@quereus/quereus 3.3.0 → 4.0.0

package/dist/src/planner/util/fd-utils.js

@@ -8,8 +8,9 @@ import { createLogger } from '../../common/logger.js';
 import { ColumnReferenceNode, ParameterReferenceNode } from '../nodes/reference.js';
 import { BetweenNode, BinaryOpNode, CastNode, CollateNode, LiteralNode, UnaryOpNode } from '../nodes/scalar.js';
 import { InNode } from '../nodes/subquery.js';
-import { compareSqlValues } from '../../util/comparison.js';
+import { compareSqlValues, normalizeCollationName } from '../../util/comparison.js';
 import { flipComparison } from '../analysis/predicate-shape.js';
+import { effectiveBetweenBoundCollation, effectiveComparisonCollation, effectiveInCollation, isValueDiscriminatingEquality, operandCollation } from '../analysis/comparison-collation.js';
 const log = createLogger('planner:fd');
 /**
  * Per-node cap on the number of FDs we materialize. The propagation rules
@@ -51,6 +52,8 @@ export function computeClosure(attrs, fds) {
  * column indices, then concatenate with the existing FDs. For a class
  * `{c0, c1, ..., ck}` this emits `{ci} → {cj}` for every distinct ordered pair
  * — enough for `computeClosure` to derive every member from any one of them.
+ * EC-derived FDs are pure value claims (`kind: 'determination'`) — an equality
+ * never implies row-uniqueness of either endpoint.
  */
 export function expandEcsToFds(ecs, fds) {
     const out = fds.slice();
@@ -61,7 +64,7 @@ export function expandEcsToFds(ecs, fds) {
             for (let j = 0; j < cls.length; j++) {
                 if (i === j)
                     continue;
-                out.push({ determinants: [cls[i]], dependents: [cls[j]] });
+                out.push({ determinants: [cls[i]], dependents: [cls[j]], kind: 'determination' });
             }
         }
     }
@@ -110,6 +113,9 @@ export function minimalCover(attrs, fds) {
 // collapse to one in `addFd` / `mergeFds`. The first-merged source wins —
 // table references merge declared-check contributions before hoisted
 // assertion contributions, so `declared-check` is preferred on collisions.
+// `kind` and `valueEquality` are likewise NOT compared; `addFd` reconciles
+// `kind` on collisions with an "'unique' wins" rule (uniqueness is a property
+// of the determinant set, so equal-determinant claims compose).
 function fdsEqual(a, b) {
     if (a.determinants.length !== b.determinants.length)
         return false;
@@ -233,33 +239,52 @@ function dependentsSubset(sub, sup) {
  * Guard-aware: FDs with different `guard` predicates are kept side-by-side
  * even when their determinants/dependents match — they are logically distinct
  * facts and may be activated by different surrounding predicates.
+ *
+ * Kind reconciliation: on a merge between entries with equal determinants and
+ * guards, the surviving entry's `kind` is `'unique'` when EITHER side claims
+ * it — uniqueness is a property of the determinant set, so equal-determinant
+ * claims compose. This includes upgrading a kept 'determination' entry in
+ * place when the subsumed newcomer is 'unique'. Equal-determinant entries with
+ * incomparable dependent sets both survive and each keeps its own kind (a
+ * sound under-claim). Object identity is preserved when nothing changes.
  */
 export function addFd(fds, next, opts = {}) {
     if (next.dependents.length === 0)
         return fds.slice();
     const result = [];
     let subsumedByExisting = false;
+    // 'unique' wins on merge: a dropped-or-subsumed 'unique' twin upgrades the survivor.
+    let nextKind = next.kind;
     for (const existing of fds) {
         if (fdsEqual(existing, next)) {
             subsumedByExisting = true;
-            result.push(existing);
+            result.push(next.kind === 'unique' && existing.kind !== 'unique'
+                ? { ...existing, kind: 'unique' }
+                : existing);
             continue;
         }
         if (determinantsEqual(existing.determinants, next.determinants) &&
             guardsEqual(existing.guard, next.guard)) {
             // Same determinants and guard: keep whichever has the larger dependent set.
             if (dependentsSubset(existing.dependents, next.dependents)) {
-                // existing ⊂ next, drop existing
+                // existing ⊂ next, drop existing — its uniqueness claim survives on next.
+                if (existing.kind === 'unique')
+                    nextKind = 'unique';
                 continue;
             }
             if (dependentsSubset(next.dependents, existing.dependents)) {
                 subsumedByExisting = true;
+                result.push(next.kind === 'unique' && existing.kind !== 'unique'
+                    ? { ...existing, kind: 'unique' }
+                    : existing);
+                continue;
             }
         }
         result.push(existing);
     }
-    if (!subsumedByExisting)
-        result.push(next);
+    if (!subsumedByExisting) {
+        result.push(nextKind === next.kind ? next : { ...next, kind: nextKind });
+    }
     return enforceCap(result, opts);
 }
 function enforceCap(fds, opts) {
@@ -275,12 +300,19 @@ function enforceCap(fds, opts) {
     };
     const preferred = fds.filter(fd => isSubsetOfAnyKey(fd.determinants));
     const other = fds.filter(fd => !isSubsetOfAnyKey(fd.determinants));
+    // Quality bias within each partition: keep 'unique' FDs ahead of plain
+    // determinations. Evicting a uniqueness witness can only cause downstream
+    // under-claims (sound), but it is cheap to avoid.
