fix: = ANY(array) on an indexed column returns no rows

[earfman]

What

col = ANY(array) silently returned zero rows when col is indexed (primary key or explicit index). This fixes it and adds a regression test. Fixes #338.

This also resolves the mystery in that thread: the reporter's repro used a SERIAL PRIMARY KEY (indexed → broken path), while the unit test they added — and the name = any(...) case that worked — used non-indexed columns (seq-scan path, correct). The divergence was the index all along.

Why

buildComparison (src/transforms/build-filter.ts) optimizes col = <constant> into a scalar index equality. When the right operand is ANY(array), the constant is the whole array, so the generated EqFilter looks up an index key equal to ['A','C'] and matches nothing. Without an index, the seq-scan fallback evaluates ANY element-wise and returns the correct rows — hence the index-only divergence.

This is a common query shape (WHERE id = ANY($1) against a primary key is exactly what Sequelize, Mikro-ORM, etc. generate), and the failure is silent — an empty result set, no error.

The fix

Skip the scalar index optimization in buildComparison when either operand is an ANY/ALL expression, so it falls back to the seq-scan path that already evaluates array membership correctly:

// `col = ANY(array)` / `col > ANY(array)` (and ALL) must NOT be optimized into a
// scalar index lookup: the ANY/ALL operand is an array, so comparing an indexed
// column against the array *value* matches no rows (silently returning nothing).
// Fall back to a seq scan, which evaluates ANY/ALL element-wise. See buildBinaryAny.
if (leftValue.isAny || rightValue.isAny) {
    return null;
}

Tests

• Added src/tests/any-indexed-column.spec.ts — two cases (ARRAY[] and a string array literal) against a primary key column. Both fail on master and pass with this change.
• Full suite: 846 → 847 pass, with identical skip/fail/error counts (the 3 pre-existing failures are the ORM integration tests that require a live Postgres, unrelated to this change).
• The fix additionally repairs > ANY and != ANY on indexed columns, which returned empty/incorrect rows before.

Scope

Minimal — an 8-line guard in build-filter.ts plus a standalone regression test. No behavior change for non-indexed columns, IN (...), = ANY(subquery), or scalar comparisons.

---

This change was prepared with AI assistance and independently verified against the project's full test suite before submission (baseline-vs-fixed diff, and the regression test confirmed to fail without the fix).

[earfman]

Coretexa gate report — pgmem-any-idx

fix: = ANY(array) on an indexed column returns no rows (pg-mem)

status: closed · doer: doer-claude · verifier: verifier-pgmem · repo: oguimbal/pg-mem@master

Acceptance criteria

• col = ANY(array) on an indexed column (primary key / explicit index) matches the same rows as on a non-indexed column
• Existing behavior unchanged for =, IN (...), = ANY(subquery), and non-indexed columns
• pg-mem's full test suite passes with no new failures; a regression test is added that fails without the fix

Evidence

check	result	by	detail
repro	✓ pass	verifier-pgmem	indexed = ANY(array['A','C']) and = ANY('{A,C}') → [A,C]; PK, explicit index, and no-index all correct
regression-guard	✓ pass	verifier-pgmem	reverted the guard → new test FAILS (Received: []); restored → passes. Genuinely protective.
full-suite	✓ pass	verifier-pgmem	baseline 846 pass / 3 fail / 3 err → fixed 847 pass / 3 fail / 3 err (+1 pass, identical failures)
no-regression	✓ pass	verifier-pgmem	=, IN, = ANY(subquery), !=, scalar comparisons all still correct; 3 pre-existing fails are ORM tests needing a live Postgres
over-delivery	✓ pass	verifier-pgmem	fix also repairs > ANY and != ANY on indexed columns (returned empty/garbage before)

All required checks have passing evidence: repro, regression-guard, full-suite, no-regression

Verification

--- VERIFIED 2026-07-12 by verifier-pgmem (independent session, not the doer) ---
Gate 1: reproduced with own scripts — indexed = ANY(array/literal/subquery) all correct; non-broken paths intact.
Gate 2: ran full suite baseline (git stash) vs fixed — exactly +1 pass, same 3 fails / 3 errors (pre-existing ORM integration tests).
Gate 3: diff is an 8-line guard at the top of buildComparison, before the isConstant short-circuit and all index EqFilter/IneqFilter paths; comment accurate; isAny is a real IValue property.
Gate 4: reverted only the guard → the new regression test fails; also confirmed > ANY and != ANY on indexed columns are repaired; ALL is a pre-existing unsupported feature (identical both sides), not a gap.

---

Verified by a second agent · claimer ≠ verifier · coretexa.dev