compiler
-- materialises a gate_mobius-rooted circuit (a signed combination over
-- certified-independent islands), answered in PTIME data complexity by the
-- standard probability path.
-- ===========================================================================
 
/**
 * @brief Materialise the safe-UCQ Möbius circuit and return its root token.
 *        Columnar (TID) interface; see ucq_mobius_provenance for the gather.
 */
CREATE OR REPLACE FUNCTION ucq_mobius_materialize_tracked(
  disjunct_nvars INT[],
  atom_disjunct INT[],
  atom_rel INT[],
  atom_vars INT[],
  atom_arity INT[],
  fact_rel INT[],
  fact_elems INT[],
  fact_arity INT[],
  fact_tokens UUID[],
  lineage UUID DEFAULT NULL)
  RETURNS UUID AS
  'provsql','ucq_mobius_materialize_tracked' LANGUAGE C VOLATILE;
 
/**
 * @brief Compile the Möbius circuit and return the lattice statistics plus the
 *        probability (the demonstrability surface).  @c cancelled_hard is the
 *        single number that makes the mechanism legible: for q9 the 1 cancelled
 *        element is \#P-hard, so the query is easy only because its hard part
 *        cancels.
 */
CREATE OR REPLACE FUNCTION ucq_mobius_compile_stats(
  IN disjunct_nvars INT[],
  IN atom_disjunct INT[],
  IN atom_rel INT[],
  IN atom_vars INT[],
  IN atom_arity INT[],
  IN fact_rel INT[],
  IN fact_elems INT[],
  IN fact_arity INT[],
  IN fact_tokens UUID[],
  OUT probability DOUBLE PRECISION,
  OUT n_components INT,
  OUT n_cnf_conjuncts INT,
  OUT lattice_size INT,
  OUT n_nonzero INT,
  OUT n_cancelled INT,
  OUT cancelled_hard BOOLEAN,
  OUT dd_size BIGINT,
  OUT memo_hits BIGINT)
  AS 'provsql','ucq_mobius_compile_stats'
  LANGUAGE C VOLATILE;
 
/**
 * @brief Pass a token through iff it is a @c gate_mobius, else return NULL.
 *
 * The Möbius-precedence dispatch (see @c make_provenance_expression) wraps the
 * Möbius call in this and then @c COALESCE\ s it before the joint-width call:
 * a Möbius *success* always roots a @c gate_mobius (the compiler wraps even a
 * thin selector around the lineage), so it short-circuits and the joint-width
 * compiler never runs; a Möbius *decline* returns the literal lineage (never a
 * @c gate_mobius), so this yields NULL and @c COALESCE falls through to
 * joint-width.  The lineage token is a plain plus/times/input, so the test is
 * unambiguous.
 */
CREATE OR REPLACE FUNCTION mobius_or_null(tok UUID)
RETURNS UUID AS $$
  SELECT CASE WHEN tok IS NOT NULL AND provsql.get_gate_type(tok) = 'mobius'
              THEN tok END
$$ LANGUAGE sql STABLE;
 
/**
 * @brief Möbius-route provenance from a descriptor (the planner-substituted
 *        entry point, and the manual one).  Same descriptor and TID gather as
 *        @c ucq_joint_provenance; on any decline (unsafe shape, cap, not TID)
 *        returns @p fallback, so a recognised query never fails.
 */
CREATE OR REPLACE FUNCTION ucq_mobius_provenance(
  descriptor JSONB, fallback UUID DEFAULT NULL)
RETURNS UUID AS $$
DECLARE
  legs TEXT; sql TEXT; saved TEXT;
  fact_rel INT[]; fact_elems INT[]; fact_arity INT[]; fact_tokens UUID[];
  dnv INT[]:='{}'; adisj INT[]:='{}'; arel INT[]:='{}';
  avars INT[]:='{}'; aarity INT[]:='{}';
  d jsonb; a jsonb; v TEXT; didx INT:=0;
BEGIN
  FOR d IN SELECT * FROM jsonb_array_elements(descriptor->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx; arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
 
  SELECT string_agg(
           format('SELECT %s, ARRAY[%s]::TEXT[], provsql FROM %s%s',
             rn - 1,
             (SELECT string_agg(format('(%I)::TEXT', c), ',')
                FROM jsonb_array_elements_text(descriptor->'elem_cols'->(rn-1)::INT) c),
             rel,
             CASE WHEN coalesce(descriptor->'rel_where'->>(rn-1)::INT,'') <> ''
                  THEN ' WHERE '||(descriptor->'rel_where'->>(rn-1)::INT)
                  ELSE '' END),
           ' UNION ALL ')
    INTO legs
    FROM jsonb_array_elements_text(descriptor->'relations') WITH ORDINALITY t(rel, rn);
 
  sql := format($q$
    WITH facts(rel,elems,tok) AS (%s),
         ord AS (SELECT row_number() OVER () AS ord, rel, elems, tok FROM facts),
         dict AS (SELECT val, (dense_rank() OVER (ORDER BY val))-1 AS id
                    FROM (SELECT DISTINCT unnest(elems) AS val FROM facts) u)
    SELECT (SELECT array_agg(rel ORDER BY ord) FROM ord),
           (SELECT array_agg(cardinality(elems) ORDER BY ord) FROM ord),
           (SELECT array_agg(tok ORDER BY ord) FROM ord),
           (SELECT array_agg(dd.id ORDER BY o.ord, e.k)
              FROM ord o, LATERAL unnest(o.elems) WITH ORDINALITY e(val,k)
              JOIN dict dd ON dd.val = e.val)
  $q$, legs);
 
  saved := current_setting('provsql.active', true);
  PERFORM set_config('provsql.active','off', true);
  EXECUTE sql INTO fact_rel, fact_arity, fact_tokens, fact_elems;
  PERFORM set_config('provsql.active', saved, true);
 
  -- Pass the normal-provenance fallback as the lineage: it is carried on the
  -- gate_mobius so the token still answers Shapley / semiring / PROV on the
  -- literal lineage (the Möbius combination is a probability-only shortcut).
  RETURN ucq_mobius_materialize_tracked(dnv,adisj,arel,avars,aarity,
    fact_rel,fact_elems,fact_arity,fact_tokens, fallback);
EXCEPTION WHEN OTHERS THEN
  RETURN fallback;
END;
$$ LANGUAGE plpgsql VOLATILE;
 
/**
 * @brief Möbius lattice statistics + probability from a descriptor: the
 *        demonstrability SRF.  Gathers
 *        the same TID facts as @c ucq_mobius_provenance, then runs the columnar
 *        @c ucq_mobius_compile_stats.
 */
CREATE OR REPLACE FUNCTION mobius_compile_stats(
  IN descriptor JSONB,
  OUT probability DOUBLE PRECISION,
  OUT n_components INT,
  OUT n_cnf_conjuncts INT,
  OUT lattice_size INT,
  OUT n_nonzero INT,
  OUT n_cancelled INT,
  OUT cancelled_hard BOOLEAN,
  OUT dd_size BIGINT,
  OUT memo_hits BIGINT)
RETURNS RECORD AS $$
DECLARE
  legs TEXT; sql TEXT; saved TEXT;
  fact_rel INT[]; fact_elems INT[]; fact_arity INT[]; fact_tokens UUID[];
  dnv INT[]:='{}'; adisj INT[]:='{}'; arel INT[]:='{}';
  avars INT[]:='{}'; aarity INT[]:='{}';
  d jsonb; a jsonb; v TEXT; didx INT:=0;
BEGIN
  FOR d IN SELECT * FROM jsonb_array_elements(descriptor->'disjuncts') LOOP
    dnv := dnv || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      adisj := adisj || didx; arel := arel || (a->>'rel')::INT;
      aarity := aarity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        avars := avars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
 
  SELECT string_agg(
           format('SELECT %s, ARRAY[%s]::TEXT[], provsql FROM %s%s',
             rn - 1,
             (SELECT string_agg(format('(%I)::TEXT', c), ',')
                FROM jsonb_array_elements_text(descriptor->'elem_cols'->(rn-1)::INT) c),
             rel,
             CASE WHEN coalesce(descriptor->'rel_where'->>(rn-1)::INT,'') <> ''
                  THEN ' WHERE '||(descriptor->'rel_where'->>(rn-1)::INT)
                  ELSE '' END),
           ' UNION ALL ')
    INTO legs
    FROM jsonb_array_elements_text(descriptor->'relations') WITH ORDINALITY t(rel, rn);
 
  sql := format($q$
    WITH facts(rel,elems,tok) AS (%s),
         ord AS (SELECT row_number() OVER () AS ord, rel, elems, tok FROM facts),
         dict AS (SELECT val, (dense_rank() OVER (ORDER BY val))-1 AS id
                    FROM (SELECT DISTINCT unnest(elems) AS val FROM facts) u)
    SELECT (SELECT array_agg(rel ORDER BY ord) FROM ord),
           (SELECT array_agg(cardinality(elems) ORDER BY ord) FROM ord),
           (SELECT array_agg(tok ORDER BY ord) FROM ord),
           (SELECT array_agg(dd.id ORDER BY o.ord, e.k)
              FROM ord o, LATERAL unnest(o.elems) WITH ORDINALITY e(val,k)
              JOIN dict dd ON dd.val = e.val)
  $q$, legs);
 
  saved := current_setting('provsql.active', true);
  PERFORM set_config('provsql.active','off', true);
  EXECUTE sql INTO fact_rel, fact_arity, fact_tokens, fact_elems;
  PERFORM set_config('provsql.active', saved, true);
 
  SELECT s.probability, s.n_components, s.n_cnf_conjuncts, s.lattice_size,
         s.n_nonzero, s.n_cancelled, s.cancelled_hard, s.dd_size, s.memo_hits
    INTO probability, n_components, n_cnf_conjuncts, lattice_size,
         n_nonzero, n_cancelled, cancelled_hard, dd_size, memo_hits
    FROM ucq_mobius_compile_stats(dnv,adisj,arel,avars,aarity,
      fact_rel,fact_elems,fact_arity,fact_tokens) s;
END;
$$ LANGUAGE plpgsql VOLATILE;
 
/**
 * @brief Internal gather for the per-answer joint route: parse @p descriptor
 *        into the columnar UCQ arrays and gather every fact (relation index,
 *        dense element ids, provenance token) with the value dictionary.
 *
 * Used only by the planner-substituted @c ucq_joint_provenance_answer (the C
 * single-DP entry point), which calls it ONCE per query and then computes all
 * answers in one sweep.  No head pinning: the single DP discovers the answers.
 * @c val_by_id maps a dense element id back to its TEXT value (so an answer's
 * head ids can be matched to the @c GROUP @c BY head TEXT).
 */
CREATE OR REPLACE FUNCTION ucq_joint_gather(
  descriptor JSONB,
  OUT disjunct_nvars INT[], OUT atom_disjunct INT[], OUT atom_rel INT[],
  OUT atom_vars INT[], OUT atom_arity INT[],
  OUT fact_rel INT[], OUT fact_elems INT[], OUT fact_arity INT[],
  OUT fact_tokens UUID[], OUT val_by_id TEXT[])
AS $$
DECLARE
  legs TEXT; sql TEXT; saved TEXT; d jsonb; a jsonb; v TEXT; didx INT := 0;
BEGIN
  disjunct_nvars:='{}'; atom_disjunct:='{}'; atom_rel:='{}';
  atom_vars:='{}'; atom_arity:='{}';
  FOR d IN SELECT * FROM jsonb_array_elements(descriptor->'disjuncts') LOOP
    disjunct_nvars := disjunct_nvars || (d->>'n_vars')::INT;
    FOR a IN SELECT * FROM jsonb_array_elements(d->'atoms') LOOP
      atom_disjunct := atom_disjunct || didx;
      atom_rel := atom_rel || (a->>'rel')::INT;
      atom_arity := atom_arity || jsonb_array_length(a->'vars');
      FOR v IN SELECT * FROM jsonb_array_elements_text(a->'vars') LOOP
        atom_vars := atom_vars || v::INT;
      END LOOP;
    END LOOP;
    didx := didx + 1;
  END LOOP;
 
  SELECT string_agg(
           format('SELECT %s, ARRAY[%s]::TEXT[], provsql FROM %s%s', rn - 1,
             (SELECT string_agg(format('(%I)::TEXT', c), ',')
                FROM jsonb_array_elements_text(descriptor->'elem_cols'->(rn-1)::INT) c),
             rel,
             CASE WHEN coalesce(descriptor->'rel_where'->>(rn-1)::INT,'') <> ''
                  THEN ' WHERE '||(descriptor->'rel_where'->>(rn-1)::INT)
                  ELSE '' END),
           ' UNION ALL ')
    INTO legs
    FROM jsonb_array_elements_text(descriptor->'relations') WITH ORDINALITY t(rel, rn);
 
  sql := format($q$
    WITH facts(rel,elems,tok) AS (%s),
         ord AS (SELECT row_number() OVER () AS ord, rel, elems, tok FROM facts),
         dict AS (SELECT val, (dense_rank() OVER (ORDER BY val))-1 AS id
                    FROM (SELECT DISTINCT unnest(elems) AS val FROM facts) u)
    SELECT (SELECT array_agg(rel ORDER BY ord) FROM ord),
           (SELECT array_agg(cardinality(elems) ORDER BY ord) FROM ord),
           (SELECT array_agg(tok ORDER BY ord) FROM ord),
           (SELECT array_agg(dd.id ORDER BY o.ord, e.k)
              FROM ord o, LATERAL unnest(o.elems) WITH ORDINALITY e(val,k)
              JOIN dict dd ON dd.val = e.val),
           (SELECT array_agg(val ORDER BY id) FROM dict)
  $q$, legs);
 
  saved := current_setting('provsql.active', true);
  PERFORM set_config('provsql.active','off', true);
  EXECUTE sql INTO fact_rel, fact_arity, fact_tokens, fact_elems, val_by_id;
  PERFORM set_config('provsql.active', saved, true);
END;
$$ LANGUAGE plpgsql VOLATILE;
 
