docs(pm-dispatch): six measured platform-reading rows — search zeroing, labels AND/OR, the MCP read-path trap, landing by content

[claude]

Fixes#11193
Fixes#11271

Six measured platform facts land as one-line-per-entry rows in
.claude/skills/pm-dispatch/references/platform-readings.md. Dispatched as a fold of
two graded cards, one PR, one commit per member.

Member commits

commit	card	rows
39f6a6a1	#11193	search_issues session-level zeroing + control-term probe; REST list endpoints as the correct fallback (never MCP list_issues hand-scans); the labels AND/OR asymmetry; the qualifier-syntax candidate recorded as REFUTED
7fcab57f	#11271	the MCP read path eating angle-bracket fragments inside inline code, with its discriminator; landing judged by content on origin/main

Folded into existing entries, not appended

Per the grading note, the file was read in full first. Four of the six facts had an entry
that already owned their subject, so they were folded in rather than duplicated:

• The labels row already carried the OR half (MCP list_issues). This adds the REST
  labels=a,b AND counterpart and the two independent same-day measurements.
• The issue_read row claimed read-side artifacts were uniformly reversible. The new
  fact contradicts that in a way worth stating: a dropped fragment cannot be decoded back,
  so that row gets the correction plus a pointer to the full rule.
• The truncation-judgment row already described the API-reads-short direction. The new leg
  is the same trap with the read paths swapped, so it extends that entry.
• The landing-check row already required two readings; it now says the origin/main
  reading is a content reading, and why the two tempting shortcuts are both wrong.

Only one genuinely new entry was needed (the search-zeroing operating procedure).

How the ceiling was paid

The file sits at its line-ratchet ceiling with zero headroom (134/134). Payment follows
the remedy guidance in check-skill-line-ratchet.mjs's own header — compress in place,
never raise a ceiling (raising needs a maintainer ruling, and none was sought or given).
Each commit adds exactly one physical line and pays for it by re-flowing one loosely
wrapped bullet onto a single line, no words dropped; both re-flowed bullets land well
inside the file's established line-length convention. The ceiling constant is untouched and
the file is 134 lines at both commits, so the ratchet is green at each one, not just at
the tip.

Gate verdict at the final head 7fcab57f, quoted as the gate printed it:

check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md is 134 lines (ceiling 134; headroom 0).

⛔ #11106 remains open and is deliberately not addressed here. The line-vs-byte metering
question is in the decision inbox; this PR pays purely in lines per the current line-ratchet
remedy and neither resolves nor preempts it.

Governed surface

.claude/** is a governed surface: this PR opens as draft, is not to be flipped
ready by an agent, carries no auto-merge, and is for human merge only.
skip-changeset applies by convention — the diff is one instruction-surface reference file
and publishes nothing.

Verification

Gate families derived mechanically from the actual change set with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (re-derived after the
final commit; same six families, no new members). All seven run green at 7fcab57f:

gate	verdict
check:pm-skill-ratchet	134 lines, ceiling 134, headroom 0
check:pm-skill-id-lint	17 file(s) clean (pattern for issue-number references)
check:doc-authoring	389 files clean — no bare metadata literals
check:skill-frame-sync	4 copies of the decision frame structurally isomorphic across 3 files
check:doc-formula-expressions	22 record-scoped formula example(s) across 416 files / 1447 TS blocks clean
check:pm-governed-merges	self-test, 119 assertions
check:nul-bytes	6332 text file(s), no raw ASCII control bytes

The new rows carry no issue numbers, per the file's own header rule and check:pm-skill-id-lint.

One fact verified first-hand while writing this PR

Fact 5 was confirmed in this session before it was written down, on this very fold: #11271's
body read through MCP showed an empty inline-code span and an identifier written without
angle brackets, while curl against api.github.com returned the same body carrying five
angle-bracket-bearing inline spans intact. The card was not damaged and was not
"repaired" — which is exactly the near-miss the row exists to prevent.

Declared narrowings

• The @objectstack/formula dependency closure was built with dts generation skipped. That
  gate loads formula at runtime and consumes no .d.ts, and this diff contains no
  TypeScript, so no type judgment rests on it.
• Repo-wide pnpm lint was not run; CI owns that run. Nothing here is a narrowing of a
  named family — all six derived families plus check:nul-bytes ran in full.

Reported at draft-PR time per the dispatch contract; CI convergence is the PM's to read.

Generated by Claude Code

---

Generated by Claude Code

[github-actions]

⛔ merge queue 构建失败 — 先分诊，再决定要不要重排

队列构建 32629104199 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集），
所以失败的测试可能在本 PR 没碰过的包里 —— 那不是重排能修的。每次盲目重排都会让排在后面的所有 PR 重建一轮。

失败的 job（日志抽取，best effort）：

• Test Core (5/6) — 失败步骤: Verify pnpm version

  AssertionError [ERR_ASSERTION]: The expression evaluated to a falsy value:
  

↳ 失败原因 是判读的关键：超时（Test timed out in … / Hook timed out in …）多半是负载/时序，不是本 PR 的回归；
断言（AssertionError: …）才指向真实的行为改变。两者的 FAIL 行长得一模一样，只有这一行能区分。

跨 PR 相同签名（24h，按失败测试文件聚合）：

• ⚠️本次没有可用的聚合签名（日志里没有能解析出测试文件名的 FAIL 行）—— 这不是「没有同签名的其他 PR」，是这一轮没测到。跨 PR 聚合本次不可用，请手工比对其他 PR 的同类评论。
• ⚠️ 24h 评论账本没读完（超过 5 页仍未读到窗口尽头），所以上面的「不同 PR 数」是下界，不是全量。

历史信号：

• 本 PR 过去 24h 无队列失败记录（首次）。
• 过去 24h 队列共有 71 个失败构建（不含本次）。

分诊清单：

1. 失败测试在本 PR 改动的包里 → 真回归，修 PR。
2. 失败测试与本 PR 无关 → 看上面的「跨 PR 相同签名」；已有汇总 issue ⇒ flaky/环境问题实锤，去那张 issue 上谈，修好前重排只会再烧一轮全队列。
3. 两者都不是 → 可能与同组 PR 语义冲突；等前面的 PR 落地或失败出队后再重排一次即可，不要连续重排。

---

Generated by Claude Code · merge-queue-triage workflow (#4859)

[os-zhuang]

Queue-kick triage (skills seat, session session_01RMTpSRF5CjMmQBFfPtPCwJ): re-queuing once, as-is. Basis from the triage comment above: the failing step is Verify pnpm version in Test Core (5/6) — a setup-step assertion that died before any test body ran, which is the one class the re-run rule sanctions; and this PR's diff is a single .claude/** markdown reference file with no causal path into Test Core (checklist branch 2, no aggregated cross-PR signature available this round — recorded as "not measured", not "no siblings"). Not touching the signature ledger (a suspected new infra flake gets a mention only — promotion to the ledger is a human's). If the SAME signature kicks this PR again, that second failure is treated as real per the rules: no further re-queues, re-diagnosis instead.

---

Generated by Claude Code