refs(pm-dispatch): bank five measured platform readings, paid for by same-file compression

[os-litant]

Fixes#12486
Fixes#12534
Fixes#12604
Fixes#12628
Fixes#12640

Five measured platform facts become fact-table rows in
.claude/skills/pm-dispatch/references/platform-readings.md. That file is the only
surface this PR touches.

Zero-headroom accounting. The line ratchet pins the file at 314 and it was sitting at
exactly 314, so every added line is paid for by compression inside the same file. No
row was re-wrapped to squeeze a line out — each cut removes content, and the ledger below
names where that content survives. Final state: 314/314, headroom 0; widest table row 0
bytes (pin 0).

Governed surface (.claude/**, Markdown): stays draft for a maintainer merge. No
changeset — this publishes nothing (skip-changeset).

---

The five rows

1 · A free-text negation is inert in search_pull_requests

API 配额 section, new row, 5 lines.

A leading - on a quoted phrase is silently dropped: -"Part of" and "Part of" return
the same total_count (197) and the same top hit, and 197 + 197 does not reconcile against
the unfiltered 216 — so the two are not a partition of anything. The qualifier half of the
same query keeps exact arithmetic (linked:issue 2903 + -linked:issue 216 = 3119). One
call, correct arithmetic from the qualifiers, silent nonsense from the free text. The row
carries the discipline that follows: any stratification built on a free-text term must run
its complement and sum both against the unfiltered population; if they do not reconcile,
the text half is not filtering.

Placed after the two search_issues reliability rows, so the search instruments read
together.

2 · The reachable positive enqueue instrument, and auto_merge as unstable

队列成员资格与 auto-merge section, existing timeline row 4 → 8 lines.

The table named what cannot tell you a PR is enqueued and pointed at a webhook for the
positive answer — leaving a seat with a blind instrument and a ⛔ rule that depends on the
reading it cannot make, which is how an already-enqueued PR got re-armed.

Per the filing seat's own 2026-08-26 correction, the webhook is reachable from a
subscribed session (observed within ~1s of the timeline event), so the row states the
division rather than a gap:

• timeline (added_to_merge_queue / removed_from_merge_queue) is the reading you
  make — queryable on demand, both directions from one call, and it separates "never
  enqueued" from "enqueued then ejected";
• the pull_request.enqueued webhook is push-only corroboration a fresh context or an
  unsubscribed seat cannot recover;
• auto_merge is upgraded from absent to UNSTABLE — the same PR read set and then
  None within a minute, so "not armed" is never a conclusion from it.

This sits beside the just-landed mergeable_state / update_pull_request /
draft-dequeue reshaping. None of that chain's cuts is reopened.

3 · The session proxy's /search 403 parses as a clean zero

API 配额 section, new row, 5 lines.

Unauthenticated REST from inside a session reaches repository-scoped paths only; anything
under /search/ is refused. The refusal body is valid JSON, so a caller reading
total_count off it gets None and prints total: None — one character from a real empty
result set. Three duplicate-check queries produced three clean-looking zeros and never ran.
The instrument for /search questions is the MCP search tools; repo-scoped enumeration is
unaffected.

Placed directly under the session-zeroing row so that row's 见下条 pointer resolves, and
carrying the pairing the card asked for: run the positive control FIRST, because a zero
that matches your expectations is the one you never re-check.

4 · Tier-quota termination, and re-deriving the mandate per file surface

读数陷阱 section, two new rows, 8 lines.

A subagent that terminates mid-edit because the account's model tier hit its limit is a
third death shape the table had no name for: the host reports the limit, so it is
neither an ordinary subagent death nor a maintainer abort — the dispatch is dead, the
premise is intact, and re-dispatch is blocked on a resource. Handling is dead-claim
recovery plus worktree rescue, not premise re-verification and not an escalation. No
pre-flight probe exists, so the mitigation is dev-side WIP-committing.

The second row is what keeps a lane moving. A mandated tier must not be downgraded because
it is unavailable — but "this lane is fable-mandated" is usually recalled, and the
recalled version is too wide. Re-deriving with dispatch-gates --tierper file surface
splits a blocked fold instead of stopping it (measured: SKILL.md and .claude/agents/**
are mandated; scripts/pm/** and references/** carry no path mandate — a fully-blocked
three-card fold became one blocked half and three runnable ones). The path line is a
floor, never a clearance.

5 · issue_read get_labels cannot resolve a PR number

API 配额 section, new row, 3 lines.

get_labels resolves the GraphQL Issue type only, so a pull request number errors with
"Could not resolve to an Issue" — the REST convention that a PR is also an issue does not
carry over. Measured twice, on different PRs, by different dev runs in the same hour. Label
read-backs on PRs are a mandated step, so every dev walks this path; the row names the
working read (pull_request_read get, whose response carries the labels array) and rules
out spending a search-channel call on a single-PR question.

Placed with the other label-semantics traps so the read-back rule and its instrument sit
together.