/**
 * @brief Per-answer joint-width provenance via the TOP-DOWN single DP
 *        (planner-substituted, C).
 *
 * The transparent per-answer rewrite substitutes one call per output group.
 * On the FIRST call of a query the function gathers the facts once
 * (@c ucq_joint_gather), runs the single DP, and materialises EVERY answer's
 * certified d-D into the store, caching @c head_vals -> token in @c fn_extra;
 * each subsequent group call is an O(1) lookup -- so the whole GROUP BY costs
 * one gather + one decomposition + one sweep, not @p k of each.  On any
 * decline (joint width too large) the @p fallback token (the normal
 * per-answer provenance) is returned, so the query never fails.  The answer's
 * marginal / Shapley / expectation is then the standard evaluation on the
 * returned token -- one pipeline for the whole system.
 */
CREATE OR REPLACE FUNCTION ucq_joint_provenance_answer(
  descriptor JSONB, head_vars INT[], head_vals TEXT[], fallback UUID DEFAULT NULL)
RETURNS UUID AS 'provsql','ucq_joint_provenance_answer'
LANGUAGE C STABLE;
 
/**
 * @brief Per-answer safe-UCQ Möbius provenance (planner-substituted): one
 *        head-pinned Möbius circuit per output group.  On the first call the
 *        facts are gathered once (ucq_joint_gather) and cached; each group pins
 *        @p head_vars to @p head_vals and compiles, caching head -> token.  On
 *        any decline returns @p fallback.  STABLE: it caches per fn-call
 *        context, so it is not re-evaluated within one scan.
 */
CREATE OR REPLACE FUNCTION ucq_mobius_provenance_answer(
  descriptor JSONB, head_vars INT[], head_vals TEXT[], fallback UUID DEFAULT NULL)
RETURNS UUID AS 'provsql','ucq_mobius_provenance_answer'
LANGUAGE C STABLE;
 
 
/**
 * @brief Compile and materialise the reachability provenance of every
 * vertex (columnar form, internal)
 *
 * All-targets variant of @c reachability_evaluate(): compiles, along a
 * tree decomposition of the data graph, one certified provenance
 * circuit per vertex reachable from some source in the all-edges-present
 * world, materialises the (shared, linear-size) circuits in the
 * provenance store -- @c plus / @c times gates carrying the d-DNNF
 * certificate, negated edges as @c monus(one, edge) -- and returns one
 * @c (vertex, token) row per such vertex.  Sources form a possibly
 * *probabilistic source set*: each source arc is gated by the source
 * tuple's token, the nil UUID marking a certain (always present)
 * source.  This is the engine behind the rewriter's
 * recursive-reachability route; the returned tokens are ordinary
 * provenance tokens usable with the whole evaluation surface, wrapped
 * in the 'absorptive' assumption marker (the compiled circuit is the
 * exact Boolean lineage but only the absorptive quotient of the
 * infinite recursive semiring provenance: probability and absorptive
 * semiring evaluations -- e.g. nonnegative min-plus -- are exact,
 * counting and why-provenance refuse).
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param block_keys per-edge BID key variable (nil UUID = independent
 *        tuple; alternatives sharing a key are mutually exclusive, e.g.
 *        from repair_key)
 * @param block_indices per-edge outcome index within its block
 * @param source_vertices the source vertices
 * @param source_tokens per-source provenance token (nil UUID = certain)
 * @param source_probabilities per-source probability
 * @param directed if false, each edge can be traversed both ways
 * @param[out] vertex a vertex reachable from some source
 * @param[out] token the materialised reachability provenance token of @c vertex
 */
CREATE OR REPLACE FUNCTION reachability_materialize(
  IN sources INT[],
  IN destinations INT[],
  IN tokens UUID[],
  IN probabilities DOUBLE PRECISION[],
  IN block_keys UUID[],
  IN block_indices INT[],
  IN source_vertices INT[],
  IN source_tokens UUID[],
  IN source_probabilities DOUBLE PRECISION[],
  IN directed BOOLEAN,
  OUT vertex INT,
  OUT token UUID)
  RETURNS SETOF RECORD AS
  'provsql','reachability_materialize' LANGUAGE C VOLATILE;
 
 
/**
 * @brief Bounded-hop variant of @c reachability_materialize() (internal)
 *
 * Compiles, along a tree decomposition of the data graph, one certified
 * provenance circuit per (vertex, walk length) pair achievable within
 * @p hop_bound edges -- the rows a hop-counting recursive CTE derives,
 * row @c (v,h) meaning "some *walk* of exactly @c h edges connects a
 * present source to @c v" -- and returns them as @c (vertex, hops,
 * token) with @p hop_seed added to the lengths (the CTE base arm's hop
 * constant).  Also pre-creates, per vertex, the certified gate that a
 * hop-discarding query's deduplication will address, wired to the
 * compilation's native within-bound root, so the natural "within k
 * hops" probability evaluates through the linear certified route.
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param block_keys per-edge BID key variable (nil UUID = independent)
 * @param block_indices per-edge outcome index within its block
 * @param source_vertices the source vertices
 * @param source_tokens per-source provenance token (nil UUID = certain)
 * @param source_probabilities per-source probability
 * @param directed if false, each edge can be traversed both ways
 * @param hop_bound maximum walk length
 * @param hop_seed hop value of the base arm (added to reported lengths)
 * @param[out] vertex a reachable vertex
 * @param[out] hops the walk length at which @c vertex is reached
 * @param[out] token the materialised provenance token of the @c (vertex, hops) pair
 */
CREATE OR REPLACE FUNCTION reachability_materialize_hops(
  IN sources INT[],
  IN destinations INT[],
  IN tokens UUID[],
  IN probabilities DOUBLE PRECISION[],
  IN block_keys UUID[],
  IN block_indices INT[],
  IN source_vertices INT[],
  IN source_tokens UUID[],
  IN source_probabilities DOUBLE PRECISION[],
  IN directed BOOLEAN,
  IN hop_bound INT,
  IN hop_seed INT,
  OUT vertex INT,
  OUT hops INT,
  OUT token UUID)
  RETURNS SETOF RECORD AS
  'provsql','reachability_materialize_hops' LANGUAGE C VOLATILE;
 
 
/**
 * @brief Per-group "some member reachable" compilation (columnar form,
 * internal)
 *
 * For each distinct group in the parallel @p group_ids /
 * @p member_vertices arrays, compiles the certified circuit of "some
 * member vertex is reachable from a present source" along the data
 * decomposition -- the disjunction over the group's *correlated*
 * per-vertex reachability events, deterministic by construction
 * through the set-reachability state bit -- materialises it, and
 * returns one @c (group_id, token) row per group.  Engine behind the
 * rewriter's cross-vertex aggregation planting.
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param block_keys per-edge BID key variable (nil UUID = independent)
 * @param block_indices per-edge outcome index within its block
 * @param source_vertices the source vertices
 * @param source_tokens per-source provenance token (nil UUID = certain)
 * @param source_probabilities per-source probability
 * @param directed if false, each edge can be traversed both ways
 * @param group_ids group identifier of each member row
 * @param member_vertices member vertex of each member row
 * @param[out] group_id a group whose every member is reachable
 * @param[out] token the materialised all-members-reachable provenance token of
 *        @c group_id
 */
CREATE OR REPLACE FUNCTION reachability_materialize_any(
  IN sources INT[],
  IN destinations INT[],
  IN tokens UUID[],
  IN probabilities DOUBLE PRECISION[],
  IN block_keys UUID[],
  IN block_indices INT[],
  IN source_vertices INT[],
  IN source_tokens UUID[],
  IN source_probabilities DOUBLE PRECISION[],
  IN directed BOOLEAN,
  IN group_ids INT[],
  IN member_vertices INT[],
  OUT group_id INT,
  OUT token UUID)
  RETURNS SETOF RECORD AS
  'provsql','reachability_materialize_any' LANGUAGE C VOLATILE;
 
/**
 * @brief Compile and materialise the "every member vertex reachable"
 * (k-terminal / coverage) circuit (columnar form, internal)
 *
 * Arguments as @c reachability_materialize_any() with a single member
 * set: compiles the certified circuit of "every member vertex is
 * reachable from a present source" -- the conjunction over the
 * members' *correlated* per-vertex events, deterministic by
 * construction through the pending rescuer-set congruence --
 * materialises it, and returns its token, wrapped in the
 * @c 'absorptive' assumption marker.  Probability evaluation gives the
 * k-terminal reliability; nonnegative min-plus the cost of the
 * cheapest covering subgraph (directed Steiner cost), shared edges
 * paid once.  A member vertex absent from the graph is unreachable:
 * the circuit is then constant false.
 *
 * @param sources source vertex of each edge (dense INTEGER IDs)
 * @param destinations destination vertex of each edge
 * @param tokens provenance token of each edge tuple
 * @param probabilities probability of each edge tuple
 * @param block_keys per-edge BID key variable (nil UUID = independent)
 * @param block_indices per-edge outcome index within its block
 * @param source_vertices the source vertices
 * @param source_tokens per-source provenance token (nil UUID = certain)
 * @param source_probabilities per-source probability
 * @param directed if false, each edge can be traversed both ways
 * @param member_vertices the member vertices (dense IDs)
 */
CREATE OR REPLACE FUNCTION reachability_materialize_cover(
  sources INT[],
  destinations INT[],
  tokens UUID[],
  probabilities DOUBLE PRECISION[],
  block_keys UUID[],
  block_indices INT[],
  source_vertices INT[],
  source_tokens UUID[],
  source_probabilities DOUBLE PRECISION[],
  directed BOOLEAN,
  member_vertices INT[])
  RETURNS UUID AS
  'provsql','reachability_materialize_cover' LANGUAGE C VOLATILE;
 
/**
 * @brief Plant certified any-member-reachable gates for a grouped
 * reachability aggregation (internal)
 *
 * Called (at plan time, over SPI) by the recursive-CTE lowering when
 * the outer query aggregates a reachability working table by a column
 * of a joined, untracked member relation: @c GROUP @c BY collapses
 * each group's per-vertex reach tokens with @c provenance_plus, whose
 * disjuncts are correlated (they share edges) and would otherwise
 * leave the certified route.  For each multi-member group this
 * pre-creates, at the canonical address of the group's token multiset,
 * a certified single-child plus over the group's native
 * any-member-reachable circuit (@c reachability_materialize_any), so
 * the natural aggregation stays on the linear evaluation route.
 * Best-effort: any failure leaves the generic path untouched (notice
 * under verbosity 10).
 *
 * @param work_name the lowered CTE's working table
 * @param node_attribute its vertex column
 * @param member_rel the joined member relation (must be untracked)
 * @param member_attribute the member relation's join column
 * @param group_attribute the member relation's grouping column
 * @param edge_rel the tracked edge relation (as for eval_reachability)
 * @param source_attribute name of the source-vertex column
 * @param destination_attribute name of the destination-vertex column
 * @param source_value the base arm's constant, as TEXT
 * @param directed if false, each edge can be traversed both ways
 * @param edge_quals optional deterministic filter over edge columns
 * @param source_rel source relation of a multi-source base arm
 * @param source_rel_attribute the source relation's vertex column
 * @param edge_sql deparsed edge subquery (join-defined edges)
 * @param member_quals optional deterministic filter over the member
 *        relation's columns (table-qualified as @c t.column), restricting
 *        which members participate in each group
 */
CREATE OR REPLACE FUNCTION plant_reach_any_groups(
  work_name TEXT,
  node_attribute TEXT,
  member_rel REGCLASS,
  member_attribute TEXT,
  group_attribute TEXT,
  edge_rel REGCLASS,
  source_attribute TEXT,
  destination_attribute TEXT,
  source_value TEXT,
  directed BOOLEAN,
  edge_quals TEXT DEFAULT NULL,
  source_rel REGCLASS DEFAULT NULL,
  source_rel_attribute TEXT DEFAULT NULL,
  edge_sql TEXT DEFAULT NULL,
  member_quals TEXT DEFAULT NULL)
  RETURNS VOID AS
$$
DECLARE
  e RECORD;
  grp RECORD;
  m RECORD;
  sv TEXT[];
  st UUID[];
  sp double precision[];
  gids INT[] := ARRAY[]::INT[];
  mids INT[] := ARRAY[]::INT[];
  vid INT;
  canonical UUID;
  verbosity INT := coalesce(current_setting('provsql.verbose_level', true)::INT, 0);
BEGIN
  BEGIN
    -- A tracked member relation would make the aggregated tokens
    -- per-row products, not the bare reach tokens: nothing to plant.
    IF EXISTS (SELECT 1 FROM pg_attribute
               WHERE attrelid = member_rel AND attname = 'provsql'
                 AND atttypid = 'UUID'::REGTYPE AND NOT attisdropped) THEN
      RETURN;
    END IF;
 
    IF source_rel IS NOT NULL THEN
      SELECT g.source_values, g.source_tokens, g.source_probabilities
        INTO sv, st, sp
        FROM provsql.gather_reachability_sources(source_rel,
                                                 source_rel_attribute) g;
      IF sv IS NULL THEN
        sv := ARRAY[]::TEXT[];
        st := ARRAY[]::UUID[];
        sp := ARRAY[]::float8[];
      END IF;
    ELSE
      sv := ARRAY[source_value];
      st := ARRAY['00000000-0000-0000-0000-000000000000'::UUID];
      sp := ARRAY[1.0::float8];
    END IF;
 
    e := provsql.gather_reachability_edges(edge_rel, source_attribute,
                                           destination_attribute,
                                           sv, edge_quals, edge_sql);
 
    -- The groups, replicating the user's join semantics: per group, the
    -- member vertices and the multiset of their reach tokens (with the
    -- multiplicity the join produces).  Single-member groups need no
    -- planting (provenance_plus passes a single token through).
    -- Two steps: materialise the joined rows with their per-row tokens
    -- (tracked CTAS, then strip the automatic provsql column), and only
    -- then aggregate the now-plain table -- aggregating provenance()
    -- inside a grouped tracked query would be rewritten as a
    -- provenance-aware aggregation, which is not what the planting
    -- needs.
    DROP TABLE IF EXISTS provsql_reach_any_flat_tmp;
    EXECUTE format(
      'CREATE TEMP TABLE provsql_reach_any_flat_tmp AS '
      || 'SELECT w.%1$I::TEXT AS node_val, provsql.provenance() AS tok, '
      || '       t.%5$I AS grp_key '
      || 'FROM %2$I w JOIN %3$s t ON w.%1$I = t.%4$I'
      -- The member-relation filter restricts which members participate
      -- (deparsed table-qualified as t.column); the working table side
      -- carries no provenance distinction here.
      || coalesce(' WHERE ' || member_quals, ''),
      node_attribute, work_name, member_rel::TEXT, member_attribute,
      group_attribute);
    PERFORM provsql.remove_provenance('provsql_reach_any_flat_tmp');
    DROP TABLE IF EXISTS provsql_reach_any_groups_tmp;
    CREATE TEMP TABLE provsql_reach_any_groups_tmp AS
      SELECT (row_number() OVER ())::INT AS gid, members, toks FROM (
        SELECT array_agg(node_val) AS members, array_agg(tok) AS toks
        FROM provsql_reach_any_flat_tmp
        GROUP BY grp_key HAVING count(*) >= 2) g;
    DROP TABLE provsql_reach_any_flat_tmp;
 