+    const uniqueFirst = (list) => [
+        ...list.filter(fd => fd.kind === 'unique'),
+        ...list.filter(fd => fd.kind !== 'unique'),
+    ];
     let kept;
     if (preferred.length >= cap) {
-        kept = preferred.slice(0, cap);
+        kept = uniqueFirst(preferred).slice(0, cap);
     }
     else {
-        kept = preferred.concat(other.slice(0, cap - preferred.length));
+        kept = preferred.concat(uniqueFirst(other).slice(0, cap - preferred.length));
     }
     log('FD cap reached: dropped %d FD(s) from %d', fds.length - kept.length, fds.length);
     return kept;
@@ -307,6 +339,14 @@ export function mergeFds(a, b, opts = {}) {
  * Guarded FDs additionally require every column referenced in `guard.clauses`
  * to be in the mapping — if any guard column is dropped the guard becomes
  * unobservable and the FD can never be re-activated downstream.
+ *
+ * Rebuilds via spread so `kind` / `source` / `valueEquality` survive verbatim.
+ * Preserving `kind` is sound: a projection maps rows 1:1 (no merge, no
+ * duplication), so determinant row-uniqueness survives whenever the
+ * determinants survive — which the determinant-loss drop above already
+ * requires. The empty-determinant exception keeps its kind too: a 'unique'
+ * singleton stays ≤1-row under projection; a 'determination' constant pin
+ * stays a mere pin.
  */
 export function projectFds(fds, mapping) {
     const result = [];
@@ -339,8 +379,8 @@ export function projectFds(fds, mapping) {
             newGuard = mappedGuard;
         }
         result.push(newGuard
-            ? { determinants: newDet, dependents: newDep, guard: newGuard }
-            : { determinants: newDet, dependents: newDep });
+            ? { ...fd, determinants: newDet, dependents: newDep, guard: newGuard }
+            : { ...fd, determinants: newDet, dependents: newDep });
     }
     return result;
 }
@@ -395,12 +435,18 @@ function projectClause(clause, mapping) {
         }
     }
 }
-/** Shift all column indices in `fds` (including any `guard` columns) by `offset`. */
+/**
+ * Shift all column indices in `fds` (including any `guard` columns) by `offset`.
+ * Rebuilds via spread so `kind` / `source` / `valueEquality` survive verbatim —
+ * a shift is a pure column relabel, so every claim (uniqueness included) holds
+ * unchanged on the relabeled columns.
+ */
 export function shiftFds(fds, offset) {
     if (offset === 0)
         return fds.slice();
     return fds.map(fd => {
         const shifted = {
+            ...fd,
             determinants: fd.determinants.map(d => d + offset),
             dependents: fd.dependents.map(d => d + offset),
         };
@@ -428,14 +474,18 @@ function shiftClause(clause, offset) {
     }
 }
 /**
- * Return the unconditional twin of `fd` — drop the guard but keep determinants
- * and dependents. Used by Filter activation when the surrounding predicate
- * entails the guard.
+ * Return the unconditional twin of `fd` — drop the guard but keep every other
+ * field (`kind` / `source` / `valueEquality` survive verbatim). Used by Filter
+ * activation when the surrounding predicate entails the guard. Preserving
+ * 'unique' is sound at the activating Filter: its rows all satisfy the guard,
+ * and filtering only shrinks the row set — fan-out hazards are handled by the
+ * join-side downgrade, not here.
  */
 export function stripGuard(fd) {
     if (fd.guard === undefined)
         return fd;
-    return { determinants: fd.determinants, dependents: fd.dependents };
+    const { guard: _guard, ...rest } = fd;
+    return rest;
 }
 /** Shift all column indices in `classes` by `offset`. */
 export function shiftEquivClasses(classes, offset) {
@@ -491,6 +541,11 @@ export function addEquivalence(classes, a, b) {
  * way to encode "K is a unique key on a relation": K determines every other
  * output column. K = ∅ produces the "at-most-one-row" singleton FD.
  *
+ * Emits `kind: 'unique'` — every caller passes a genuine key (declared or
+ * projected keys, fan-out-aware join `preservedKeys`, the aggregate group key,
+ * the set-op data-columns key, lens key obligations, TVF-declared keys), so
+ * the relation has at most one row per determinant tuple at the minting site.
+ *
  * Returns undefined when K covers every column (the all-columns case has no
  * non-trivial encoding — that case is communicated via `RelationType.isSet`
  * instead).
