Found while landing the rest-channel.md header correction (issue 13008), which asked in
its own body whether any other file quotes the read order and drifted the same way. It
does. Filed unassigned and not fixed in that PR — see "Why it was not folded in" below.
The drift
.claude/skills/pm-dispatch/references/review-checklist.md line 33 carries:
- **报告 `tests`/证据里有 git 本可回答的 API 读吗?**(dev 契约的读序是 git → REST →
MCP/GraphQL)判据:MCP `list_issues`/`search_issues` 一类 GraphQL 读,或重跑派发词已下发的去重读数
Three tiers, and the zero-quota payload tier is missing — the same omission, in the same
direction, as the header just corrected in rest-channel.md.
Why it was not folded into that PR
The bounded in-place fix exemption in the dev contract needs all four of its conditions.
Condition 2 — the correct form is pinned by existing evidence — does not hold here,
and that is the whole reason this is a separate card rather than a one-word insertion:
rest-channel.md's header names its source (platform-readings.md) and quotes that
file's policy line, so the correction was a mechanical copy of a spelling that already
exists on disk.- This line names a different source: 「dev 契约的读序」 — the dev contract, which is
.claude/agents/os-dev.md. That file does not state a compact ordered ladder in the
git → X → Y spelling at all; it states the discipline as prose (probe repo-scoped REST
first, take the zero-quota payload channel for bulk reads, keep MCP for writes plus the
one dedupe search). So what the corrected spelling should be is a judgement about how to
summarise that prose as a ladder, not a copy of an existing one.
Deciding that inside a PR claimed for one file, on a second zero-headroom governed surface,
is exactly the unreviewable scope creep the exemption exists to prevent.
Constraint on whoever takes it
review-checklist.md is governed and ratcheted at 84 lines, ceiling 84, headroom 0 —
so net 0 lines.- Good news on the byte budget, measured rather than assumed: line 33 is 105 bytes
against the 120-byte cap. The insertion that the sibling card needed costs 12 bytes,
which would land it at 117 — under the cap, so unlike the sibling this one very likely
needs no re-wrap at all. Re-measure before writing; the arithmetic above is this
seat's, not a gate's.
- The real work is the judgement in the section above, not the edit.
The general gap, worth its own consideration
Nothing mechanically holds a quoted policy equal to its source across these files. Both
sites were found by reading, not by a gate, and the second one only because the first card's
body thought to ask. A third site can appear the same way and stay wrong indefinitely. There
may be a cheap gate here — a pinned pairing of "this file quotes that file's ladder" — but
that is a design question, not part of this fix.
Generated by Claude Code
Found while landing the
rest-channel.mdheader correction (issue 13008), which asked inits own body whether any other file quotes the read order and drifted the same way. It
does. Filed unassigned and not fixed in that PR — see "Why it was not folded in" below.
The drift
.claude/skills/pm-dispatch/references/review-checklist.mdline 33 carries:Three tiers, and the zero-quota payload tier is missing — the same omission, in the same
direction, as the header just corrected in
rest-channel.md.Why it was not folded into that PR
The bounded in-place fix exemption in the dev contract needs all four of its conditions.
Condition 2 — the correct form is pinned by existing evidence — does not hold here,
and that is the whole reason this is a separate card rather than a one-word insertion:
rest-channel.md's header names its source (platform-readings.md) and quotes thatfile's policy line, so the correction was a mechanical copy of a spelling that already
exists on disk.
.claude/agents/os-dev.md. That file does not state a compact ordered ladder in thegit → X → Yspelling at all; it states the discipline as prose (probe repo-scoped RESTfirst, take the zero-quota payload channel for bulk reads, keep MCP for writes plus the
one dedupe search). So what the corrected spelling should be is a judgement about how to
summarise that prose as a ladder, not a copy of an existing one.
Deciding that inside a PR claimed for one file, on a second zero-headroom governed surface,
is exactly the unreviewable scope creep the exemption exists to prevent.
Constraint on whoever takes it
review-checklist.mdis governed and ratcheted at 84 lines, ceiling 84, headroom 0 —so net 0 lines.
against the 120-byte cap. The insertion that the sibling card needed costs 12 bytes,
which would land it at 117 — under the cap, so unlike the sibling this one very likely
needs no re-wrap at all. Re-measure before writing; the arithmetic above is this
seat's, not a gate's.
The general gap, worth its own consideration
Nothing mechanically holds a quoted policy equal to its source across these files. Both
sites were found by reading, not by a gate, and the second one only because the first card's
body thought to ask. A third site can appear the same way and stay wrong indefinitely. There
may be a cheap gate here — a pinned pairing of "this file quotes that file's ladder" — but
that is a design question, not part of this fix.
Generated by Claude Code