    FOR grp IN SELECT gid, members FROM provsql_reach_any_groups_tmp LOOP
      FOR m IN SELECT DISTINCT unnest(grp.members) AS val LOOP
        vid := array_position(e.vertices, m.val);
        IF vid IS NOT NULL THEN
          gids := gids || grp.gid;
          mids := mids || vid;
        END IF;
      END LOOP;
    END LOOP;
    IF cardinality(gids) = 0 THEN
      DROP TABLE provsql_reach_any_groups_tmp;
      RETURN;
    END IF;
 
    FOR grp IN
      SELECT a.group_id, a.token AS any_token, t.toks
      FROM provsql.reachability_materialize_any(
             e.sources, e.destinations, e.tokens, e.probabilities,
             e.block_keys, e.block_indices, e.extra_ids, st, sp,
             directed, gids, mids) a
      JOIN provsql_reach_any_groups_tmp t ON t.gid = a.group_id
    LOOP
      canonical := public.uuid_generate_v5(
        provsql.uuid_ns_provsql(),
        concat('plus-canonical',
               (SELECT array_agg(tok ORDER BY tok)
                FROM unnest(grp.toks) tok)));
      PERFORM provsql.create_gate(canonical, 'plus', ARRAY[grp.any_token]);
      PERFORM provsql.set_infos(canonical, 1);
    END LOOP;
    DROP TABLE provsql_reach_any_groups_tmp;
    IF verbosity >= 20 THEN
      -- Lift the function-level client_min_messages = warning for the
      -- one RAISE; the function-level SET restores the caller's value.
      PERFORM set_config('client_min_messages', 'notice', true);
      RAISE NOTICE 'ProvSQL: certified any-member gates planted for the aggregation of "%" by %.%',
        work_name, member_rel, group_attribute;
      PERFORM set_config('client_min_messages', 'warning', true);
    END IF;
  EXCEPTION WHEN OTHERS THEN
    IF verbosity >= 10 THEN
      PERFORM set_config('client_min_messages', 'notice', true);
      RAISE NOTICE 'ProvSQL: any-member planting for "%" skipped (%)',
        work_name, SQLERRM;
      PERFORM set_config('client_min_messages', 'warning', true);
    END IF;
  END;
END
-- No SET search_path: the deparsed edge subquery must resolve against
-- the caller's path; ProvSQL internals are schema-qualified.
$$ LANGUAGE plpgsql SET client_min_messages = warning;
 
/**
 * @brief Plant the certified all-members-reachable gate for a
 * reachability self-join conjunction (internal)
 *
 * Called (at plan time, over SPI) by the recursive-CTE lowering when
 * the outer query self-joins a reachability working table with one
 * constant node binding per reference -- "are these k vertices all
 * reachable" -- whose row provenance @c provenance_times() computes as
 * the product of *correlated* per-vertex reach tokens (they share
 * edges).  This pre-creates, at the times-canonical address of that
 * token multiset, a certified single-child times over the native
 * all-members-reachable circuit (@c reachability_materialize_cover),
 * so the natural conjunction stays on the linear certified route --
 * with the joint-worlds semantics: probability evaluation gives the
 * k-terminal reliability, and nonnegative min-plus the cost of the
 * cheapest covering subgraph (directed Steiner cost), shared edges
 * paid once where the raw product would pay them once per factor.
 * Best-effort: any failure leaves the generic path untouched (notice
 * under verbosity 10).
 *
 * @param work_name the lowered CTE's working table
 * @param node_attribute its vertex column
 * @param edge_rel the tracked edge relation (as for eval_reachability)
 * @param source_attribute name of the source-vertex column
 * @param destination_attribute name of the destination-vertex column
 * @param source_value the base arm's constant, as TEXT
 * @param directed if false, each edge can be traversed both ways
 * @param node_values the constant node bindings, as TEXT (multiset:
 *        one per self-join reference)
 * @param edge_quals optional deterministic filter over edge columns
 * @param source_rel source relation of a multi-source base arm
 * @param source_rel_attribute the source relation's vertex column
 * @param edge_sql deparsed edge subquery (join-defined edges)
 */
CREATE OR REPLACE FUNCTION plant_reach_cover(
  work_name TEXT,
  node_attribute TEXT,
  edge_rel REGCLASS,
  source_attribute TEXT,
  destination_attribute TEXT,
  source_value TEXT,
  directed BOOLEAN,
  node_values TEXT[],
  edge_quals TEXT DEFAULT NULL,
  source_rel REGCLASS DEFAULT NULL,
  source_rel_attribute TEXT DEFAULT NULL,
  edge_sql TEXT DEFAULT NULL)
  RETURNS VOID AS
$$
DECLARE
  e RECORD;
  sv TEXT[];
  st UUID[];
  sp double precision[];
  val TEXT;
  vid INT;
  vids INT[] := ARRAY[]::INT[];
  tok UUID;
  toks UUID[] := ARRAY[]::UUID[];
  cover_token UUID;
  canonical UUID;
  verbosity INT := coalesce(current_setting('provsql.verbose_level', true)::INT, 0);
BEGIN
  BEGIN
    IF source_rel IS NOT NULL THEN
      SELECT g.source_values, g.source_tokens, g.source_probabilities
        INTO sv, st, sp
        FROM provsql.gather_reachability_sources(source_rel,
                                                 source_rel_attribute) g;
      IF sv IS NULL THEN
        sv := ARRAY[]::TEXT[];
        st := ARRAY[]::UUID[];
        sp := ARRAY[]::float8[];
      END IF;
    ELSE
      sv := ARRAY[source_value];
      st := ARRAY['00000000-0000-0000-0000-000000000000'::UUID];
      sp := ARRAY[1.0::float8];
    END IF;
 
    e := provsql.gather_reachability_edges(edge_rel, source_attribute,
                                           destination_attribute,
                                           sv, edge_quals, edge_sql);
 
    -- The bound vertices and their per-row reach tokens, with the
    -- multiplicity the self-join produces.  A vertex absent from the
    -- graph, or from the working table, means the join is empty: no
    -- row will exist, nothing to plant.
    FOREACH val IN ARRAY node_values LOOP
      vid := array_position(e.vertices, val);
      IF vid IS NULL THEN
        RETURN;
      END IF;
      vids := vids || vid;
      EXECUTE format('SELECT provsql FROM %I WHERE %I::TEXT = $1',
                     work_name, node_attribute)
        INTO tok USING val;
      IF tok IS NULL THEN
        RETURN;
      END IF;
      toks := toks || tok;
    END LOOP;
 
    cover_token := provsql.reachability_materialize_cover(
      e.sources, e.destinations, e.tokens, e.probabilities,
      e.block_keys, e.block_indices, e.extra_ids, st, sp,
      directed, vids);
 
    SELECT public.uuid_generate_v5(
             provsql.uuid_ns_provsql(),
             concat('times-canonical', array_agg(t ORDER BY t)))
    FROM unnest(toks) t
    INTO canonical;
    PERFORM provsql.create_gate(canonical, 'times', ARRAY[cover_token]);
    PERFORM provsql.set_infos(canonical, 1);
    IF verbosity >= 20 THEN
      -- Lift the function-level client_min_messages = warning for the
      -- one RAISE; the function-level SET restores the caller's value.
      PERFORM set_config('client_min_messages', 'notice', true);
      RAISE NOTICE 'ProvSQL: certified all-members gate planted for the self-join of "%"',
        work_name;
      PERFORM set_config('client_min_messages', 'warning', true);
    END IF;
  EXCEPTION WHEN OTHERS THEN
    IF verbosity >= 10 THEN
      PERFORM set_config('client_min_messages', 'notice', true);
      RAISE NOTICE 'ProvSQL: all-members planting for "%" skipped (%)',
        work_name, SQLERRM;
      PERFORM set_config('client_min_messages', 'warning', true);
    END IF;
  END;
END
-- No SET search_path: the deparsed edge subquery must resolve against
-- the caller's path; ProvSQL internals are schema-qualified.
$$ LANGUAGE plpgsql SET client_min_messages = warning;
 
/**
 * @brief Input leaves of a conjunction-shaped provenance token (internal)
 *
 * Descends a token's circuit through the conjunctive gate types
 * (@c times, and the pass-through @c project / @c eq where-provenance
 * wrappers) down to @c input leaves.  Returns the distinct leaves, or
 * NULL when the circuit contains any other gate type (a disjunctive or
 * aggregate shape, which is not a conjunction of independent tuples).
 * Used by the reachability gathering to accept join-defined edges:
 * a derived edge whose token is a pure conjunction of base tuples.
 *
 * @param token the provenance token
 */
CREATE OR REPLACE FUNCTION token_conjunctive_leaves(token UUID)
  RETURNS UUID[] AS
$$
WITH RECURSIVE walk(g) AS (
  SELECT token
  UNION
  SELECT c FROM walk w, unnest(provsql.get_children(w.g)) AS c
  WHERE provsql.get_gate_type(w.g) IN ('times', 'project', 'eq', 'annotation')
)
SELECT CASE WHEN bool_and(provsql.get_gate_type(g)
                          IN ('times', 'project', 'eq', 'annotation', 'input'))
            THEN array_agg(DISTINCT g)
                   FILTER (WHERE provsql.get_gate_type(g) = 'input')
            ELSE NULL END
FROM walk;
$$ LANGUAGE sql STABLE;
 
/**
 * @brief Gather the edges of a tracked relation in the columnar form
 * expected by reachability_evaluate (internal)
 *
 * Materializes the edge relation with its provenance tokens and
 * probabilities, maps arbitrary vertex values (compared as TEXT) onto
 * dense INTEGER IDs, and checks that every edge tuple carries a base
 * input token (independent tuples): reachability compilation along the
 * data is only correct when the edges are independent events, so views
 * or query results with derived provenance are rejected.
 *
 * @param rel the provenance-tracked edge relation
 * @param source_attribute name of the source-vertex column
 * @param destination_attribute name of the destination-vertex column
 * @param extra_vertices vertex values (as TEXT) that must be part of
 *        the dense ID space even when they touch no edge -- the source
 *        set in particular; their IDs come back in @c extra_ids
 *        (aligned with the input)
 * @param edge_quals optional deterministic filter over the edge
 *        relation's columns (SQL TEXT, deparsed by the rewriter from
 *        the recursive arm's WHERE clause), restricting which edges
 *        participate
 * @param rel_sql deparsed edge subquery to gather from instead of
 *        @p rel (join-defined edges); the tokens are then conjunctions
 *        of base tuples, validated for shape and disjoint supports
 *
 * The @c vertices output maps the dense IDs back to the original
 * vertex values (as TEXT, 1-indexed), for callers that need to label
 * per-vertex results.
 *
 * @param[out] sources source vertex (dense ID) of each gathered edge
 * @param[out] destinations destination vertex (dense ID) of each edge
 * @param[out] tokens provenance token of each edge tuple
 * @param[out] probabilities probability of each edge tuple
 * @param[out] block_keys per-edge BID key variable (nil UUID = independent)
 * @param[out] block_indices per-edge outcome index within its block
 * @param[out] extra_ids dense IDs assigned to the @p extra_vertices
 * @param[out] vertices dense-ID-to-original-value map (TEXT, 1-indexed)
 */
CREATE OR REPLACE FUNCTION gather_reachability_edges(
  IN rel REGCLASS,
  IN source_attribute TEXT,
  IN destination_attribute TEXT,
  IN extra_vertices TEXT[],
  IN edge_quals TEXT DEFAULT NULL,
  IN rel_sql TEXT DEFAULT NULL,
  OUT sources INT[],
  OUT destinations INT[],
  OUT tokens UUID[],
  OUT probabilities DOUBLE PRECISION[],
  OUT block_keys UUID[],
  OUT block_indices INT[],
  OUT extra_ids INT[],
  OUT vertices TEXT[])
AS
$$
DECLARE
  tkind TEXT;
  bkey_expr TEXT;
  sel_probs TEXT;
  sel_bkeys TEXT;
  sel_bidx TEXT;
  verbosity INT := coalesce(current_setting('provsql.verbose_level', true)::INT, 0);
BEGIN
  -- Consult the per-table characterisation registry (TID / BID / OPAQUE,
  -- maintained by add_provenance / repair_key and the CTAS lineage hook):
  -- a TID relation is certified all-independent-inputs, a BID relation
  -- holds input or mulinput rows with the block structure given by the
  -- registry's key columns.  Derived (OPAQUE), unregistered, or
  -- subquery-defined edges take the fully dynamic per-token path.
  IF rel IS NOT NULL AND rel_sql IS NULL THEN
    tkind := (provsql.get_table_info(rel::oid)).kind;
  END IF;
  IF tkind NOT IN ('tid', 'bid') THEN
    tkind := NULL;
  END IF;
  IF tkind = 'bid' THEN
    SELECT string_agg(quote_ident(a.attname) || '::TEXT', ' || '','' || '
                      ORDER BY k.ord)
      INTO bkey_expr
      FROM unnest((provsql.get_table_info(rel::oid)).block_key)
             WITH ORDINALITY AS k(attnum, ord)
      JOIN pg_attribute a ON a.attrelid = rel AND a.attnum = k.attnum;
    -- An empty registry key means the whole table is one block.
    bkey_expr := coalesce(bkey_expr, quote_literal(''));
  END IF;
  IF tkind IS NOT NULL AND verbosity >= 20 THEN
    -- The function-level client_min_messages = warning (which silences
    -- the CTAS / DROP TABLE chatter) would also swallow this notice;
    -- lift it for the one RAISE.  The function-level SET restores the
    -- caller's value at exit regardless.
    PERFORM set_config('client_min_messages', 'notice', true);
    RAISE NOTICE 'ProvSQL: catalog characterises % as %', rel, upper(tkind);
    PERFORM set_config('client_min_messages', 'warning', true);
  END IF;
 