@@ -504,14 +559,19 @@ export function superkeyToFd(key, columnCount) {
     }
     if (dependents.length === 0)
         return undefined;
-    return { determinants: key.slice(), dependents };
+    return { determinants: key.slice(), dependents, kind: 'unique' };
 }
 /**
- * True iff the closure of `attrs` under `fds` covers `{0..columnCount-1}` —
- * i.e., `attrs` is a superkey of the relation. Replaces the legacy "covers a
- * `uniqueKeys` entry" check; FDs are the canonical surface now.
+ * True iff the closure of `attrs` under `fds` covers `{0..columnCount-1}`.
+ * COVERAGE ONLY — this is a pure value-determination claim and says NOTHING
+ * about row-uniqueness: over a bag, a determination-only closure path can
+ * cover every column while the relation still holds duplicate rows. For the
+ * uniqueness question ("at most one row per attrs-tuple?") use
+ * `isUniqueDeterminant`. (Renamed from `isSuperkey`, whose name read as a
+ * uniqueness claim and invited exactly that misuse — ticket
+ * `fd-determination-reader-side-rule`.)
  */
-export function isSuperkey(attrs, fds, columnCount) {
+export function closureCoversAll(attrs, fds, columnCount) {
     if (columnCount <= 0)
         return true;
     const closure = computeClosure(attrs, fds ?? []);
@@ -521,16 +581,52 @@ export function isSuperkey(attrs, fds, columnCount) {
     }
     return true;
 }
+/**
+ * True iff `attrs` is provably row-unique on the relation: its FD closure
+ * covers every column AND uniqueness is reachable —
+ *   - the relation is a set (two rows agreeing on `attrs` would agree on all
+ *     columns = a duplicate, impossible in a set), or
+ *   - some unguarded `kind: 'unique'` FD has determinants ⊆ closure(attrs)
+ *     (rows agreeing on `attrs` agree on that unique determinant set; ≤1 row
+ *     per its tuple ⇒ ≤1 row per attrs-tuple).
+ *
+ * Coverage alone (a determination-only closure path over a bag) proves
+ * nothing — that is the over-claim family the kind-aware readers exist to
+ * prevent (ticket `fd-determination-reader-side-rule`). This is the single
+ * reader-side uniqueness primitive; producers no longer gate determinations.
+ *
+ * Guarded FDs participate in neither branch: `computeClosure` skips them, and
+ * only UNguarded 'unique' FDs can witness (a guarded uniqueness claim holds
+ * only under its predicate, which this layer cannot see).
+ */
+export function isUniqueDeterminant(attrs, fds, columnCount, isSet) {
+    const fdList = fds ?? [];
+    const closure = computeClosure(attrs, fdList);
+    for (let i = 0; i < columnCount; i++) {
+        if (!closure.has(i))
+            return false;
+    }
+    if (isSet)
+        return true;
+    return fdList.some(fd => fd.guard === undefined &&
+        fd.kind === 'unique' &&
+        fd.determinants.every(d => closure.has(d)));
+}
 /**
  * Enumerate the minimal full-cover key sets discoverable from `fds`: for each
- * FD `K → Y` whose closure covers all columns, return `K` (greedily minimized
- * within `K`). Deduplicated by set equality.
+ * FD `K → Y` where `K` is a provably row-unique determinant
+ * (`isUniqueDeterminant` — coverage AND uniqueness reachability), return `K`
+ * (greedily minimized within `K`). Deduplicated by set equality.
+ *
+ * `minimalCover` preserves the closure, so the minimized key keeps both
+ * coverage and the unique witness (witness determinants ⊆ the unchanged
+ * closure).
  *
- * Excludes the trivial "all-columns is a superkey" tautology — only FDs with
- * `K ⊊ all_cols` are considered, since the all-cols case is encoded via
- * `RelationType.isSet`.
+ * Excludes the trivial "all-columns is a key of a set" tautology — only FDs
+ * with `K ⊊ all_cols` are considered, since the all-cols case is encoded via
+ * `RelationType.isSet` (the `keysOf` fallback).
  */
-export function deriveKeysFromFds(fds, columnCount) {
+export function deriveKeysFromFds(fds, columnCount, isSet) {
     if (!fds || fds.length === 0)
         return [];
     const results = [];
@@ -541,11 +637,9 @@ export function deriveKeysFromFds(fds, columnCount) {
         if (fd.determinants.length >= columnCount)
             continue;
         const det = new Set(fd.determinants);
-        if (!isSuperkey(det, fds, columnCount))
+        if (!isUniqueDeterminant(det, fds, columnCount, isSet))
             continue;
         const minimal = minimalCover(det, fds);
-        // Ensure the minimal cover still covers all columns (it should — minimalCover
-        // only drops attrs whose removal doesn't change closure).
         const sorted = Array.from(minimal).sort((a, b) => a - b);
         const key = sorted.join(',');
         if (seen.has(key))
@@ -557,35 +651,42 @@ export function deriveKeysFromFds(fds, columnCount) {
 }
 /**
  * True iff the FD set encodes any non-trivial key — i.e., there exists some
- * FD whose determinants form a superkey of `columnCount` columns with the
- * determinant set strictly smaller than all columns. This is the FD-surface
- * replacement for "the relation has a known unique key smaller than its full
- * column list" (the old `uniqueKeys.length > 0` check), excluding the
- * tautological all-columns case which carries no information.
