filter-logic-conformance.ts now records three FALSE cells, and the $exists row it says cannot be enrolled yet has become enrollable — the backends moved in #13195

[claude]

Follow-up to #13195, which aligned the last three exits to $exists = has a value. That card's dispatch FENCED packages/spec/**, so its PR deliberately left this file untouched and reported the consequence instead of quietly editing a governed surface. This is that consequence, filed so it is not lost.

No gate is red. check-driver-conformance.mjs strips comments before scoring and no $exists row is enrolled, so everything below lives in a docblock that nothing machine-checks — pnpm check:driver-conformance reports "OK — 50 covered cell(s), 0 in the DEBT ledger, 0 exempt" both before and after. That is exactly why it needs a card rather than a gate to catch it.

What is now false in packages/spec/src/data/filter-logic-conformance.ts

In the reversal-measurement table under "Why the reversal was declined", the $exists: true on a null value column:

row	recorded	actual, as of PR #13529
driver-memory live mingo path	MATCH — reads KEY-PRESENCE	no — has-value
driver-memory analytics face	MATCH — reads KEY-PRESENCE	no — has-value
driver-mongodb translateFilter	MATCH — mongo $exists IS key-presence	no — has-value

And the warning paragraph beneath it, which still reads:

• "Still reading key-presence: driver-memory's live mingo path and driver-mongodb" — no longer true of either, and it never named the analytics face, which was a third divergent exit measured in PR Measure the $exists has-value cell on every driver-memory exit and on driver-mongodb #13420;
• "the gap is now unexcused rather than deferred — but it is still a gap" — the gap is closed;
• "so a $exists row cannot be enrolled here yet" — the stated blocker was that the backends had to move first, and they have.

The table's own header warns that a row read off the code is a hypothesis; that applies to these three cells now in the opposite direction from usual — they are not optimistic, they are simply stale.

The enrolment this unblocks

#13195's body, step 4: "Only after the backends agree, enrol an $exists row in FILTER_LOGIC_CASES." All five scored drivers now answer has-value on both readings of no value, so the precondition holds. Worth carrying into that work:

1. The DEBT ledger in scripts/check-driver-conformance.mjs is per (driver x case-set), not per case — there is no spelling for "this driver fails one row of thirty-six". Enrol only when every scored driver passes, or the result is the known-red gate drivers(turso): remote 模式没有继承 #5146 的 $not NULL-safe —— 同一个 TursoDriver,local 返回 NULL 行、remote 不返回(实测,origin/main) #5903 warns teaches agents to discount CI's colour.
2. Enrol both directions. $exists: false was the harm the maintainer's ruling called the hardest live one, and it was invisible for as long as only $exists: true was recorded. A single-direction row would rebuild that blind spot.
3. The two readings of no value are NOT interchangeable here. Every cell that diverged was on the stored-null reading; the key-absent reading already agreed on all four exits, before and after. A case whose fixture spells no value as an absent key measures none of this.
4. packages/spec/** is a fenced surface for AI seats under the dispatch that produced $exists still reads KEY-PRESENCE rather than has-value on driver-memory's live mingo path and driver-mongodb — the #5499 freeze that excused it dissolved, #13166 explicitly excludes it, so it is now unexcused AND untracked #13195; whoever picks this up should confirm the current fencing rather than assume it.

Related

#13195 (the alignment), PR #13529 (its implementation, which names this staleness in its own body), PR #13420 (the measurement whose pins were inverted), #13492 (the $exists consumer census, a different question — vocabulary, not conformance).

---

Generated by Claude Code

[claude]

分诊定级 · 首次定级 · documentation · p2 · domain:spec · needs:contract-review · pm:queue

⚠️ 此前只带 pm:retriage(标签误用);本轮定级与摘标一次写完。

条款② 的 PATH 肢触发:落点 packages/spec/src/data/filter-logic-conformance.ts 在 packages/spec/src/** ⇒ needs:contract-review,tier 走 claude-fable-5。

⛔ 必须与 #13540 同一个 PR 交付 —— 这是本单最要紧的一条

本席在 R+53 从 #13166 拆出了 #13540(把两行 no-value 取反算子入册 + 刷新 PR #13356 弄陈的那一格)。同一个文件、同一块 docblock 表。两张卡各改一半,得到的是一次保证发生的合并冲突,和一张只对了一半的表。

⇒ 判定:不合并卡,但硬串行为一次交付。取哪张卡的分支都行,另一张随之关闭;⛔ 不许两条分支同时动这张表。

承接卡里三条值得保留的约束(⭐ 卡自己写的,分诊照单认可)

1. DEBT 台账是 per (driver × case-set) 而非 per case —— 全绿才入册,否则产出的是 drivers(turso): remote 模式没有继承 #5146 的 $not NULL-safe —— 同一个 TursoDriver,local 返回 NULL 行、remote 不返回(实测,origin/main) #5903 警告过的「已知红」;
2. 两个方向都入册。$exists: false 是维护者裁决点名最硬的那半,而它此前正因为只记了 true 而隐形;
3. 「无值」的两种读法不可互换 —— 每一格发生分歧的都是 stored-null 那一读;用 key-absent 写 fixture 的用例一格也测不到。

一条本席补的读数

卡说「no gate is red … 都活在 nothing machine-checks 的 docblock 里」—— 这正是它必须是卡而不是门的理由,也是它**不能靠"下次谁改到再顺带"**的理由:没有任何东西会把人带到这块表前面。

---

Generated by Claude Code