  -- Materialize the edges with their tokens; the planner hook resolves
  -- provenance() over the tracked relation, and remove_provenance strips
  -- the automatic provsql column so the later aggregation is plain SQL.
  -- For a BID relation the synthetic per-block key (a v5 UUID over the
  -- registry key columns' values) is computed here, while the columns
  -- are in scope.
  DROP TABLE IF EXISTS provsql_reachability_edges_tmp;
  EXECUTE format(
    'CREATE TEMP TABLE provsql_reachability_edges_tmp AS '
    || 'SELECT %1$I::TEXT AS u, %2$I::TEXT AS v, '
    || 'provsql.strip_annotations(provsql.provenance()) AS token%5$s '
    || 'FROM %3$s WHERE %1$I IS NOT NULL AND %2$I IS NOT NULL%4$s',
    source_attribute, destination_attribute,
    CASE WHEN rel_sql IS NULL THEN rel::TEXT
         ELSE '(' || rel_sql || ') AS provsql_edge_subquery' END,
    CASE WHEN edge_quals IS NULL THEN ''
         ELSE ' AND (' || edge_quals || ')' END,
    CASE WHEN tkind = 'bid'
         THEN ', public.uuid_generate_v5(provsql.uuid_ns_provsql(), '
              || quote_literal('bidblock' || rel::TEXT || ':')
              || ' || ' || bkey_expr || ') AS bkey'
         ELSE ', NULL::UUID AS bkey' END);
  PERFORM provsql.remove_provenance('provsql_reachability_edges_tmp');
 
  DROP TABLE IF EXISTS provsql_reachability_support_tmp;
  IF tkind IS NULL THEN
    -- Dynamic path: validate the token shapes and, for conjunction-shaped
    -- (join-defined) tokens, the pairwise disjointness of their supports.
    IF EXISTS (SELECT 1 FROM provsql_reachability_edges_tmp
               WHERE provsql.get_gate_type(token) NOT IN ('input', 'mulinput', 'times',
                                                  'project', 'eq')) THEN
      DROP TABLE provsql_reachability_edges_tmp;
      RAISE EXCEPTION 'reachability: the provenance of % must consist of base input, repair_key, or conjunctive join tokens', coalesce(rel::TEXT, 'the edge query');
    END IF;
    CREATE TEMP TABLE provsql_reachability_support_tmp AS
      SELECT t.token, l.leaf
      FROM (SELECT DISTINCT token FROM provsql_reachability_edges_tmp
            WHERE provsql.get_gate_type(token) IN ('times', 'project', 'eq')) t,
           LATERAL unnest(provsql.token_conjunctive_leaves(t.token)) AS l(leaf);
    IF EXISTS (SELECT 1
               FROM (SELECT DISTINCT token FROM provsql_reachability_edges_tmp) t
               WHERE provsql.get_gate_type(t.token) IN ('times', 'project', 'eq')
                 AND provsql.token_conjunctive_leaves(t.token) IS NULL) THEN
      DROP TABLE provsql_reachability_support_tmp;
      DROP TABLE provsql_reachability_edges_tmp;
      RAISE EXCEPTION 'reachability: a join-defined edge token is not a pure conjunction of base tuples';
    END IF;
    IF EXISTS (SELECT 1 FROM (
                 SELECT leaf FROM provsql_reachability_support_tmp
                 UNION ALL
                 SELECT DISTINCT token FROM provsql_reachability_edges_tmp
                 WHERE provsql.get_gate_type(token) = 'input'
               ) all_leaves
               GROUP BY leaf HAVING count(*) > 1) THEN
      DROP TABLE provsql_reachability_support_tmp;
      DROP TABLE provsql_reachability_edges_tmp;
      RAISE EXCEPTION 'reachability: join-defined edges share base tuples (their supports overlap), so they are not independent';
    END IF;
  END IF;
 
  -- Per-kind classification expressions for the final aggregation: a TID
  -- relation needs no per-row gate introspection at all; a BID relation
  -- one get_gate_type per row (the input/mulinput split), block keys from
  -- the precomputed column-derived key and indices by numbering within
  -- the block; the dynamic path reads the gates.
  IF tkind = 'tid' THEN
    sel_probs := 'coalesce(provsql.get_prob(e.token), 1.0)';
    sel_bkeys := $sql$'00000000-0000-0000-0000-000000000000'::UUID$sql$;
    sel_bidx  := '0';
  ELSIF tkind = 'bid' THEN
    sel_probs := 'coalesce(provsql.get_prob(e.token), 1.0)';
    sel_bkeys := $sql$CASE WHEN provsql.get_gate_type(e.token) = 'mulinput'
                      THEN e.bkey
                      ELSE '00000000-0000-0000-0000-000000000000'::UUID END$sql$;
    sel_bidx  := 'e.bidx';
  ELSE
    sel_probs := $sql$CASE WHEN provsql.get_gate_type(e.token) IN ('times','project','eq')
                      THEN (SELECT CASE WHEN bool_or(coalesce(provsql.get_prob(s.leaf),1.0) = 0)
                                        THEN 0.0
                                        ELSE exp(sum(ln(coalesce(provsql.get_prob(s.leaf),1.0)))) END
                            FROM provsql_reachability_support_tmp s
                            WHERE s.token = e.token)
                      ELSE coalesce(provsql.get_prob(e.token), 1.0) END$sql$;
    sel_bkeys := $sql$CASE WHEN provsql.get_gate_type(e.token) = 'mulinput'
                      THEN (provsql.get_children(e.token))[1]
                      ELSE '00000000-0000-0000-0000-000000000000'::UUID END$sql$;
    sel_bidx  := $sql$CASE WHEN provsql.get_gate_type(e.token) = 'mulinput'
                      THEN (provsql.get_infos(e.token)).info1 ELSE 0 END$sql$;
  END IF;
 
  EXECUTE format(
    $sql$
    WITH verts AS (
      SELECT u AS x FROM provsql_reachability_edges_tmp
      UNION SELECT v FROM provsql_reachability_edges_tmp
      UNION SELECT unnest($1)),
    ids AS (
      SELECT x, (row_number() OVER (ORDER BY x))::INT AS id FROM verts)
    SELECT array_agg(iu.id), array_agg(iv.id),
           array_agg(e.token),
           array_agg(%s),
           array_agg(%s),
           array_agg(%s),
           (SELECT array_agg(i.id ORDER BY ev.ord)
              FROM unnest($1) WITH ORDINALITY AS ev(x, ord)
              JOIN ids i ON i.x = ev.x),
           (SELECT array_agg(x ORDER BY id) FROM ids)
      FROM (SELECT t.*,
                   (row_number() OVER (PARTITION BY t.bkey))::INT AS bidx
            FROM provsql_reachability_edges_tmp t) e
      JOIN ids iu ON iu.x = e.u
      JOIN ids iv ON iv.x = e.v
    $sql$, sel_probs, sel_bkeys, sel_bidx)
    INTO sources, destinations, tokens, probabilities, block_keys,
         block_indices, extra_ids, vertices
    USING extra_vertices;
 
  DROP TABLE provsql_reachability_edges_tmp;
  DROP TABLE IF EXISTS provsql_reachability_support_tmp;
END
-- No SET search_path: the deparsed edge subquery (and the REGCLASS
-- rendering) must resolve against the caller's search_path; the ProvSQL
-- calls above are schema-qualified instead.
$$ LANGUAGE plpgsql SET client_min_messages = warning;
 
 
/**
 * @brief Gather a source relation's vertices, tokens and probabilities
 * (internal)
 *
 * For a provenance-tracked source relation, every tuple must carry a
 * base @c input token (a *probabilistic source set*); for an untracked
 * relation the sources are certain and the tokens come back as the nil
 * UUID.  Vertex values are returned as TEXT, for the shared dense-ID
 * mapping of @c gather_reachability_edges().
 *
 * @param rel the source relation
 * @param source_attribute name of the vertex column
 * @param[out] source_values vertex value of each source tuple (as TEXT)
 * @param[out] source_tokens per-source base @c input token (nil UUID = certain)
 * @param[out] source_probabilities per-source probability
 */
CREATE OR REPLACE FUNCTION gather_reachability_sources(
  IN rel REGCLASS,
  IN source_attribute TEXT,
  OUT source_values TEXT[],
  OUT source_tokens UUID[],
  OUT source_probabilities DOUBLE PRECISION[])
AS
$$
DECLARE
  tracked BOOLEAN;
  tkind TEXT;
BEGIN
  SELECT EXISTS (
    SELECT 1 FROM pg_attribute
    WHERE attrelid = rel AND attname = 'provsql'
      AND atttypid = 'UUID'::REGTYPE AND NOT attisdropped)
  INTO tracked;
 
  -- Registry consultation: a TID source relation is certified
  -- all-base-input, so the per-row gate check can be skipped; a BID one
  -- holds block-correlated tuples, which a probabilistic source set
  -- cannot model -- reject it before gathering anything.
  IF tracked THEN
    tkind := (get_table_info(rel::oid)).kind;
    IF tkind = 'bid' THEN
      RAISE EXCEPTION 'reachability: % is block-independent (repair_key); block-correlated source sets are not supported', rel;
    END IF;
  END IF;
 
  DROP TABLE IF EXISTS provsql_reachability_sources_tmp;
  IF tracked THEN
    EXECUTE format(
      'CREATE TEMP TABLE provsql_reachability_sources_tmp AS '
      || 'SELECT %1$I::TEXT AS x, provenance() AS token '
      || 'FROM %2$s WHERE %1$I IS NOT NULL',
      source_attribute, rel);
    PERFORM remove_provenance('provsql_reachability_sources_tmp');
    IF tkind IS DISTINCT FROM 'tid'
       AND EXISTS (SELECT 1 FROM provsql_reachability_sources_tmp
                   WHERE get_gate_type(token) <> 'input') THEN
      DROP TABLE provsql_reachability_sources_tmp;
      RAISE EXCEPTION 'reachability: the provenance of % must consist of base input tokens (independent tuples); views or query results are not supported', rel;
    END IF;
    SELECT array_agg(x), array_agg(token),
           array_agg(coalesce(get_prob(token), 1.0))
      INTO source_values, source_tokens, source_probabilities
      FROM provsql_reachability_sources_tmp;
    DROP TABLE provsql_reachability_sources_tmp;
  ELSE
    EXECUTE format(
      'CREATE TEMP TABLE provsql_reachability_sources_tmp AS '
      || 'SELECT DISTINCT %1$I::TEXT AS x FROM %2$s WHERE %1$I IS NOT NULL',
      source_attribute, rel);
    SELECT array_agg(x),
           array_agg('00000000-0000-0000-0000-000000000000'::UUID),
           array_agg(1.0::float8)
      INTO source_values, source_tokens, source_probabilities
      FROM provsql_reachability_sources_tmp;
    DROP TABLE provsql_reachability_sources_tmp;
  END IF;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp,public SET client_min_messages = warning;
 
/**
 * @brief Fixpoint driver for the recursive reachability shape:
 * decomposition-aligned compilation with fallback to eval_recursive
 *
 * Called (at plan time, over SPI) by the recursive-CTE lowering when
 * the provenance class is 'absorptive' or 'BOOLEAN'
 * (@c provsql.provenance) and the CTE matches the linear
 * reachability shape over a tracked base edge relation.  Attempts the
 * decomposition-aligned route -- gather the edges, compile every
 * reachable vertex's certified provenance circuit along a tree
 * decomposition of the data graph, materialise them, and fill the
 * working table with one tokenised row per reachable vertex.  On any
 * failure (data treewidth above the cap, per-node state bound, edges
 * that are not independent base tuples...), falls back to the generic
 * @c eval_recursive() fixpoint, preserving its behaviour exactly.
 *
 * @param edge_rel the provenance-tracked edge relation
 * @param source_attribute name of the source-vertex column
 * @param destination_attribute name of the destination-vertex column
 * @param source_value the base arm's constant, as TEXT
 * @param directed if false, each edge can be traversed both ways
 * @param work_name name of the working temp table (the CTE name)
 * @param colnames comma-separated user column names (for the fallback)
 * @param coldef column definitions of the working table
 * @param coltype type of the CTE's single column
 * @param body_sql deparsed CTE body (for the fallback)
 * @param edge_quals optional deterministic filter over edge columns
 *        (deparsed from the recursive arm's WHERE clause)
 * @param source_rel source relation of a multi-source base arm
 *        (@c SELECT col FROM sources), NULL for the constant form;
 *        tracked sources form a probabilistic source set, untracked
 *        ones are certain
 * @param source_rel_attribute the source relation's vertex column
 * @param edge_sql deparsed edge subquery when the recursive arm joins a
 *        derived (join-defined) edge relation instead of a base one;
 *        NULL for the REGCLASS form
 * @param hop_bound maximum number of recursive steps for the
 *        hop-counting CTE shape (NULL for plain reachability)
 * @param hop_seed the base arm's hop constant (hop-counting shape)
 * @param hops_position 1-based position of the hop column among the
 *        CTE's two columns (hop-counting shape)
 */
CREATE OR REPLACE FUNCTION eval_reachability(
  edge_rel REGCLASS,
  source_attribute TEXT,
  destination_attribute TEXT,
  source_value TEXT,
  directed BOOLEAN,
  work_name TEXT,
  colnames TEXT,
  coldef TEXT,
  coltype TEXT,
  body_sql TEXT,
  edge_quals TEXT DEFAULT NULL,
  source_rel REGCLASS DEFAULT NULL,
  source_rel_attribute TEXT DEFAULT NULL,
  edge_sql TEXT DEFAULT NULL,
  hop_bound INT DEFAULT NULL,
  hop_seed INT DEFAULT NULL,
  hops_position INT DEFAULT NULL)
  RETURNS VOID AS
$$
DECLARE
  e RECORD;
  sv TEXT[];
  st UUID[];
  sp double precision[];
  verbosity INT := coalesce(current_setting('provsql.verbose_level', true)::INT, 0);
BEGIN
  BEGIN
    IF source_rel IS NOT NULL THEN
      -- Multi-source: gather the source relation (probabilistic when
      -- tracked, certain otherwise).
      SELECT g.source_values, g.source_tokens, g.source_probabilities
        INTO sv, st, sp
        FROM provsql.gather_reachability_sources(source_rel,
                                                 source_rel_attribute) g;
      IF sv IS NULL THEN
        sv := ARRAY[]::TEXT[];
        st := ARRAY[]::UUID[];
        sp := ARRAY[]::float8[];
      END IF;
    ELSE
      -- Constant base arm: one certain source.
      sv := ARRAY[source_value];
      st := ARRAY['00000000-0000-0000-0000-000000000000'::UUID];
      sp := ARRAY[1.0::float8];
    END IF;
 
