fix(drivers): undefined 比较数一律拒收 —— 闸落在任何发射器/守卫之前,两个毛病同闸消灭 (#6050) - #6142

… or guard reads it (#6050)

`undefined` in a comparand position had two answers in one driver, chosen by
the url a TursoDriver was constructed with. Measured on origin/main
(cba7454), four rows with `d` valued on 1-2 and NULL on 3-4:

  { d: undefined }                    LOCAL: knex "Undefined binding(s)"  REMOTE: ['3','4']
  { d: { $eq: undefined } }           LOCAL: knex "Undefined binding(s)"  REMOTE: ['3','4']
  { $not: { d: undefined } }          LOCAL: knex "Undefined binding(s)"  REMOTE: ['1','2']
  { d: { $ne: undefined } }           LOCAL: ['1','2']                    REMOTE: ['1','2']
  { $not: { d: { $ne: undefined } } } LOCAL: []                           REMOTE: ['3','4']
  { d: { $in: [undefined] } }         LOCAL: knex "Undefined binding(s)"  REMOTE: []
  { d: { $gt: undefined } }           LOCAL: knex "Undefined binding(s)"  REMOTE: []

Two defects. (A) The throwing rows carried no ADR-0112 envelope: knex's error
has neither `code` nor `status`, so mapDataError served an opaque 500 for a
caller mistake in a filter (#1116/#4436 catalogued the shape; this one was
missing from the list). (B) driver-sql's `$ne` emitter read `coerced == null`
while its own polarity tables (operatorIsNullTotal, nullValueSatisfiesOperator)
read `=== null`, so nullGuardForFieldSpec wrapped an already-TOTAL predicate in
`d IS NULL OR d IS NOT NULL` and its negation answered []. That is #5298's
invariant — a polarity table pins the spelling of its own emitter — broken at
its own definition.

Ruled REFUSED on 2026-08-07 (#6050 ruling B, the #5347-A disposition):
FieldOperatorsSchema declares no undefined comparand, `{ f: undefined }` is
indistinguishable from an absent key while meaning the opposite, and undefined
cannot survive JSON so it is always an in-process authoring bug — the
`{ owner_id: ctx.user?.id }` that silently matched every env-wide row.

One gate, placed before any emitter or guard, kills both: knex never sees an
undefined binding, and guard-vs-emitter disagreement about undefined becomes
unreachable rather than repaired.

- driver-sql: on the reduceFilterKey validation walk, beside the $null/$exists
  refusals, plus applyFilters' plain-map branch (which the walk cannot see —
  `{ d: undefined }` cannot make hasMongoOperators true). One function, two
  call sites.
- driver-turso: a whole-subtree pre-walk at buildWhereSQL entry, which is what
  makes `{ $not: { d: undefined } }` hold — compiling key by key would reach
  nullSafeNegationOperand with the undefined still in it.
- Both sides' `== null` / `|| === undefined` spellings tightened to `=== null`
  for #5347's reason: a lenient test silently resumes answering if the gate
  ever moves.

Positions enumerated, not swept: direct comparand, single-value operators, and
members of $in/$nin/$between arrays, at every depth. $null/$exists keep their
own boolean-domain refusal (#5240: one condition, one wording, both ways). The
two drivers' refusal sentences are verbatim identical.

null is untouched in every position: { f: null }, $eq/$ne: null, $null, and
null as an $in member all compile exactly as before, pinned by control blocks
in all three test files.

Callers measured, per the issue's precondition: full repo build (71 tasks) and
full repo test (134/135) pass, both example apps build, and the dogfood suite
boots the showcase stack — no production caller depends on the old behaviour.

Other no-value surfaces measured and NOT changed: formula reads undefined as
key-absence, read-scope-sql compiles `= NULL`, driver-memory reads it as null.
Three further readings, each a separate semantic ruling with its own envelope —
filed as #6125 rather than widened into this diff.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WyvqvKMG6asi9aXjKE6xtx