+ * FD whose determinants are a provably row-unique determinant set
+ * (`isUniqueDeterminant`) strictly smaller than all columns. This is the
+ * FD-surface replacement for "the relation has a known unique key smaller
+ * than its full column list" (the old `uniqueKeys.length > 0` check),
+ * excluding the tautological all-columns case which carries no information.
  */
-export function hasAnyKey(fds, columnCount) {
+export function hasAnyKey(fds, columnCount, isSet) {
     if (!fds || fds.length === 0)
         return false;
     return fds.some(fd => fd.guard === undefined &&
         fd.determinants.length < columnCount &&
-        isSuperkey(new Set(fd.determinants), fds, columnCount));
+        isUniqueDeterminant(new Set(fd.determinants), fds, columnCount, isSet));
 }
 /**
- * True iff the relation has at-most-one-row — i.e., some FD `∅ → Y` exists
- * whose closure covers every column. Replaces the legacy `[[]]` singleton
- * marker on `uniqueKeys`.
+ * True iff the relation is provably at-most-one-row from its FD surface —
+ * `isUniqueDeterminant(∅, …)`: the closure of the empty set covers every
+ * column AND uniqueness is reachable (the relation is a set, or an unguarded
+ * 'unique' FD witnesses). Replaces the legacy `[[]]` singleton marker on
+ * `uniqueKeys`.
+ *
+ * On a bag, constant pins (`∅ → col` determinations from `where a = 1`) no
+ * longer over-claim ≤1-row; on a set, pinning every column IS a sound ≤1-row
+ * derivation. Zero-column relations return false — the `∅ → all_cols` FD is
+ * unrepresentable there, so the ≤1-row claim rides `estimatedRows` instead
+ * (see `characteristics.guaranteesUniqueRows`).
  */
-export function hasSingletonFd(fds, columnCount) {
-    if (!fds)
+export function hasSingletonFd(fds, columnCount, isSet) {
+    if (columnCount <= 0)
         return false;
-    return fds.some(fd => fd.guard === undefined &&
-        fd.determinants.length === 0 &&
-        isSuperkey(new Set(), fds, columnCount));
+    return isUniqueDeterminant(new Set(), fds, columnCount, isSet);
 }
 /**
  * Build the singleton FD `∅ → {0..columnCount-1}` that encodes
- * "at-most-one-row". Returns undefined when `columnCount === 0` (no
- * dependents).
+ * "at-most-one-row". `kind: 'unique'` — ∅ row-unique ⟺ ≤1 row, which is
+ * exactly what every caller asserts. Returns undefined when
+ * `columnCount === 0` (no dependents).
  */
 export function singletonFd(columnCount) {
     if (columnCount <= 0)
@@ -593,39 +694,21 @@ export function singletonFd(columnCount) {
     const dependents = [];
     for (let i = 0; i < columnCount; i++)
         dependents.push(i);
-    return { determinants: [], dependents };
+    return { determinants: [], dependents, kind: 'unique' };
 }
 /**
- * True iff `attrs` is asserted to be a unique key by the FD set — i.e., there
- * exists some FD whose determinants are a subset of `attrs` and whose closure
- * covers all columns. Stricter than `isSuperkey`: the trivial "all-cols is a
- * superkey of itself" tautology does NOT count, because no FD makes that claim.
+ * Fold the singleton FD `∅ → {0..columnCount-1}` ("at-most-one-row") into `fds`
+ * via `addFd`. The canonical producer-side spelling of the ≤1-row fact — every
+ * `computePhysical` site that proves a relation emits ≤1 row should reach for
+ * this rather than open-coding `singletonFd` + `addFd`.
  *
- * Use this when you need a positive uniqueness claim (e.g., the
- * sort/window strict-monotonicOn check). For "would attrs functionally
- * determine the rest of the relation under closure?" use `isSuperkey` directly.
+ * A no-op returning a copy of `fds` when `columnCount === 0` (since
+ * `singletonFd(0)` is `undefined` — a zero-column relation cannot carry the
+ * marker). Pairs with the `hasSingletonFd` / `isAtMostOneRow` read surface.
  */
-export function isAssertedKey(attrs, fds, columnCount) {
-    if (!fds || fds.length === 0)
-        return false;
-    for (const fd of fds) {
-        if (fd.guard !== undefined)
-            continue;
-        // Determinants must be a subset of attrs.
-        let subset = true;
-        for (const d of fd.determinants) {
-            if (!attrs.has(d)) {
-                subset = false;
-                break;
-            }
-        }
-        if (!subset)
-            continue;
-        // Determinants closure must cover all columns.
-        if (isSuperkey(new Set(fd.determinants), fds, columnCount))
-            return true;
-    }
-    return false;
+export function addSingletonFd(fds, columnCount) {
+    const singleton = singletonFd(columnCount);
+    return singleton ? addFd(fds, singleton) : fds.slice();
 }
 /**
  * Normalize a list of candidate keys to the minimal, deduped set:
@@ -703,12 +786,14 @@ export function keysOf(rel) {
     for (const key of type.keys) {
         keys.push(key.map(ref => ref.index));
     }
-    // 2. `∅ → all_cols` ⇒ at-most-one-row ⇒ the empty key.