    e := provsql.gather_reachability_edges(edge_rel, source_attribute,
                                           destination_attribute,
                                           sv, edge_quals, edge_sql);
    IF to_regclass(work_name) IS NOT NULL THEN
      EXECUTE format('DROP TABLE %I', work_name);
    END IF;
    EXECUTE format('CREATE TEMP TABLE %I (%s, provsql UUID)', work_name, coldef);
    IF hop_bound IS NULL THEN
      EXECUTE format(
        'INSERT INTO %I SELECT ($1::TEXT[])[m.vertex]::%s, m.token '
        || 'FROM provsql.reachability_materialize($2, $3, $4, $5, $6, $7, $8, $9, $10, $11) m',
        work_name, coltype)
        USING e.vertices, e.sources, e.destinations, e.tokens, e.probabilities,
              e.block_keys, e.block_indices, e.extra_ids, st, sp, directed;
    ELSE
      -- Hop-counting shape: one row per (vertex, walk length), the hop
      -- column in its CTE position.
      EXECUTE format(
        'INSERT INTO %I SELECT %s, m.token '
        || 'FROM provsql.reachability_materialize_hops($2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13) m',
        work_name,
        CASE WHEN hops_position = 1
             THEN format('m.hops, ($1::TEXT[])[m.vertex]::%s', coltype)
             ELSE format('($1::TEXT[])[m.vertex]::%s, m.hops', coltype) END)
        USING e.vertices, e.sources, e.destinations, e.tokens, e.probabilities,
              e.block_keys, e.block_indices, e.extra_ids, st, sp, directed,
              hop_bound, hop_seed;
    END IF;
    IF verbosity >= 20 THEN
      RAISE NOTICE 'ProvSQL: recursive CTE "%" compiled along a tree decomposition of %',
        work_name, coalesce(edge_rel::TEXT, 'the join-defined edge query');
    END IF;
  EXCEPTION WHEN OTHERS THEN
    IF verbosity >= 10 THEN
      RAISE NOTICE 'ProvSQL: reachability route for "%" fell back to the generic fixpoint (%)',
        work_name, SQLERRM;
    END IF;
    PERFORM provsql.eval_recursive(body_sql, work_name, colnames, coldef);
  END;
END
$$ LANGUAGE plpgsql;
 
 
 
/** @} */
 
/** @defgroup provenance_output Provenance output
 *  Functions for visualizing and exporting provenance circuits
 *  in various formats.
 *  @{
 */
 
/**
 * @brief Return a DOT or TEXT visualization of the provenance circuit
 *
 * @param token root provenance token
 * @param token2desc mapping table for gate descriptions
 * @param dbg debug level (0 = normal)
 */
CREATE OR REPLACE FUNCTION view_circuit(
  token UUID,
  token2desc REGCLASS,
  dbg INT = 0)
  RETURNS TEXT AS
  'provsql','view_circuit' LANGUAGE C;
 
/**
 * @brief Return a DOT visualisation of the d-DNNF compiled from the
 * provenance circuit
 *
 * Runs the requested external knowledge compiler and renders the
 * resulting d-DNNF as a GraphViz digraph.
 *
 * @param token root provenance token
 * @param compiler external compiler or in-process meta-route to invoke;
 *   empty (the default) picks the highest-preference available compiler
 */
CREATE OR REPLACE FUNCTION compile_to_ddnnf_dot(
  token UUID,
  compiler TEXT = '')
  RETURNS TEXT AS
  'provsql','compile_to_ddnnf_dot' LANGUAGE C;
 
/**
 * @brief Return the compiled d-DNNF of a provenance circuit in the
 * c2d / d4 ".nnf" TEXT interchange format.
 *
 * Companion to compile_to_ddnnf_dot (DOT, for viewing): this is the
 * machine-readable form, suitable for feeding to an external d-DNNF
 * reasoner / verifier or saving next to tseytin_cnf (same variable
 * numbering). Accepts the same compiler / meta-route names.
 *
 * @param token root provenance token
 * @param compiler compiler or in-process meta-route to use; empty (the
 *   default) picks the highest-preference available compiler
 */
CREATE OR REPLACE FUNCTION compile_to_ddnnf(
  token UUID,
  compiler TEXT = '')
  RETURNS TEXT AS
  'provsql','compile_to_ddnnf' LANGUAGE C;
 
/**
 * @brief Structural statistics of the d-DNNF a compiler produces for a
 * provenance circuit.
 *
 * Compiles the circuit with the given compiler / meta-route (same names
 * as compile_to_ddnnf_dot: d4, d4v2, c2d, minic2d, dsharp, panini-*,
 * tree-decomposition, interpret-as-dd, default) and returns a jsonb
 * object: nodes, edges, and / or / not / inputs counts, smooth, depth
 * (longest path), treewidth (null when not computable), and compile_ms.
 * Lets clients compare what each compiler produces on the same circuit.
 *
 * @param token root provenance token
 * @param compiler compiler or in-process meta-route to use; empty (the
 *   default) picks the highest-preference available compiler
 */
CREATE OR REPLACE FUNCTION ddnnf_stats(
  token UUID,
  compiler TEXT = '')
  RETURNS jsonb AS
  'provsql','ddnnf_stats' LANGUAGE C;
 
/**
 * @brief Return the DIMACS CNF (Tseytin transformation) of the provenance circuit
 *
 * Returns the same encoding the extension writes to a temp file before
 * invoking d4 / c2d / minic2d / dsharp. With @c weighted true (the
 * default), per-input probability weights are appended as @c w lines.
 *
 * @param token root provenance token
 * @param weighted include probability weights when true
 * @param mapping prepend "c input <var> <UUID> <prob>" comment lines
 *        documenting which provenance input each variable stands for
 */
CREATE OR REPLACE FUNCTION tseytin_cnf(
  token UUID,
  weighted BOOLEAN = TRUE,
  mapping BOOLEAN = TRUE)
  RETURNS TEXT AS
  'provsql','tseytin_cnf' LANGUAGE C;
 
/**
 * @brief Map each DIMACS variable of tseytin_cnf back to its
 *        provenance input.
 *
 * Returns one row per input gate: the variable index (matching
 * tseytin_cnf and compile_to_ddnnf's NNF), the original-circuit UUID
 * of that input, and its probability. Lets a satisfying assignment or
 * weighted model count obtained from an external tool be read against
 * the provenance circuit.
 *
 * @param token root provenance token
 */
CREATE OR REPLACE FUNCTION tseytin_cnf_mapping_json(token UUID)
  RETURNS jsonb AS
  'provsql','tseytin_cnf_mapping_json' LANGUAGE C;
 
CREATE OR REPLACE FUNCTION tseytin_cnf_mapping(token UUID)
  RETURNS TABLE(variable INT, gate UUID, probability FLOAT8) AS $$
  SELECT variable, gate, probability
  FROM jsonb_to_recordset(tseytin_cnf_mapping_json(token))
       AS x(variable INT, gate UUID, probability FLOAT8)
  ORDER BY variable
$$ LANGUAGE SQL STABLE;
 
/**
 * @brief Return a DOT visualisation of the tree decomposition of the
 * provenance circuit
 *
 * Computes the min-fill decomposition used by the in-process
 * knowledge compiler. The first line of the output is a comment of
 * the form @c "// treewidth=<n>".
 *
 * @param token root provenance token
 */
CREATE OR REPLACE FUNCTION tree_decomposition_dot(
  token UUID)
  RETURNS TEXT AS
  'provsql','tree_decomposition_dot' LANGUAGE C;
 
/**
 * @brief Report whether an external tool is on the backend's resolved PATH
 *
 * Uses the same @c find_external_tool() helper that the compilers
 * (d4 / c2d / minic2d / dsharp / panini), model counters (ganak /
 * sharpsat-td / dpmc via htb+dmc / weightmc), and visualisation
 * wrappers (graph-easy, dot) themselves consult, so the result
 * reflects exactly what a subsequent @c probability_evaluate or
 * @c view_circuit call would see, including the
 * @c provsql.tool_search_path GUC prepended to @c $PATH.
 *
 * Names with a slash are treated as paths and tested directly via
 * @c access(X_OK); bare names are resolved through @c /bin/sh's
 * @c command -v under the backend's PATH.
 *
 * @param name bare executable (e.g. @c 'd4') or an absolute path
 * @return true iff the tool resolves to an executable file
 */
CREATE OR REPLACE FUNCTION tool_available(name TEXT)
  RETURNS BOOLEAN AS
  'provsql','tool_available' LANGUAGE C STRICT;
 
/* ----------------------------------------------------------------------
 * External-tool registry
 *
 * A catalog of the external tools ProvSQL can invoke (the knowledge
 * compilers, weighted model counters, and the graph-easy DOT renderer).
 * The default tools and their invocations are compiled in (seeded in C), so
 * out-of-the-box behaviour is unchanged with no configuration.
 *
 * Administrators may add / repoint / reorder / disable tools at run time;
 * those changes are persisted in the @c provsql.tool_overrides table below
 * and overlaid on the compiled seed, so they survive across sessions and
 * backends (and dump/restore).  An empty overrides table means exactly the
 * compiled defaults.  The mutators are superuser-only because a tool RECORD
 * names an executable run as the PostgreSQL OS user (the same trust level as
 * provsql.tool_search_path).
 * ---------------------------------------------------------------------- */
 
/**
 * @brief Persistent overrides overlaid on the compiled-in tool seed.
 *
 * Each row is the complete desired RECORD for a tool (added or modified) keyed
 * by logical @c name, or a tombstone (@c removed = true) hiding a seeded
 * default.  The effective registry is the compiled seed with tombstoned names
 * removed and the remaining rows upserted over it.  Written only by the
 * superuser-only register_tool / unregister_tool / set_tool_* functions;
 * read back into each backend's in-memory registry on demand.  Marked as a
 * configuration table so pg_dump carries an operator's registrations.
 */
CREATE TABLE IF NOT EXISTS tool_overrides(
  name           TEXT PRIMARY KEY,
  removed        BOOLEAN NOT NULL DEFAULT false,
  kind           TEXT,
  executable     TEXT,
  operations     TEXT[],
  input_formats  TEXT[],
  output_format  TEXT,
  parser         TEXT,
  preference     INT,
  enabled        BOOLEAN,
  dependencies   TEXT[],
  argtpl         TEXT,
  argtpl_circuit TEXT,
  endpoint       TEXT
);
SELECT pg_catalog.pg_extension_config_dump('tool_overrides', '');
 
/**
 * @brief Set-returning listing backing the @c provsql.tools view.
 *
 * @c operations / @c input_formats / @c output_format use the KCMCP
 * shared-registry names (see the KCMCP server protocol), so a CLI RECORD and
 * a future kcmcp-server RECORD are comparable; @c parser is the CLI-only tag
 * for how to decode the tool's raw output.  @c argtpl is the command template
 * ({in}/{out}/... placeholders).  @c available is true iff @c executable
 * (when set) and every dependency currently resolve on the backend's PATH.
 */
CREATE OR REPLACE FUNCTION tool_registry_list()
  RETURNS TABLE(name TEXT, kind TEXT, executable TEXT, operations TEXT[],
                input_formats TEXT[], output_format TEXT, parser TEXT,
                preference INT, enabled BOOLEAN, argtpl TEXT,
                argtpl_circuit TEXT, endpoint TEXT, available BOOLEAN) AS
  'provsql','tool_registry_list' LANGUAGE C STABLE;
 
/**
 * @brief Read-only view of the registered tools.
 */
CREATE OR REPLACE VIEW tools AS
  SELECT name, kind, executable, operations, input_formats, output_format,
         parser, preference, enabled, argtpl, argtpl_circuit, endpoint,
         available
  FROM tool_registry_list();
 
/**
 * @brief Register a tool, or replace the RECORD with the same logical name.
 *
 * @param name        logical id (e.g. @c 'd4-jm62300'); also the value
 *                    @c provsql.fallback_compiler / the wmc tool selector use
 * @param executable  executable to resolve on PATH (defaults to @c name)
 * @param kind          @c 'cli' (spawn @c executable) or @c 'kcmcp' (talk to
 *                      the KCMCP server at @c endpoint)
 * @param operations    capabilities (KCMCP names): @c 'compile' / @c 'wmc'
 *                      (and ProvSQL-local @c 'render')
 * @param input_formats accepted inputs (KCMCP names): @c 'dimacs-cnf',
 *                      @c 'circuit-bcs12' (listing @c 'circuit-bcs12' enables
 *                      the native-circuit fast path)
 * @param output_format result encoding (KCMCP names): @c 'ddnnf-nnf',
 *                      @c 'decimal', @c 'rational', ... (local @c 'panini-dd'
 *                      / @c 'ascii' where KCMCP has no code)
 * @param parser        CLI-only decode tag: @c 'nnf' (the tolerant d4 / c2d
 *                      NNF reader), @c 'panini-dd', @c 'wmc-line',
 *                      @c 'weightmc', @c 'ascii'
 * @param argtpl        command template; placeholders @c {in} / @c {out}
 *                      (and @c {binary} / @c {tmpdir} / @c {pivotAC}).  When
 *                      it omits @c {binary}, the executable is prepended.
 * @param argtpl_circuit command used when the @c 'circuit-bcs12' input is
 *                      selected (a BC-S1.2 circuit rather than a CNF); only a
 *                      tool accepting that input needs it
 * @param preference    ordering within an operation (higher first)
 * @param enabled       whether the dispatchers may select it
 * @param endpoint      for a @c 'kcmcp' RECORD, the server address:
 *                      @c 'unix:/path' or @c 'host:port'
 *
 * Superuser-only: a CLI RECORD runs an arbitrary command as the PostgreSQL
 * OS user, and a kcmcp RECORD names a socket the server connects to.
 */
CREATE OR REPLACE FUNCTION register_tool(
  name TEXT,
  executable TEXT DEFAULT NULL,
  kind TEXT DEFAULT 'cli',
  operations TEXT[] DEFAULT NULL,
  input_formats TEXT[] DEFAULT NULL,
  output_format TEXT DEFAULT NULL,
  parser TEXT DEFAULT NULL,
  argtpl TEXT DEFAULT NULL,
  argtpl_circuit TEXT DEFAULT NULL,
  preference INT DEFAULT 0,
  enabled BOOLEAN DEFAULT true,
  endpoint TEXT DEFAULT NULL)
  RETURNS VOID AS
  'provsql','tool_registry_register' LANGUAGE C;
 
