drivers(memory, mongodb): the `$contains` family still folds case — the last two backends left on the wrong side of #4706 Q2 = A

[os-zhuang]

Finding, filed unassigned from the #6518 implementation. Not a request to unfreeze anything — it exists so the two FILTER_TEXT_CASES DEBT rows that survive #6518 have a successor to point at, instead of pointing at a closed issue.

What #6518 left behind

#4706 Q2 = A rules the $contains family ($contains / $notContains / $startsWith / $endsWith) case-SENSITIVE on every backend. After #6518 the scoreboard is:

face	{ name: { $contains: 'acme' } } against ACME Corp	verdict
driver-sql / -sqlite-wasm / -turso (local + remote)	does NOT match	✅ ruled behaviour (#6518)
@objectstack/formula matchesFilterCondition	does NOT match	✅ already compliant (String.includes)
objectql having	does NOT match	✅ already compliant
service-analytics (both SQL compilers)	does NOT match	✅ already compliant on the dialect they emit for
driver-memory — query path + analytics face	MATCHES	❌ new RegExp(escapeRegex(v), 'i')
driver-memory — reference matcher	does NOT match	⚠️ this package answers ONE operator two ways
driver-mongodb	MATCHES	❌ $regex with a hard-coded $options: 'i'

Both defects are over-matching — rows the filter excludes are returned — and both fold the whole Unicode range, so they also overshoot the #4706 Q1 = A ASCII-only boundary that $icontains is held to. On an RLS read scope a wider predicate is over-reach, not a loose filter (#3948).

The driver-memory row is the sharper one: its two faces disagree with each other, which is the divergence class #5374 closed between the other two faces of that package.

Why #6518 did not do it

The #5499 freeze (maintainer, 2026-08-05) suspends implementation investment in driver-memory and driver-mongodb. #6518 was dispatched with that freeze restated as a ruling, so both packages were explicitly out of scope; their FILTER_TEXT_CASES cells stay DEBT rather than being cleared or quietly widened.

Why this is separate from #6520

#6520 is the $icontains programme (requirement 1: implement a new operator, gated on taking $icontains into the spec FILTER_OPERATORS allowlist). This is requirement 2: an operator both packages already implement returns the wrong rows. Different requirement, different risk — #6520's cost is the vocabulary change rippling through allowlist consumers, while this one is a one-line flag on each side plus the row-set pins that flip with it.

They are, however, natural PR-mates: both touch the same two translateFieldOperators / normalizeFieldOperators sites, and clearing a FILTER_TEXT_CASES cell needs BOTH requirements answered before the case-set can be imported (scripts/check-driver-conformance.mjs judges a cell by that import, so a half-answered cell must not import it).

Suggested scope

1. driver-memory: drop the 'i' flag from the $contains-family lowering on the query path and the analytics face, and align the reference matcher — one answer per operator, per driver-memory analytics 面的 $notContains 编译成裸 mingo {$not: 'x'},该谓词不约束任何行 —— 结果被放大到全表 #5374.
2. driver-mongodb: drop the hard-coded $options: 'i' in mongodb-filter.ts's translateFieldOperators. escapeRegex already makes the comparand literal, so the literal-comparand cases keep holding.
3. Flip the row-set pins that currently encode the folded answer, asserting the NEW substance rather than deleting the assertion.
4. Re-grade the two FILTER_TEXT_CASES DEBT rows in scripts/check-driver-conformance.mjs — they can only be deleted once requirement 1 (JS 求值面全体拒收 $icontains(driver-memory 两面 / driver-mongodb / objectql having / formula）—— SQL 族已实现,同一 filter 在内存 double 上抛错 #6520) is answered too, since coverage is judged by importing the whole case-set.

Anchors verified on origin/main @ d13f627: driver-memory/src/memory-driver.ts (normalizeFieldOperators, the RegExp-'i' query path), driver-memory/src/memory-matcher.ts (match(), String.includes), driver-mongodb/src/mongodb-filter.ts (translateFieldOperators).

[os-zhuang]

Triage: pm:on-hold (routing domain:drivers already in place — kept).

Classification rationale. This is a real over-matching defect (both remaining backends fold case on the $contains family, against the #4706 Q2 = A ruling), but the fix lands entirely inside driver-memory and driver-mongodb, and the maintainer's 2026-08-05 #5499 investment freeze explicitly suspends implementation work in both packages — the card itself says it is not a request to unfreeze. That is a decision already made ("not now"), so the correct state is a hold, not pm:queue (undispatchable under the freeze) and not finding (a findings round could only re-discover the same freeze every pass).

Hold pair — date: 2026-08-08; reason: #5499 driver-memory / driver-mongodb investment freeze (restated as a ruling in the #6518 dispatch); restart condition: the maintainer lifts or narrows the #5499 freeze — at that point re-queue together with the #6520 requirement-1 decision, since a FILTER_TEXT_CASES cell can only be cleared when both requirements are answered.

Anchor verification (objectstack origin/main @ d6d1a50): driver-memory/src/memory-driver.ts:784-790,945 still builds new RegExp(escapeRegex(...), 'i') on the $contains family; driver-mongodb/src/mongodb-filter.ts:532-546 still hard-codes $options: 'i'; driver-memory/src/memory-matcher.ts:213-216 uses case-sensitive String.includes — the one-package-two-answers divergence described in the body is live.

Dedup: repo-scoped search finds no duplicate. #6520 is requirement 1 ($icontains vocabulary), objectui #6337 is the filter-UI half — both already cross-referenced in the body; distinct dispatch entries.

Not target:v17: both packages are under the #5499 freeze — the RC shipping without this fix is the standing ruling, not an oversight.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code