-    if (hasSingletonFd(fds, columnCount)) {
+    // 2. Provable ≤1-row (kind-aware) ⇒ the empty key.
+    if (hasSingletonFd(fds, columnCount, type.isSet)) {
         keys.push([]);
     }
     // 3. FD-derived keys (already bounded to FDs with det.length < columnCount).
-    for (const k of deriveKeysFromFds(fds, columnCount)) {
+    // `isSet` threads through so a determination-only covering determinant over
+    // a set (e.g. above a DISTINCT) derives a genuine key.
+    for (const k of deriveKeysFromFds(fds, columnCount, type.isSet)) {
         keys.push(k);
     }
     const normalized = normalizeKeys(keys);
@@ -727,14 +812,14 @@ export function keysOf(rel) {
  *   - `cols` is a (non-strict) superset of some `keysOf(rel)` entry (covers
  *     declared keys, the ≤1-row empty key, FD-derived keys, and the
  *     all-columns/set key), OR
- *   - `cols` is a **proper subset** of the columns whose FD closure covers all
- *     columns (`isSuperkey`) — this proves a superkey even when it is absent
- *     from the minimal `keysOf` list.
+ *   - `cols` is a provably row-unique determinant (`isUniqueDeterminant`:
+ *     closure coverage AND uniqueness reachability) — this proves a superkey
+ *     even when it is absent from the minimal `keysOf` list.
  *
- * The closure branch is deliberately restricted to proper subsets: the closure
- * of the full column set is trivially the full set, so without the guard a bag
- * would be falsely reported unique on its all-columns set. The all-columns case
- * is handled soundly by the `keysOf` branch above, which gates it on `isSet`.
+ * No all-columns guard is needed on the closure branch: an all-columns probe
+ * on a bag fails `isUniqueDeterminant` on its own (no unique FD ⇒ false; if a
+ * unique FD exists the relation cannot hold duplicate rows, so true is
+ * correct).
  */
 export function isUnique(cols, rel) {
     const type = rel.getType();
@@ -744,10 +829,24 @@ export function isUnique(cols, rel) {
         if (key.every(c => colSet.has(c)))
             return true;
     }
-    if (colSet.size < columnCount && isSuperkey(colSet, rel.physical?.fds, columnCount)) {
-        return true;
-    }
-    return false;
+    return isUniqueDeterminant(colSet, rel.physical?.fds, columnCount, type.isSet);
+}
+/**
+ * The single named spelling of the node-level "at-most-one-row" predicate:
+ * true iff `rel` is provably ≤1-row. Defined as `isUnique([], rel)` — the empty
+ * key is a subset of every column set, so a relation carrying it (via a declared
+ * empty key, the `∅ → all_cols` singleton FD, or any other channel `keysOf`
+ * reconciles) reports unique on the empty column list.
+ *
+ * Use this at node / rule level. The FD-only `hasSingletonFd` is the lower-level
+ * test `keysOf` itself calls; `isAtMostOneRow` is the surface consumers (joins,
+ * sort elimination) should reach for. Note it does **not** capture the
+ * zero-column `estimatedRows === 1` case — a zero-column relation has no
+ * representable empty key — so consumers needing that fallback keep their own
+ * check (see `characteristics.guaranteesUniqueRows`).
+ */
+export function isAtMostOneRow(rel) {
+    return isUnique([], rel);
 }
 /**
  * Walk `predicate` (assumed to be a normalized conjunction) and extract FDs,
@@ -765,6 +864,19 @@ export function isUnique(cols, rel) {
  *     equivalence pair `[col1, col2]`.
  *
  * Non-equality conjuncts contribute nothing.
+ *
+ * **Collation gate.** Every extracted fact is a VALUE-level claim (a pinned
+ * column has one value across rows; `col1 = col2` rows are value-equal), so a
+ * conjunct only contributes when its comparison is value-discriminating
+ * (`isValueDiscriminatingEquality`): for textual operands the effective
+ * comparison collation must be BINARY. A NOCASE/RTRIM comparison — via a
+ * `COLLATE` wrapper on either side or a non-BINARY declared column collation —
+ * passes value-DIFFERENT rows ('Bob' = 'bob' NOCASE), so its facts would
+ * over-claim (false ≤1-row keys, false EC-driven inferences, wrong insert
+ * defaults — ticket `collation-blind-equality-fact-extraction`). Declared-
+ * collation covered-key ≤1-row detection is NOT lost: it flows through the
+ * independent (and collation-sound) `extractConstraints` path in
+ * `FilterNode.computePhysical`.
  */
 export function extractEqualityFds(predicate, attrIdToIndex) {
     const fds = [];
@@ -782,6 +894,9 @@ export function extractEqualityFds(predicate, attrIdToIndex) {
         }
         if (op !== '=')
             continue;
+        // Value-level facts only from value-discriminating comparisons (see doc).