/** @brief Unregister a tool; errors on an unknown tool name. Superuser-only. */
CREATE OR REPLACE FUNCTION unregister_tool(name TEXT)
  RETURNS VOID AS
  'provsql','tool_registry_unregister' LANGUAGE C STRICT;
 
/** @brief Enable/disable a tool; errors on an unknown tool name. Superuser-only. */
CREATE OR REPLACE FUNCTION set_tool_enabled(name TEXT, enabled BOOLEAN)
  RETURNS VOID AS
  'provsql','tool_registry_set_enabled' LANGUAGE C STRICT;
 
/** @brief Set a tool's preference; errors on an unknown tool name. Superuser-only. */
CREATE OR REPLACE FUNCTION set_tool_preference(name TEXT, preference INT)
  RETURNS VOID AS
  'provsql','tool_registry_set_preference' LANGUAGE C STRICT;
 
-- The mutators guard at the C level too, but revoke from PUBLIC so the
-- superuser requirement is visible in the catalog.
REVOKE ALL ON FUNCTION register_tool(TEXT, TEXT, TEXT, TEXT[], TEXT[], TEXT, TEXT, TEXT, TEXT, INT, BOOLEAN, TEXT) FROM PUBLIC;
REVOKE ALL ON FUNCTION unregister_tool(TEXT) FROM PUBLIC;
REVOKE ALL ON FUNCTION set_tool_enabled(TEXT, BOOLEAN) FROM PUBLIC;
REVOKE ALL ON FUNCTION set_tool_preference(TEXT, INT) FROM PUBLIC;
 
/**
 * @brief Return an XML representation of the provenance circuit
 *
 * @param token root provenance token
 * @param token2desc optional mapping table for gate descriptions
 */
CREATE OR REPLACE FUNCTION to_provxml(
  token UUID,
  token2desc REGCLASS = NULL)
  RETURNS TEXT AS
  'provsql','to_provxml' LANGUAGE C;
 
/** @brief Return the provenance token of the current query result tuple */
CREATE OR REPLACE FUNCTION provenance() RETURNS UUID AS
 'provsql', 'provenance' LANGUAGE C;
 
/**
 * @brief Compute where-provenance for a result tuple
 *
 * Returns a TEXT representation showing which input columns
 * contributed to each output column.
 */
CREATE OR REPLACE FUNCTION where_provenance(token UUID)
  RETURNS TEXT AS
  'provsql','where_provenance' LANGUAGE C;
 
/** @} */
 
/** @defgroup circuit_init Circuit initialization
 *  Functions and statements executed at extension load time to
 *  reset internal caches and create the constant zero/one gates.
 *  @{
 */
 
/** @brief Reset the internal cache of OID constants used by the query rewriter */
CREATE OR REPLACE FUNCTION reset_constants_cache()
  RETURNS VOID AS
  'provsql', 'reset_constants_cache' LANGUAGE C;
 
SELECT reset_constants_cache();
 
SELECT create_gate(gate_zero(), 'zero');
SELECT create_gate(gate_one(), 'one');
 
/** @} */
 
/** @brief Types of update operations tracked for temporal provenance */
CREATE TYPE QUERY_TYPE_ENUM AS ENUM (
    'INSERT',  -- Row was inserted
    'DELETE',  -- Row was deleted
    'UPDATE',  -- Row was updated
    'UNDO'     -- Previous operation was undone
    );
 
/** @defgroup compiled_semirings Compiled semirings
 *  Definitions of compiled semirings
 *  @{
 */
 
/** @brief Evaluate provenance as a symbolic formula (e.g., "a ⊗ b ⊕ c") */
-- The mapping is optional (as for sr_boolexpr): formula renders whatever
-- circuit it is given, and the measure-carrier circuits it is most useful
-- on (random variables, arithmetic, mixtures) have no leaf mapping at all.
-- Without one, input leaves render as the semiring's 𝟙.
CREATE FUNCTION sr_formula(token ANYELEMENT, token2value REGCLASS = NULL)
  RETURNS VARCHAR AS
$$
BEGIN
  IF token IS NULL THEN
    RETURN NULL;
  END IF;
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'formula',
    '𝟙'::VARCHAR
  );
END
$$ LANGUAGE plpgsql PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the counting semiring (ℕ) */
CREATE FUNCTION sr_counting(token ANYELEMENT, token2value REGCLASS)
  RETURNS INT AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'counting',
    1
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance as why-provenance (set of witness sets) */
CREATE FUNCTION sr_why(token ANYELEMENT, token2value REGCLASS)
  RETURNS VARCHAR AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'why',
    '{}'::VARCHAR
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance as how-provenance (canonical polynomial provenance ℕ[X], universal commutative-semiring provenance) */
CREATE FUNCTION sr_how(token ANYELEMENT, token2value REGCLASS)
  RETURNS VARCHAR AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'how',
    '{}'::VARCHAR
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance as which-provenance (lineage: a single set of contributing labels) */
CREATE FUNCTION sr_which(token ANYELEMENT, token2value REGCLASS)
  RETURNS VARCHAR AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'which',
    '{}'::VARCHAR
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance as a Boolean expression
 *
 * The optional @p token2value mapping labels the leaves of the
 * formula: when omitted, leaves are rendered as bare @c x@<id@>
 * placeholders.
 */
CREATE FUNCTION sr_boolexpr(token ANYELEMENT, token2value REGCLASS = NULL)
  RETURNS VARCHAR AS
$$
BEGIN
  IF token IS NULL THEN
    RETURN NULL;
  END IF;
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'boolexpr',
    '⊤'::VARCHAR
  );
END
$$ LANGUAGE plpgsql PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the Boolean semiring (true/false) */
CREATE FUNCTION sr_boolean(token ANYELEMENT, token2value REGCLASS)
  RETURNS BOOLEAN AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'BOOLEAN',
    TRUE
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Structural universal-zero test (C backend of nonzero's default mode) */
CREATE FUNCTION true_nonzero(token UUID)
  RETURNS BOOLEAN AS
  'provsql', 'true_nonzero' LANGUAGE C PARALLEL SAFE STABLE;
 
/**
 * @brief Test whether a provenance annotation is nonzero.
 *
 * Returns false only on a *proof* that the annotation is zero; true
 * otherwise, so filtering with <tt>WHERE nonzero(provenance())</tt> never
 * discards a row whose annotation could be nonzero.
 *
 * The default mode (@p semiring NULL) tests *universal* zero-ness: zero in
 * every (m-)semiring under every leaf valuation, decided by sound
 * structural rules (zero propagation through the gates; a comparison gate
 * whose satisfying-world set is empty).  Filtering on it can never
 * contradict any downstream semiring evaluation.
 *
 * A named @p semiring evaluates the circuit there and tests against that
 * semiring's zero: 'BOOLEAN' is presence in the vanilla SQL answer on this
 * instance (the mode that filters, e.g., the null-padded arm of a
 * difference), 'counting' is bag multiplicity.  An absent @p mapping reads
 * every leaf as the semiring's one (true / 1); with a mapping, leaves take
 * their mapped values.
 *
 * A NULL @p token reads as the neutral 1 (an untracked row): true.
 *
 * @param token    provenance token to test
 * @param semiring NULL (universal zero test), 'BOOLEAN', or 'counting'
 * @param mapping  optional mapping table from tokens to leaf values
 */
CREATE FUNCTION nonzero(token UUID,
                        semiring TEXT DEFAULT NULL,
                        mapping REGCLASS DEFAULT NULL)
  RETURNS BOOLEAN AS
$$
BEGIN
  IF token IS NULL THEN
    RETURN true;
  END IF;
  IF semiring IS NULL THEN
    RETURN provsql.true_nonzero(token);
  ELSIF semiring = 'BOOLEAN' THEN
    RETURN provsql.provenance_evaluate_compiled(token, mapping, 'BOOLEAN', TRUE);
  ELSIF semiring = 'counting' THEN
    RETURN provsql.provenance_evaluate_compiled(token, mapping, 'counting', 1) <> 0;
  ELSE
    RAISE EXCEPTION 'nonzero: unsupported semiring "%" (supported: BOOLEAN, counting; NULL for the universal zero test)', semiring;
  END IF;
END
$$ LANGUAGE plpgsql PARALLEL SAFE STABLE;
 
/**
 * @brief Presence in the vanilla SQL answer on this instance.
 *
 * Shorthand for <tt>nonzero(token, 'BOOLEAN')</tt> with every leaf true:
 * <tt>WHERE present(provenance())</tt> restores the result set the query
 * has without provenance tracking, filtering the zero-annotated extras
 * (antijoin arms, failed HAVING groups, unknown comparisons) that the
 * rewriting keeps visible.
 */
CREATE FUNCTION present(token UUID)
  RETURNS BOOLEAN AS
$$
  SELECT provsql.nonzero(token, 'BOOLEAN');
$$ LANGUAGE sql PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the tropical (min-plus) m-semiring
 *
 * Inputs are read as %float8 cost values; the additive identity
 * is <tt>'Infinity'::%float8</tt> and the multiplicative identity is 0.
 * Returns the cost of the cheapest derivation.
 *
 * With @p nonnegative, input costs are checked nonnegative and the
 * semiring is *absorptive*: evaluation then also accepts circuits
 * carrying the @c 'absorptive' assumption marker -- notably cyclic
 * recursive queries truncated at the absorptive value fixpoint, giving
 * exact min-cost reachability on cyclic data.
 */
CREATE FUNCTION sr_tropical(token ANYELEMENT, token2value REGCLASS,
                            nonnegative BOOLEAN = false)
  RETURNS FLOAT AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    CASE WHEN nonnegative THEN 'tropical_nonneg' ELSE 'tropical' END,
    0::FLOAT
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the Viterbi (max-times) m-semiring
 *
 * Inputs are read as %float8 probability values in @f$[0,1]@f$.
 * Returns the probability of the most likely derivation.
 */
CREATE FUNCTION sr_viterbi(token ANYELEMENT, token2value REGCLASS)
  RETURNS FLOAT AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'viterbi',
    1::FLOAT
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the Łukasiewicz fuzzy m-semiring
 *
 * Inputs are read as %float8 graded-truth values in @f$[0,1]@f$.
 * Addition is @f$\max@f$; multiplication is the Łukasiewicz t-norm
 * @f$\max(a + b - 1, 0)@f$, which preserves crisp truth and avoids
 * the near-zero collapse of long product chains.
 */
CREATE FUNCTION sr_lukasiewicz(token ANYELEMENT, token2value REGCLASS)
  RETURNS FLOAT AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'lukasiewicz',
    1::FLOAT
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the min-max m-semiring on a user ENUM
 *
 * Inputs are read as values of a user-defined ENUM carrier; addition
 * is ENUM-min, multiplication is ENUM-max. Bottom and top of the ENUM
 * are derived from @c pg_enum.enumsortorder. The third argument is a
 * sample value of the carrier ENUM, used only for type inference; its
 * value is ignored.
 *
 * The security shape: alternative derivations combine to the least
 * sensitive label, joins combine to the most sensitive label.
 *
 * @param token Provenance token to evaluate.
 * @param token2value Mapping from input gates to ENUM values.
 * @param element_one Sample value of the carrier ENUM (any value works).
 */
CREATE FUNCTION sr_minmax(token UUID, token2value REGCLASS, element_one ANYENUM)
  RETURNS ANYENUM AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'minmax',
    element_one
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @brief Evaluate provenance over the max-min m-semiring on a user ENUM
 *
 * Dual of :sqlfunc:`sr_minmax`: addition is ENUM-max, multiplication
 * is ENUM-min. The fuzzy / availability / trust shape: alternatives
 * combine to the most permissive label, joins combine to the strictest
 * label. The third argument is a sample value of the carrier ENUM,
 * used only for type inference; its value is ignored.
 *
 * @param token Provenance token to evaluate.
 * @param token2value Mapping from input gates to ENUM values.
 * @param element_one Sample value of the carrier ENUM (any value works).
 */
CREATE FUNCTION sr_maxmin(token UUID, token2value REGCLASS, element_one ANYENUM)
  RETURNS ANYENUM AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'maxmin',
    element_one
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/** @} */
 
/** @defgroup choose_aggregate choose aggregate
 *  Choose one value among many, used in particular to code a mutually
 * exclusive choice as an aggregate.
 *  @{
 */
 
/** @brief Transition function for the choose aggregate (keeps first non-NULL value) */
CREATE FUNCTION choose_function(state ANYELEMENT, data ANYELEMENT)
  RETURNS ANYELEMENT AS
$$
BEGIN
  IF state IS NULL THEN
    RETURN data;
  ELSE
    RETURN state;
  END IF;
END
$$ LANGUAGE plpgsql PARALLEL SAFE IMMUTABLE;
 
/** @brief Aggregate that returns an arbitrary non-NULL value from a group */
CREATE AGGREGATE choose(ANYELEMENT) (
  SFUNC = choose_function,
  STYPE = ANYELEMENT
);
 