---

Cut ledger — every cut, and where its content survives

row cut	lines	what was removed, and its surviving home
search_issues session-zeroing	12 → 10	不是工具/平台级 (survives as 故障是会话级) · the perPage pointer (its own dedicated row above) · 极易半途而废 (survives as the 226/100 specimen) · the undetermined-mechanism tail (survives as ⛔ 已推翻别再追)
读数五坑 gh clause	7 → 6	the command -v gh exit-1 evidence (survives as 容器里没有 gh) · the pipe-exit-code cross-reference (survives in os-dev.md's gate-reading clause)
auto-tier classifier	5 → 4	the label-POST anecdote, compressed to 判定还跨天翻面 — the cross-day fact is kept, the story is not
quota-per-account	7 → 6	端点自身不计费 (survives at the 开轮先读配额 clause's 免费) · the used-10461/core-7 specimen (survives as the independent-pools reading in the 默认读序 row)
MCP read-side entity	7 → 5	the write-side shape enumeration (survives in the 写侧实测行为 rows it pointed at) · the MCP-or-API attribution aside; the & boundary is kept
REST-availability	12 → 11	shift matrix trimmed to the two readings that prove 门态逐容器变; the per-entry details survive at the credential-discriminator clause and the gh clause
probe-branch	8 → 7	the pre-claim-check restatement (survives as the named failure direction) · the undiagnosed-cause enumeration; the boundary itself is kept
truncation diagnosis	12 → 10	the WebFetch restatement (merged into the ⛔ clause) · the three-seat anecdote · the 0-span/10-tag control example; the 判据 is verbatim
sanitizer author rule	7 → 6	the comment-exemption history, compressed to ⛔ 评论不豁免 (mechanism survives in the comment row below)
sanitizer issue body	9 → 7	two fenced-behaviour readings merged into one dated clause · the negative-lookbehind aside · 同类标记在评论里存活 (rule survives as ⛔ 评论不豁免)
sanitizer comment	8 → 6	the second-order anecdote · the safe-authoring recipe (verbatim in the author-rule row) · the 人看着像正常收尾 restatement; shape, fuse and worst-case boundary all kept
silent-downgrade pair	9 → 8	the host-fork row was entirely a modifier on the row above, so it becomes that row's ⚠️ clause; both host readings, both dates and the transcript fallback survive
blocked-download	3 → 2	the unverified-verdict anecdote · the zero-hit cross-reference; os-dev.md carries the same rule with the same evidence
enable-sequence	8 → 7	the stale-PR-reading specimen, whose fact (contradictory merged booleans) is the 二义读数 row's own subject
default-read-order	8 → 7	one of two specimens for independent metering
git-first	11 → 10	the container-has-no-gh duplicate (the 读数五坑 row asserts it) · the search/core independence duplicate (the row directly above states it)
MCP parameters	10 → 9	the weaker of two OR-side specimens; the sharper one (an "intersection" wider than one of its own inputs) is kept
list_issues assignees	3 → 2	the trailing ⛔ restatement, which repeats 清单只是候选名单 in imperative form
MCP list/search	3 → 2	the navigational pointer to the session-gate row printed directly above it
org-authorization	3 → 2	the valid-JSON-403 mechanism now lives in row 3 above, so this row cites it instead of restating it

Net: +25 added / −25 paid, landing back at 314. Every figure in this table is read
back from git diff -U0 d79c6025f..HEAD hunk headers, not from notes taken while editing —
that re-derivation caught one: the first commit's message records the gh clause as
7 -> 5; it is 7 -> 6, as above. The commit is already pushed and this branch is
never force-pushed, so the correction lives here and in the dev report rather than in
rewritten history. The totals above are the diff's own.

---

Verification

Derived union, not recalled: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack
with no paths, re-run after git fetch so the stale-tree warning cleared. It reads the
change set from the merge base itself — exactly one path, this file. Eight families.
All runs below are at 887592d04, the branch head, which is also the merge commit that
brings origin/main in.

Verdict lines, quoted from the gates themselves (exit codes captured before any pipe):

✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md is 314 lines (ceiling 314; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md: widest table row is 0 bytes (pin 0; headroom 0).
✓ check-skill-line-ratchet self-test: 111 cases pass.
✓ check-skill-id-lint: 22 file(s) clean (pattern /#[0-9]{3,}/g).
✓ check-skill-id-lint self-test: 14 cases pass.
✓ check-agent-test-spelling: 0 violations — 386 file(s) · 4166 bare `--` token(s) · 1190 launcher-rooted run(s) · 9 separator(s) JUDGED
✓ doc authoring guard: 390 files clean — no bare metadata literals.
✓ doc authoring guard: 48 published skill files clean — no internal issue-id references.
✓ check-governed-merges --self-test: 193 assertions
✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 422 files / 1450 TS blocks judged clean by @objectstack/formula.
✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
✓ check-pm-governed-merges (self-test above; the sweep itself exited 0 in the same run)

check:doc-formula-expressions first answered PREREQUISITE NOT MET (@objectstack/formula
then @objectstack/lint not built) — read as NOT MEASURED, not as red; both were built and
the gate then produced the verdict line above.

One family is not measurable here and CI owns it:scripts/pm/check-governed-queue-guard.mjs
reads GITHUB_EVENT_PATH and refuses without a workflow event payload
("could not look" must never exit 0 here). That is a declared CI-only input, not a
finding.

Declared narrowing: the repo-wide pnpm lint sweep was not run locally; the derived
union above is the local scope, and CI runs the whole farm exactly once regardless.

Generated by Claude Code

---

Generated by Claude Code