+        if (!isValueDiscriminatingEquality(n.left, n.right))
+            continue;
         const lIsCol = n.left instanceof ColumnReferenceNode;
         const rIsCol = n.right instanceof ColumnReferenceNode;
         const lConst = constantValueOf(n.left);
@@ -790,8 +905,10 @@ export function extractEqualityFds(predicate, attrIdToIndex) {
             const lIdx = attrIdToIndex.get(n.left.attributeId);
             const rIdx = attrIdToIndex.get(n.right.attributeId);
             if (lIdx !== undefined && rIdx !== undefined && lIdx !== rIdx) {
-                fds.push({ determinants: [lIdx], dependents: [rIdx] });
-                fds.push({ determinants: [rIdx], dependents: [lIdx] });
+                // Mirror FDs from `col1 = col2` are pure value claims — equality
+                // never makes either endpoint row-unique.
+                fds.push({ determinants: [lIdx], dependents: [rIdx], kind: 'determination' });
+                fds.push({ determinants: [rIdx], dependents: [lIdx], kind: 'determination' });
                 equivPairs.push([lIdx, rIdx]);
             }
             continue;
@@ -799,7 +916,9 @@ export function extractEqualityFds(predicate, attrIdToIndex) {
         if (lIsCol && rConst !== undefined) {
             const lIdx = attrIdToIndex.get(n.left.attributeId);
             if (lIdx !== undefined) {
-                fds.push({ determinants: [], dependents: [lIdx] });
+                // `∅ → col` from a constant pin is deliberately 'determination': a
+                // pinned column does NOT imply the relation has at most one row.
+                fds.push({ determinants: [], dependents: [lIdx], kind: 'determination' });
                 constantBindings.push({ attrs: [lIdx], value: rConst });
             }
             continue;
@@ -807,7 +926,7 @@ export function extractEqualityFds(predicate, attrIdToIndex) {
         if (rIsCol && lConst !== undefined) {
             const rIdx = attrIdToIndex.get(n.right.attributeId);
             if (rIdx !== undefined) {
-                fds.push({ determinants: [], dependents: [rIdx] });
+                fds.push({ determinants: [], dependents: [rIdx], kind: 'determination' });
                 constantBindings.push({ attrs: [rIdx], value: lConst });
             }
             continue;
@@ -815,7 +934,44 @@ export function extractEqualityFds(predicate, attrIdToIndex) {
     }
     return { fds, equivPairs, constantBindings };
 }
-function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
+/**
+ * Collect per-column facts from a filter predicate for guard discharge.
+ *
+ * **Collation gate (per conjunct).** A fact may only discharge a guard when
+ * the filter's runtime comparison keeps filter-rows ⊆ guard-scope-rows. The
+ * guard predicate is evaluated under the column's *declared* collation at
+ * enforcement time — index maintenance for partial-UNIQUE guards, write-time
+ * CHECK evaluation for implication-form CHECK guards (constraint-builder
+ * threads declared collations into the CHECK scope types) — and guard
+ * recognition (`partial-unique-extraction.ts`, `check-extraction.ts`) only
+ * accepts bare column/literal AST shapes, so guard comparisons resolve to the
+ * declared collation. Hence:
+ *
+ *  - `col = lit` / singleton-IN facts: sound when the conjunct's effective
+ *    comparison collation is BINARY (value-equality implies equality under any
+ *    collation) OR equals the column's declared collation (same comparison the
+ *    guard scope uses; the strict `sqlValueEquals` literal match at discharge
+ *    then under-claims at worst). The previous code stripped `CollateNode`
+ *    from the literal side unconditionally, so `b = 'bob' collate nocase`
+ *    discharged a BINARY `eq-literal{b,'bob'}` guard while admitting rows
+ *    outside the partial-index scope (ticket
+ *    `collation-blind-equality-fact-extraction`, repro 4).
+ *  - `col1 = col2` facts: sound when both columns' contributed collations are
+ *    equal — then any operand-resolution order (filter conjunct or guard
+ *    spelling) lands on the same collation.
+ *  - range facts over TEXT bounds: the discharge subset check
+ *    (`filterRangeSubsetOfGuardRange`) compares bounds under BINARY, so the
+ *    filter's effective collation AND the column's declared collation must
+ *    both be BINARY (a BINARY bound comparison does not bound a NOCASE-ordered
+ *    row set). Non-text bounds are collation-inert and stay ungated. This is
+ *    deliberately stricter than the equality gate — a completeness loss for
+ *    collated text ranges, never a soundness one.
+ *  - plain `col IN (lits)`: the runtime (`emitIn`) compares under the
+ *    condition operand's collation, which for the only recognized shape (a
+ *    bare column) IS the declared collation; listed values' COLLATE wrappers
+ *    are inert. Gated uniformly anyway for future-proofing.
+ */
+function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric, declaredCollationOf) {
     const literalEqs = new Map();
     const columnEqs = new Map();
     const isNullCols = new Set();
@@ -844,6 +1000,13 @@ function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
         }
         return undefined;
     };
+    // Equality-fact gate: BINARY, or matching the column's declared collation
+    // (the collation guard scopes are evaluated under). See the function doc.