/** @brief Explodes a table column containing aggregated provenance into multiple rows.
 *
 *  For each row in the input table, this function unnests the children of the
 *  specified aggregate token column and produces one output row per child.
 *  It reconstructs the corresponding value and provenance (`provsql`) for
 *  each resulting row.
 *
 *  The original table is replaced by the transformed table.
 *
 *  @param _tbl Name of the table to transform.
 *  @param AGG_TOKEN Name of the column containing the aggregate to explode.
 */
CREATE OR REPLACE FUNCTION explode_table(_tbl TEXT, AGG_TOKEN TEXT)
RETURNS VOID AS $$
DECLARE
  _nsp TEXT;
BEGIN
    -- Resolve the schema actually holding _tbl so the rebuilt table is
    -- recreated in place (the provsql helper functions are schema-qualified
    -- so this works whatever the caller's search_path is).
    SELECT n.nspname INTO _nsp
    FROM pg_class c JOIN pg_namespace n ON n.oid = c.relnamespace
    WHERE c.oid = _tbl::REGCLASS;
 
    EXECUTE format('
    CREATE TABLE %1$I.temp_exploded AS
    SELECT
        %2$I.*,
        provsql.get_extra(children[2]) AS new_t,
        provsql.provenance_times(children[1], provsql) AS new_provsql
    FROM %1$I.%2$I,
    LATERAL (
        SELECT provsql.get_children(sm) AS children
        FROM UNNEST(provsql.get_children(%3$I)) AS sm
    ) AS sub', _nsp, _tbl, AGG_TOKEN);
    EXECUTE format('DROP TABLE %I.%I', _nsp, _tbl);
    EXECUTE format('ALTER TABLE %I.temp_exploded DROP COLUMN %I, DROP COLUMN provsql', _nsp, AGG_TOKEN);
    EXECUTE format('ALTER TABLE %I.temp_exploded RENAME COLUMN new_t TO %I', _nsp, AGG_TOKEN);
    EXECUTE format('ALTER TABLE %I.temp_exploded RENAME COLUMN new_provsql TO provsql', _nsp);
    EXECUTE format('ALTER TABLE %I.temp_exploded RENAME TO %I', _nsp, _tbl);
END;
$$ LANGUAGE plpgsql;
 
/** @} */
 
/**
 * @brief Append @c provsql to this database's default search_path, if missing.
 *
 * ProvSQL's operators and functions live in the @c provsql schema and
 * are resolved through @c search_path.  When @c provsql is absent from
 * the path some surfaces fail with a clear error (RV/AGG_TOKEN
 * arithmetic), but others can be silently misrouted by an implicit
 * cross-domain cast.  This helper makes the common case painless: it
 * reads the current <em>database-level</em> search_path setting from
 * @c pg_db_role_setting, appends @c provsql if not already present
 * (never replacing or reordering the existing entries), and applies the
 * result with @c ALTER @c DATABASE.  It is idempotent and emits a
 * @c NOTICE describing what it did.
 *
 * Only @b new sessions pick up the change; the calling session keeps its
 * current path.  Role-level settings (if any) take precedence over the
 * database-level setting and are left untouched.  The caller must be the
 * database owner or a superuser (the privilege model of @c ALTER
 * @c DATABASE).  Returns the resulting search_path value.
 */
CREATE OR REPLACE FUNCTION setup_search_path()
  RETURNS TEXT
  LANGUAGE plpgsql AS $$
DECLARE
  db        TEXT := current_database();
  cfg       TEXT[];
  cur       TEXT;        -- existing database-level search_path value
  new_path  TEXT;
BEGIN
  -- setrole = 0 selects the database-wide default, not a per-role override.
  SELECT s.setconfig INTO cfg
    FROM pg_db_role_setting s
    JOIN pg_database d ON d.oid = s.setdatabase
   WHERE d.datname = db AND s.setrole = 0;
 
  IF cfg IS NOT NULL THEN
    SELECT substr(e, length('search_path=') + 1) INTO cur
      FROM unnest(cfg) AS e
     WHERE e LIKE 'search_path=%';
  END IF;
 
  IF cur IS NULL THEN
    -- No database-level search_path at all: install the documented
    -- default with provsql appended.
    new_path := '"$user", public, provsql';
    EXECUTE format('ALTER DATABASE %I SET search_path = %s', db, new_path);
    RAISE NOTICE 'ProvSQL: set search_path = % for database "%" (no previous database-level setting). Only new sessions are affected.',
      new_path, db;
    RETURN new_path;
  END IF;
 
  -- Already contains provsql as a path element?  Idempotent no-op.
  IF EXISTS (
       SELECT 1 FROM unnest(string_to[](cur, ',')) AS p
        WHERE btrim(btrim(p), '"') = 'provsql')
  THEN
    RAISE NOTICE 'ProvSQL: search_path for database "%" already contains provsql (= %); no change.',
      db, cur;
    RETURN cur;
  END IF;
 
  new_path := cur || ', provsql';
  EXECUTE format('ALTER DATABASE %I SET search_path = %s', db, new_path);
  RAISE NOTICE 'ProvSQL: appended provsql to search_path for database "%" (now: %). Only new sessions are affected.',
    db, new_path;
  RETURN new_path;
END;
$$;
 
GRANT USAGE ON SCHEMA provsql TO PUBLIC;
 
SET search_path TO public;
 
-- Installation-time advisory: if provsql is not in the database's default
-- search_path, point the user at setup_search_path().  reset_val reflects
-- the configured session default (postgresql.conf / ALTER DATABASE / ALTER
-- ROLE), unaffected by the SET search_path statements this script ran.
-- CREATE EXTENSION raises client_min_messages to WARNING for the duration
-- of the script, so we lower it around the RAISE NOTICE. SET LOCAL only:
-- it unwinds by itself when CREATE EXTENSION's transaction ends. An
-- explicit save/restore here would capture the WARNING clamp (already in
-- force when this block runs) and restore *that* at session level,
-- leaving the whole installing session with NOTICEs suppressed.
DO $$
DECLARE
  rp TEXT;
  has_provsql BOOLEAN;
BEGIN
  SELECT reset_val INTO rp FROM pg_settings WHERE name = 'search_path';
  SELECT bool_or(btrim(btrim(p), '"') = 'provsql')
    INTO has_provsql
    FROM unnest(string_to[](coalesce(rp, ''), ',')) AS p;
  IF NOT coalesce(has_provsql, false) THEN
    SET LOCAL client_min_messages = notice;
    RAISE NOTICE 'ProvSQL: schema "provsql" is not in your default search_path (currently: %).', rp;
    RAISE NOTICE 'ProvSQL operators and functions are resolved through search_path. Run "SELECT provsql.setup_search_path();" to add it, or set it manually (e.g. ALTER DATABASE % SET search_path = "$user", public, provsql).', quote_ident(current_database());
  END IF;
END;
$$;
 
-- Final constants-cache refresh.  The planned SELECT statements earlier in
-- this script (reset_constants_cache itself, the zero/one create_gate calls)
-- make the installing session memoize the OID constants *mid-script*, while
-- objects defined later (notably the choose aggregate, used by the
-- scalar-subquery decorrelation) do not exist yet.  Their optional lookups
-- then stay InvalidOid for the rest of the session, silently disabling the
-- corresponding rewrites (e.g. IN/NOT IN over a tracked relation would raise
-- "Subqueries ... not supported") until a new connection.  Refreshing here,
-- after every object exists, repairs the installing session's cache.
SELECT provsql.reset_constants_cache();
SET search_path TO provsql;
 
/** @defgroup update_provenance Update provenance (PostgreSQL 14+)
 *  Extended provenance tracking for INSERT, UPDATE, DELETE, and UNDO
 *  operations, including temporal validity ranges.
 *  @{
 */
 
/**
 * @brief Table recording the history of INSERT, UPDATE, DELETE, and UNDO operations
 *
 * Each row records one provenance-tracked modification, linking the
 * operation's provenance token to metadata (query TEXT, type, user,
 * TIMESTAMP) and the temporal validity range of the affected rows.
 */
CREATE TABLE update_provenance (
  provsql UUID,
  query TEXT,
  query_type QUERY_TYPE_ENUM,
  username TEXT,
  ts TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
  valid_time TSTZMULTIRANGE DEFAULT TSTZMULTIRANGE(tstzrange(CURRENT_TIMESTAMP, NULL))
);
 
/** @cond INTERNAL */
/* Enable provenance tracking on an existing table (PostgreSQL 14+ version).
 * Overrides the common version; documented via add_provenance in provsql.common.sql. */
CREATE OR REPLACE FUNCTION add_provenance(_tbl REGCLASS)
  RETURNS VOID AS
$$
BEGIN
  -- Idempotence: a second add_provenance on an already-tracked table is
  -- a no-op with a NOTICE, so setup scripts and notebook cells can be
  -- re-run freely.
  IF EXISTS (
    SELECT 1 FROM pg_attribute
    WHERE attrelid = _tbl AND attname = 'provsql' AND NOT attisdropped
  ) THEN
    RAISE NOTICE 'table % already has provenance tracking', _tbl;
    RETURN;
  END IF;
  -- See the common-version body for the rationale of dropping the
  -- column DEFAULT and UNIQUE in favour of provenance_guard + a
  -- plain index.
  EXECUTE format('ALTER TABLE %s ADD COLUMN provsql UUID', _tbl);
  EXECUTE format(
    'UPDATE %s SET provsql = public.uuid_generate_v4() WHERE provsql IS NULL',
    _tbl);
  EXECUTE format('CREATE INDEX ON %s(provsql)', _tbl);
  EXECUTE format(
    'CREATE TRIGGER provenance_guard BEFORE INSERT OR UPDATE OF provsql '
    'ON %s FOR EACH ROW EXECUTE PROCEDURE provsql.provenance_guard()',
    _tbl);
 
  EXECUTE format('CREATE TRIGGER insert_statement AFTER INSERT ON %s REFERENCING NEW TABLE AS NEW_TABLE FOR EACH STATEMENT EXECUTE PROCEDURE provsql.insert_statement_trigger()', _tbl);
  EXECUTE format('CREATE TRIGGER delete_statement AFTER DELETE ON %s REFERENCING OLD TABLE AS OLD_TABLE FOR EACH STATEMENT EXECUTE PROCEDURE provsql.delete_statement_trigger()', _tbl);
  EXECUTE format('CREATE TRIGGER update_statement AFTER UPDATE ON %s REFERENCING OLD TABLE AS OLD_TABLE NEW TABLE AS NEW_TABLE FOR EACH STATEMENT EXECUTE PROCEDURE provsql.update_statement_trigger()', _tbl);
 
  PERFORM provsql.set_table_info(_tbl::oid, 'tid');
  PERFORM provsql.set_ancestors(_tbl::oid, ARRAY[_tbl::oid]);
END
$$ LANGUAGE plpgsql SECURITY DEFINER;
/** @endcond */
 
/** @cond INTERNAL */
/* Trigger function for DELETE statement provenance tracking (PostgreSQL 14+).
 * Overrides the common version; documented via delete_statement_trigger in provsql.common.sql. */
CREATE OR REPLACE FUNCTION delete_statement_trigger()
  RETURNS TRIGGER AS
$$
DECLARE
  query_text TEXT;
  delete_token UUID;
  old_token UUID;
  new_token UUID;
  r RECORD;
  enable_trigger BOOL;
BEGIN
  enable_trigger := current_setting('provsql.update_provenance', true);
  IF enable_trigger = 'f' THEN
    RETURN NULL;
  END IF;
  delete_token := public.uuid_generate_v4();
 
  PERFORM create_gate(delete_token, 'update');
 
  SELECT query
  INTO query_text
  FROM pg_stat_activity
  WHERE pid = pg_backend_pid();
 
  INSERT INTO update_provenance (provsql, query, query_type, username, ts, valid_time)
  VALUES (delete_token, query_text, 'DELETE', current_user, CURRENT_TIMESTAMP, TSTZMULTIRANGE(tstzrange(CURRENT_TIMESTAMP, NULL)));
 
  PERFORM set_config('provsql.update_provenance', 'off', false);
  EXECUTE format('INSERT INTO %I.%I SELECT * FROM OLD_TABLE;', TG_TABLE_SCHEMA, TG_TABLE_NAME);
  PERFORM set_config('provsql.update_provenance', 'on', false);
 
  FOR r IN (SELECT * FROM OLD_TABLE) LOOP
    old_token := r.provsql;
    new_token := provenance_monus(old_token, delete_token);
 
    PERFORM set_config('provsql.update_provenance', 'off', false);
    EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2;', TG_TABLE_SCHEMA, TG_TABLE_NAME)
    USING new_token, old_token;
    PERFORM set_config('provsql.update_provenance', 'on', false);
  END LOOP;
 
  RETURN NULL;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp SECURITY DEFINER;
/** @endcond */
 
/**
 * @brief Trigger function for INSERT statement provenance tracking
 *
 * Records the insertion in update_provenance and multiplies provenance
 * tokens of inserted rows with the insert token.
 */
CREATE OR REPLACE FUNCTION insert_statement_trigger()
  RETURNS TRIGGER AS
$$
DECLARE
  query_text TEXT;
  insert_token UUID;
  old_token UUID;
  new_token UUID;
  r RECORD;
  enable_trigger BOOL;
BEGIN
  enable_trigger := current_setting('provsql.update_provenance', true);
  IF enable_trigger = 'f' THEN
    RETURN NULL;
  END IF;
 
  insert_token := public.uuid_generate_v4();
 
  PERFORM create_gate(insert_token, 'update');
 
  SELECT query
  INTO query_text
  FROM pg_stat_activity
  WHERE pid = pg_backend_pid();
 
  INSERT INTO update_provenance (provsql, query, query_type, username, ts, valid_time)
  VALUES (insert_token, query_text, 'INSERT', current_user, CURRENT_TIMESTAMP, TSTZMULTIRANGE(tstzrange(CURRENT_TIMESTAMP, NULL)));
 
  FOR r IN (SELECT * FROM NEW_TABLE) LOOP
    old_token := r.provsql;
    new_token := provenance_times(old_token, insert_token);
    PERFORM set_config('provsql.update_provenance', 'off', false);
    EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2;', TG_TABLE_SCHEMA, TG_TABLE_NAME)
    USING new_token, old_token;
    PERFORM set_config('provsql.update_provenance', 'on', false);
  END LOOP;
 
  RETURN NULL;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp SECURITY DEFINER;
 