+    const equalityCollationOk = (col, effColl) => effColl === 'BINARY' || effColl === normalizeCollationName(declaredCollationOf(col));
+    // Range-fact gate for TEXT bounds: the discharge subset check is BINARY, so
+    // both the filter comparison and the guard-scope (declared) collation must be.
+    const rangeCollationOk = (col, effColl, bound) => typeof bound !== 'string'
+        || (effColl === 'BINARY' && normalizeCollationName(declaredCollationOf(col)) === 'BINARY');
     const tightenLowerBound = (col, value, inclusive) => {
         const cur = rangeBounds.get(col);
         if (!cur) {
@@ -914,18 +1077,23 @@ function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
                 const lIdx = columnIndexOf(n.left);
                 const rIdx = columnIndexOf(n.right);
                 if (lIdx !== undefined && rIdx !== undefined) {
-                    addColumnEq(lIdx, rIdx);
+                    // Both sides are bare column refs; require their contributed
+                    // collations to agree so any resolution order matches the guard's.
+                    if (operandCollation(n.left) === operandCollation(n.right)) {
+                        addColumnEq(lIdx, rIdx);
+                    }
                     continue;
                 }
+                const effColl = effectiveComparisonCollation(n.left, n.right);
                 if (lIdx !== undefined) {
                     const lit = literalSqlValueOf(n.right);
-                    if (lit !== undefined)
+                    if (lit !== undefined && equalityCollationOk(lIdx, effColl))
                         literalEqs.set(lIdx, lit);
                     continue;
                 }
                 if (rIdx !== undefined) {
                     const lit = literalSqlValueOf(n.left);
-                    if (lit !== undefined)
+                    if (lit !== undefined && equalityCollationOk(rIdx, effColl))
                         literalEqs.set(rIdx, lit);
                 }
                 continue;
@@ -947,16 +1115,21 @@ function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
             if (op === '<' || op === '<=' || op === '>' || op === '>=') {
                 const lIdx = columnIndexOf(n.left);
                 const rIdx = columnIndexOf(n.right);
+                const effColl = effectiveComparisonCollation(n.left, n.right);
                 if (lIdx !== undefined && rIdx === undefined) {
                     const lit = literalSqlValueOf(n.right);
                     if (lit === undefined || lit === null)
                         continue;
+                    if (!rangeCollationOk(lIdx, effColl, lit))
+                        continue;
                     recordComparison(lIdx, op, lit);
                 }
                 else if (rIdx !== undefined && lIdx === undefined) {
                     const lit = literalSqlValueOf(n.left);
                     if (lit === undefined || lit === null)
                         continue;
+                    if (!rangeCollationOk(rIdx, effColl, lit))
+                        continue;
                     recordComparison(rIdx, flipComparison(op), lit);
                 }
                 continue;
@@ -971,10 +1144,15 @@ function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
                 continue;
             const lo = literalSqlValueOf(n.lower);
             const hi = literalSqlValueOf(n.upper);
-            if (lo !== undefined && lo !== null)
+            // Per-bound effective collation (emitBetween: bound wins over expr).
+            if (lo !== undefined && lo !== null
+                && rangeCollationOk(cIdx, effectiveBetweenBoundCollation(n.expr, n.lower), lo)) {
                 tightenLowerBound(cIdx, lo, true);
-            if (hi !== undefined && hi !== null)
+            }
+            if (hi !== undefined && hi !== null
+                && rangeCollationOk(cIdx, effectiveBetweenBoundCollation(n.expr, n.upper), hi)) {
                 tightenUpperBound(cIdx, hi, true);
+            }
             continue;
         }
         if (n instanceof UnaryOpNode) {
@@ -1011,6 +1189,12 @@ function buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric) {
             const cIdx = columnIndexOf(n.condition);
             if (cIdx === undefined)
                 continue;
+            // emitIn resolves ONE collation through the provenance lattice over the
+            // condition AND the listed values — a collate-wrapped element (folded
+            // to a literal whose type carries rank-3 NOCASE) can outrank the bare
+            // column recognized here, so this gate is load-bearing.
+            if (!equalityCollationOk(cIdx, effectiveInCollation(n)))
+                continue;
             const set = new Set();
             let allLiterals = true;
             for (const v of n.values) {
@@ -1088,9 +1272,13 @@ function bindingForColumn(col, bindings) {
  * sound for numeric columns since the consumer matches `eq-literal{col, 0}`
  * via strict `sqlValueEquals`, which treats TEXT `''`, BLOB, and boolean
  * `false` as unequal to integer 0.
+ *
+ * `declaredCollationOf(col)` reports the source's output column collation
+ * (normalized or not; `'BINARY'` when undeclared). Used by the per-conjunct
+ * collation gate in `buildPredicateFacts` — see its doc.
  */
-export function predicateImpliesGuard(predicate, guard, ecs, bindings, attrIdToIndex, isColumnNonNullable, isColumnNumeric) {
-    const facts = buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric);
+export function predicateImpliesGuard(predicate, guard, ecs, bindings, attrIdToIndex, isColumnNonNullable, isColumnNumeric, declaredCollationOf) {
+    const facts = buildPredicateFacts(predicate, attrIdToIndex, isColumnNumeric, declaredCollationOf);
     for (const clause of guard.clauses) {
         if (!clauseEntailed(clause, facts, ecs, bindings, isColumnNonNullable)) {
             return false;
@@ -1535,4 +1723,111 @@ export function shiftDomainConstraints(domains, offset) {
         return domains.slice();
     return domains.map(domain => ({ ...domain, column: domain.column + offset }));
 }
+/**
+ * Per-node cap on the number of INDs we materialize, mirroring
+ * `MAX_FDS_PER_NODE`. A safety valve for pathological plans; truncations are
+ * logged under the `quereus:planner:fd` logger like the FD/binding/domain caps.
+ */
+export const MAX_INDS_PER_NODE = 64;
+function sameOrderedList(a, b) {
+    if (a.length !== b.length)
+        return false;
+    for (let i = 0; i < a.length; i++)
+        if (a[i] !== b[i])
+            return false;
+    return true;
+}
+function indTargetsEqual(a, b) {
+    if (a.kind !== b.kind)
+        return false;
+    if (a.kind === 'table' && b.kind === 'table') {
+        return a.schema === b.schema && a.table === b.table && sameOrderedList(a.targetCols, b.targetCols);
+    }
+    if (a.kind === 'relation' && b.kind === 'relation') {
+        return a.relationId === b.relationId && sameOrderedList(a.targetCols, b.targetCols);
+    }
+    return false;
+}
+/**
+ * Structural equality for IND dedup. `cols` and `target.targetCols` are
+ * compared as *ordered* lists — the positional pairing between them is
+ * load-bearing (`cols[i]` is the child column matched to `targetCols[i]`), so a
+ * reordering is a different fact and must not collapse. Stricter than
+ * `fdsEqual`'s set comparison; over-keeping a reordered twin is harmless
+ * (redundancy, capped), whereas collapsing distinct facts would lose an IND.
+ */
+function indsEqual(a, b) {
+    return a.nullRejecting === b.nullRejecting
+        && sameOrderedList(a.cols, b.cols)
+        && indTargetsEqual(a.target, b.target);
+}
+/**
+ * Add a single IND, skipping it when a structurally-equal entry (per
+ * `indsEqual`) already exists. Enforces the per-node cap; truncations are
+ * logged. Mirrors `addFd` minus the determinant-subsumption logic (an IND has
+ * no determinant/dependent split to subsume).
+ */
+export function addInd(inds, next, opts = {}) {
+    const result = inds.slice();
+    if (!result.some(existing => indsEqual(existing, next))) {
+        result.push(next);
+    }
+    return enforceIndCap(result, opts);
+}
+function enforceIndCap(inds, opts) {
+    const cap = opts.cap ?? MAX_INDS_PER_NODE;
+    if (inds.length <= cap)
+        return inds;
+    const kept = inds.slice(0, cap);
+    log('IND cap reached: dropped %d IND(s) from %d', inds.length - kept.length, inds.length);
+    return kept;
+}
+/** Merge two IND lists: concat with structural dedup via `addInd`, capped. */
+export function mergeInds(a, b, opts = {}) {
+    let result = a.slice();
+    for (const ind of b) {
+        result = addInd(result, ind, opts);
+    }
+    return result;
+}
+/**
+ * Project INDs through a column mapping (oldCol → newCol). An IND's `cols` is
+ * **all-or-nothing**: drop the IND when ANY of its `cols` loses its mapping (the
+ * relation no longer carries the witnessing columns) — there is no partial-
+ * dependent survival as in `projectFds`. Survivors have their `cols` remapped to
+ * output indices; `target.targetCols` index into the *target* relation, NOT this
+ * relation's output, so they are NOT remapped. Result is deduped + capped.
+ */
+export function projectInds(inds, mapping) {
+    const out = [];
+    for (const ind of inds) {
+        const newCols = [];
+        let miss = false;
+        for (const c of ind.cols) {
+            const m = mapping.get(c);
+            if (m === undefined) {
+                miss = true;
+                break;
+            }
+            newCols.push(m);
+        }
+        if (miss)
+            continue;
+        out.push({ cols: newCols, target: ind.target, nullRejecting: ind.nullRejecting });
+    }
+    return mergeInds([], out);
+}
+/**
+ * Shift each IND's `cols` by `offset` (mirrors `shiftFds` for join column
+ * translation). `target.targetCols` are target-relative ⇒ NOT shifted.
+ */
+export function shiftInds(inds, offset) {
+    if (offset === 0)
+        return inds.slice();
+    return inds.map(ind => ({
+        cols: ind.cols.map(c => c + offset),
+        target: ind.target,
+        nullRejecting: ind.nullRejecting,
+    }));
+}
 //# sourceMappingURL=fd-utils.js.map