/**
 * @brief Trigger function for UPDATE statement provenance tracking
 *
 * Records the update in update_provenance. Multiplies new-row tokens
 * with the update token and applies monus to old-row tokens.
 */
CREATE OR REPLACE FUNCTION update_statement_trigger()
  RETURNS TRIGGER AS
$$
DECLARE
  query_text TEXT;
  update_token UUID;
  old_token UUID;
  new_token UUID;
  r RECORD;
  enable_trigger BOOL;
BEGIN
  enable_trigger := current_setting('provsql.update_provenance', true);
  IF enable_trigger = 'f' THEN
    RETURN NULL;
  END IF;
  update_token := public.uuid_generate_v4();
 
  PERFORM create_gate(update_token, 'update');
 
  SELECT query
  INTO query_text
  FROM pg_stat_activity
  WHERE pid = pg_backend_pid();
 
  INSERT INTO update_provenance (provsql, query, query_type, username, ts, valid_time)
  VALUES (update_token, query_text, 'UPDATE', current_user, CURRENT_TIMESTAMP, TSTZMULTIRANGE(tstzrange(CURRENT_TIMESTAMP, NULL)));
 
  FOR r IN (SELECT * FROM NEW_TABLE) LOOP
    old_token := r.provsql;
    new_token := provenance_times(old_token, update_token);
 
    PERFORM set_config('provsql.update_provenance', 'off', false);
    EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2;', TG_TABLE_SCHEMA, TG_TABLE_NAME)
    USING new_token, old_token;
    PERFORM set_config('provsql.update_provenance', 'on', false);
  END LOOP;
 
  PERFORM set_config('provsql.update_provenance', 'off', false);
  EXECUTE format('INSERT INTO %I.%I SELECT * FROM OLD_TABLE;', TG_TABLE_SCHEMA, TG_TABLE_NAME);
  PERFORM set_config('provsql.update_provenance', 'on', false);
 
  FOR r IN (SELECT * FROM OLD_TABLE) LOOP
    old_token := r.provsql;
    new_token := provenance_monus(old_token, update_token);
 
    PERFORM set_config('provsql.update_provenance', 'off', false);
    EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2;', TG_TABLE_SCHEMA, TG_TABLE_NAME)
    USING new_token, old_token;
    PERFORM set_config('provsql.update_provenance', 'on', false);
  END LOOP;
 
  RETURN NULL;
END
$$ LANGUAGE plpgsql SET search_path=provsql,pg_temp SECURITY DEFINER;
 
 
/** @} */
 
/** @defgroup temporal_db Temporal DB (PostgreSQL 14+)
 *  Functions for temporal database support. These use provenance
 *  evaluation over the multirange semiring to track temporal validity
 *  of tuples.
 *  @{
 */
 
SET search_path TO provsql;
 
/**
 * @brief Evaluate provenance over the temporal (interval-union) m-semiring
 *
 * Inputs are read as %TSTZMULTIRANGE validity intervals; the additive
 * identity is <tt>'{}'::%TSTZMULTIRANGE</tt> (empty), the multiplicative
 * identity is <tt>'{(,)}'::%TSTZMULTIRANGE</tt> (universal). Returns the union
 * of intervals supporting the result, computed via the compiled circuit
 * traversal.
 *
 * @param token       Provenance token to evaluate.
 * @param token2value Mapping from input gates to validity multiranges.
 */
CREATE FUNCTION sr_temporal(token ANYELEMENT, token2value REGCLASS)
  RETURNS TSTZMULTIRANGE AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'interval_union',
    '{(,)}'::TSTZMULTIRANGE
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/**
 * @brief Evaluate provenance over the interval-union m-semiring
 *        with a NUMERIC multirange carrier
 *
 * Inputs are read as %nummultirange validity ranges over a NUMERIC
 * domain (e.g. sensor measurement-validity ranges). Addition is
 * multirange union, multiplication is intersection, monus is set
 * difference; the additive identity is <tt>'{}'::%nummultirange</tt>
 * and the multiplicative identity is <tt>'{(,)}'::%nummultirange</tt>
 * (universal range).
 *
 * @param token       Provenance token to evaluate.
 * @param token2value Mapping from input gates to NUMERIC multiranges.
 */
CREATE FUNCTION sr_interval_num(token ANYELEMENT, token2value REGCLASS)
  RETURNS nummultirange AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'interval_union',
    '{(,)}'::nummultirange
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/**
 * @brief Evaluate provenance over the interval-union m-semiring
 *        with an int4 multirange carrier
 *
 * Inputs are read as %int4multirange validity ranges over the
 * integers (e.g. page or line ranges of supporting documents).
 * Addition is multirange union, multiplication is intersection,
 * monus is set difference; the additive identity is
 * <tt>'{}'::%int4multirange</tt> and the multiplicative identity is
 * <tt>'{(,)}'::%int4multirange</tt>.
 *
 * @param token       Provenance token to evaluate.
 * @param token2value Mapping from input gates to int4 multiranges.
 */
CREATE FUNCTION sr_interval_int(token ANYELEMENT, token2value REGCLASS)
  RETURNS int4multirange AS
$$
BEGIN
  RETURN provsql.provenance_evaluate_compiled(
    token,
    token2value,
    'interval_union',
    '{(,)}'::int4multirange
  );
END
$$ LANGUAGE plpgsql STRICT PARALLEL SAFE STABLE;
 
/**
 * @brief Evaluate temporal provenance as a TIMESTAMP multirange
 *
 * Thin wrapper around :sqlfunc:`sr_temporal` retained for backward
 * compatibility; both compute the same union of validity intervals.
 *
 * @param token provenance token to evaluate
 * @param token2value mapping table from tokens to temporal validity ranges
 */
CREATE OR REPLACE FUNCTION union_tstzintervals(
    token UUID,
    token2value REGCLASS
)
RETURNS TSTZMULTIRANGE AS
$$
  SELECT sr_temporal(token, token2value)
$$ LANGUAGE SQL PARALLEL SAFE STABLE;
 
/**
 * @brief Query a table as it was at a specific point in time
 *
 * Returns all rows whose temporal validity includes the given TIMESTAMP.
 *
 * @param tablename name of the provenance-tracked table
 * @param at_time the point in time to query
 */
CREATE OR REPLACE FUNCTION timetravel(
  tablename TEXT,
  at_time TIMESTAMPTZ
)
RETURNS SETOF RECORD
LANGUAGE plpgsql
AS
$$
BEGIN
    RETURN QUERY EXECUTE format(
        '
          SELECT
            %1$I.*,
            sr_temporal(provenance(), %2$L)
          FROM
            %1$I
          WHERE
            sr_temporal(provenance(), %2$L) @> %3$L::TIMESTAMPTZ
        ',
        tablename,
        'provsql.time_validity_view',
        at_time::TEXT
    );
END;
$$;
 
/**
 * @brief Query a table for rows valid during a time interval
 *
 * Returns all rows whose temporal validity overlaps the given range.
 *
 * @param tablename name of the provenance-tracked table
 * @param from_time start of the time interval
 * @param to_time end of the time interval
 */
CREATE OR REPLACE FUNCTION timeslice(
  tablename TEXT,
  from_time TIMESTAMPTZ,
  to_time TIMESTAMPTZ
)
RETURNS SETOF RECORD
LANGUAGE plpgsql
AS
$$
BEGIN
  RETURN QUERY EXECUTE format(
    '
      SELECT
        %1$I.*,
        sr_temporal(provenance(), %2$L)
      FROM
        %1$I
      WHERE
        sr_temporal(provenance(), %2$L)
        && tstzrange(%3$L::TIMESTAMPTZ, %4$L::TIMESTAMPTZ)
    ',
    tablename,
    'provsql.time_validity_view',
    from_time::TEXT,
    to_time::TEXT
  );
END;
$$;
 
/**
 * @brief Query the full temporal history of specific rows
 *
 * Returns all versions of rows matching the given column values,
 * with their temporal validity ranges.
 *
 * @param tablename name of the provenance-tracked table
 * @param col_names array of column names to filter on
 * @param col_values array of corresponding values to match
 */
CREATE OR REPLACE FUNCTION history(
  tablename TEXT,
  col_names TEXT[],
  col_values TEXT[]
)
RETURNS SETOF RECORD
LANGUAGE plpgsql
AS
$$
DECLARE
    condition TEXT := '';
    i         INT;
BEGIN
    IF array_length(col_names, 1) IS NULL
       OR array_length(col_values, 1) IS NULL
       OR array_length(col_names, 1) != array_length(col_values, 1)
    THEN
        RAISE EXCEPTION 'col_names and col_values must have the same (non-null) length';
    END IF;
 
    FOR i IN 1..array_length(col_names, 1)
    LOOP
        IF i > 1 THEN
            condition := condition || ' AND ';
        END IF;
        condition := condition || format('%I = %L', col_names[i], col_values[i]);
    END LOOP;
 
    RETURN QUERY EXECUTE format(
      '
        SELECT
          %I.*,
          sr_temporal(provenance(), %L)
        FROM
          %I
        WHERE
          %s
      ',
      tablename,
      'provsql.time_validity_view',
      tablename,
      condition
    );
END;
$$;
 
/**
 * @brief Get the valid time range for a specific tuple
 *
 * @param token provenance token of the tuple
 * @param tablename name of the table containing the tuple
 */
CREATE OR REPLACE FUNCTION get_valid_time(
  token UUID,
  tablename TEXT
)
RETURNS TSTZMULTIRANGE
LANGUAGE plpgsql
AS $$
DECLARE
  result TSTZMULTIRANGE;
BEGIN
  EXECUTE format(
    '
      SELECT
        sr_temporal(provenance(), %L)
      FROM
        %I
      WHERE
        provsql = %L
    ',
    'provsql.time_validity_view',
    tablename,
    token
  )
  INTO result;
 
  RETURN result;
END;
$$;
 
/**
 * @brief Undo a previously recorded update operation
 *
 * Traverses all provenance-tracked tables and rewrites their circuits
 * to apply monus with respect to the given update token, effectively
 * undoing the operation.
 *
 * @param c UUID of the update operation to undo (from update_provenance)
 */
CREATE OR REPLACE FUNCTION undo(
  c UUID
)
RETURNS UUID
LANGUAGE plpgsql
AS $$
DECLARE
  undo_query TEXT;
  undone_query TEXT;
  undo_token UUID;
  schema_rec RECORD;
  table_rec RECORD;
  row_rec RECORD;
  new_x UUID;
BEGIN
  SELECT query INTO undone_query
  FROM update_provenance
  WHERE provsql = c
  LIMIT 1;
 
  IF undone_query IS NULL THEN
    RAISE NOTICE 'Unable to find % in update_provenance', c;
    RETURN c;
  END IF;
 
  SELECT query
  INTO undo_query
  FROM pg_stat_activity
  WHERE pid = pg_backend_pid();
 
  undo_token := public.uuid_generate_v4();
  PERFORM create_gate(undo_token, 'update');
  INSERT INTO update_provenance(provsql, query, query_type, username, ts, valid_time)
  VALUES (
    undo_token,
    undo_query,
    'UNDO',
    current_user,
    CURRENT_TIMESTAMP,
    TSTZMULTIRANGE(tstzrange(CURRENT_TIMESTAMP, NULL))
  );
 
  PERFORM set_config('provsql.update_provenance', 'off', false);
 
  FOR schema_rec IN
    SELECT nspname
    FROM pg_namespace
    WHERE nspname NOT IN ('pg_catalog','information_schema','pg_toast','pg_temp_1','pg_toast_temp_1')
  LOOP
    FOR table_rec IN
      EXECUTE format('SELECT tablename AS tname FROM pg_tables WHERE schemaname = %L', schema_rec.nspname)
    LOOP
      IF EXISTS (
        SELECT 1
        FROM information_schema.columns
        WHERE table_schema = schema_rec.nspname
          AND table_name = table_rec.tname
          AND table_name <> 'update_provenance'
          AND column_name = 'provsql'
      ) THEN
        FOR row_rec IN
          EXECUTE format('SELECT provsql AS x FROM %I.%I', schema_rec.nspname, table_rec.tname)
        LOOP
          new_x := replace_the_circuit(row_rec.x, c, undo_token);
          EXECUTE format('UPDATE %I.%I SET provsql = $1 WHERE provsql = $2',
                         schema_rec.nspname, table_rec.tname)
          USING new_x, row_rec.x;
        END LOOP;
      END IF;
    END LOOP;
  END LOOP;
 
  PERFORM set_config('provsql.update_provenance', 'on', false);
 
  RETURN undo_token;
END;
$$;
 
/**
 * @brief Recursively rewrite a circuit to undo a specific operation
 *
 * Helper for undo(). Walks the circuit and replaces occurrences of
 * the target update gate with its monus.
 *
 * @param x provenance token to rewrite
 * @param c UUID of the update operation to undo
 * @param u UUID of the undo operation
 */
CREATE OR REPLACE FUNCTION replace_the_circuit(
  x UUID,
  c UUID,
  u UUID
)
RETURNS UUID
LANGUAGE plpgsql
AS $$
DECLARE
  nchildren UUID[];
  child UUID;
  ntoken UUID;
  ntype PROVENANCE_GATE;
BEGIN
  IF x = c THEN
    RETURN provenance_monus(c, u);
  -- update and input gates cannot have children
  ELSIF get_gate_type(x) = 'update' OR get_gate_type(x) = 'input' THEN
    RETURN x;
  ELSE
    nchildren := '{}';
    FOREACH child IN ARRAY get_children(x)
    LOOP
      nchildren := array_append(nchildren, replace_the_circuit(child, c, u));
    END LOOP;
 
    ntoken := public.uuid_generate_v4();
    ntype := get_gate_type(x);
 
    PERFORM create_gate(ntoken, ntype, nchildren);
    RETURN ntoken;
  END IF;
END;
$$;
 
-- The base validity mapping is a plain view over the data-modification log:
-- update_provenance is append-only and never has its provsql rewritten, so a
-- view stays correct (unlike a tracked table's mapping, which must be a
-- maintained mapping table -- see create_provenance_mapping(maintained)).
CREATE VIEW provsql.time_validity_view AS
  SELECT valid_time AS value, provsql AS provenance FROM provsql.update_provenance;
 
/** @} */
 
SET search_path TO public;
 
-- Final constants-cache refresh: same rationale as at the end of
-- provsql.common.sql.  On PG14+ this file is appended after the common
-- script, so this is the last statement of the generated install script;
-- the refresh must come after every object has been created for the
-- installing session's memoized constants to be complete.
SELECT provsql.reset_constants_cache();