fix(pm): a size refusal is a refusal — post-stamped stops reporting a real 422 as a missing route

[os-elon-musk]

Fixes #18843

Clause-②: no

post-stamped.mjs reported a real HTTP 422 from the platform as exit 3
EXIT_PREREQUISITE_NOT_MET — 「no route, no act at all」 — and prescribed a
transport remedy. This gives the size refusal an exit class of its own, whose
text carries the bytes SENT, the measured cap for THAT surface, and the remedy
that works; and it homes the comment cap beside ISSUE_BODY_LIMIT, which the
card says is the same card.

The fixture, verbatim, and where it came from

Built from the record on probe objectstack#18826 — comment id 5722505167,
written 2026-09-17T23:19:18Z. ⛔ No new oversized write was made anywhere to
reproduce this. Captured there from POST /issues/{n}/comments at 262,145 bytes:

{"resource":"IssueComment","code":"unprocessable","field":"data",
 "message":"Body is too long (maximum is 65536 characters)"}

The two readings the dispatch carried as unmeasured, now measured

1. The 422 message is NOT the same string on the two endpoints. The record
   is explicit: the create side answers Body is too long (maximum is 65536 characters), and PATCH /issues/comments/{id} answers the same sentence with
   a lower-case body; POST /issues says it twice. So the trigger is
   case-insensitive rather than an equality — an equality pinned to the
   create-side capital would have left the update side at exit 3. Pinned in both
   spellings, and the ablation below proves the flag is load-bearing.
   ⚠️ One correction to the dispatch's framing: PATCH /issues/comments/{id} is
   not a surface this tool ever writes. Its two write surfaces are
   POST /issues/{n}/comments (--comment) and PATCH /issues/{n} (--body).
   Both caps are declared; the update-side spelling still matters, because it is
   the second measured spelling of the sentence the trigger keys on.
2. The two caps live as two constants, beside each other in
   check-half-states.mjs. COMMENT_BODY_LIMIT is declared after
   MARKDOWN_BODY_BUDGET, with its own bisection provenance. It is homed there
   rather than in post-stamped.mjs on a measured ground: five workflow comments
   (cross-repo-issue-closer.yml, merge-queue-triage.yml,
   scheduled-full-run-card.yml, test-nightly-tiers.yml) already send a reader
   to ISSUE_BODY_LIMIT in that file for the number and warn against
   re-deriving 65,536 — the folklore grows back wherever one of two caps is kept
   somewhere else. ⛔ Two constants and not one shared number: two surfaces
   measured independently that agree is not one measurement, and the day the
   platform moves one a shared number would lie about the other.

The exit register, before and after

code	before	after
0	written, and everything sent is on the platform	unchanged
1	usage	unchanged
2	the body broke the stamp contract	unchanged
3	PREREQUISITE NOT MET — no route, no token, no act at all	unchanged, and it now prints what the platform said
4	written, and the platform did NOT store it	unchanged
5	—	TOO LARGE — the platform ANSWERED and refused the body for its size. Nothing written; shorten or split it. ⛔ Not a route problem.

EXIT_TOO_LARGE = 5 is declared in post-stamped.mjs beside EXIT_NOT_STORED.
The usage text, the not-stored text and the prerequisite text each name it, so a
reader of a 3 or a 4 learns the fifth code exists.

The reader test

A seat that sends a 300 KB comment body now reads: exit 5, 307,200 byte(s)
sent, cap 262,144 bytes for a comment, 45,056 byte(s) over it, shorten or split
the body — and ⛔ not a word about routes or tokens. Verbatim from the
--self-test fixture:

post-stamped: TOO LARGE — the platform ANSWERED this write and REFUSED it: the body is over the cap
for this surface. NOTHING WAS WRITTEN.

  Sent to objectstack-ai/objectstack#18843: 307,200 byte(s) as a comment (`POST /issues/{n}/comments`).
  The measured cap for a comment is 262,144 bytes — UTF-8 BYTES, bisected on
  objectstack#18826, one byte either side.
  This body is 45,056 byte(s) over it.

  The platform said: Validation Failed · Body is too long (maximum is 65536 characters)
  ⛔ Its number is FALSE in unit and value — the same fleet measured a 262,144-BYTE comment stored,
     four times what that sentence claims as a maximum.

What the class deliberately does NOT take

• Not every 422. The issues listing refuses deep pagination — page 99 stores,
  page 100 answers 422 (measured 2026-08-31, recorded in
  check-half-states.mjs) — and that refusal is about a cursor. It keeps its 3,
  pinned as a control, as do the same sentence on a 403, a 500 and a 200.
• Not the platform's number. The trigger stops before maximum is, so
  keying on a false clause cannot un-classify the refusal the day GitHub
  corrects its own text; a text carrying only that clause is pinned NOT to
  classify.
• Not the read path. A read sends no body, so the --body pre-read keeps its
  PREREQUISITE reading unchanged.
• Not a widened --body story. Over the cap, PATCH /issues/{n} refuses
  SILENTLY (200, old body kept, [probe] #18664 — bisecting GitHub's real issue-body refusal boundary, in BYTES #18793), which is exit 4's population; only a
  422 the platform actually answers reaches 5.

A refusal the cap does not explain

If the platform ever refuses a body at or under the measured cap, the text says
the refusal CONTRADICTS the reading and asks for a re-measurement rather than
printing an overage of zero — over is null, never 0. The ordinary over-cap
text is pinned NOT to carry that sentence.

Self-test counts

battery ledger	before	after
post-stamped --self-test cases	349	402
post-stamped batteries (floor)	13	14
new battery the size refusal: … (registered/floor)	—	53 / 53
check-half-states --self-test cases	4,963	4,978
battery ISSUE_BODY_LIMIT measured cap (registered/floor)	40 / 37	55 / 52

The check-half-states "before" is the merge base d5ccfb4e3: 4,978 minus the 15
cases this PR registers. Two independent readings agree on it — the dispatch
recorded 4,963 at PR #18911's head, and this branch measures 4,921 at its own
branch point 625db0e85, which is before #18911 landed. Both floors RISE by
exactly the number of cases added; neither is held flat.

Ablations — each mutation proven on disk, each restore proven against HEAD

All four through scripts/ablation-replace.mjs, which asserts the anchor count
fell, the blob hash changed, and after restore that the blob equals HEAD's and
git diff HEAD is empty.

mutation	expected	observed
drop the i flag from BODY_TOO_LONG_RE	red	19 of 402 cases fail by name; the lower-case pin stays green, so the ablation discriminates
put ROUTE_REMEDY in the size text's Fix: line	red	exactly the 2 held-apart pins fail
COMMENT_BODY_LIMIT 262144 → 65536	red	5 of 4,978 cases fail by name
delete UTF-8 BYTES from the comment cap's docblock	red	exactly 1 pin fails — and the sibling docblock also contains that phrase, so a mis-pointed extractor would have stayed green

⭐ The first ablation found a defect in this PR's own first draft. It came
back GREEN: the classifier was rebuilding the regex from .source with the
flags retyped at the call site, so the exported constant documented a rule it
did not carry and every case-insensitivity pin was vacuous. Fixed in
36ba41a64, after which the same ablation turns 19 cases red. A second pass
(290acfad0) made those pins fail by NAME rather than crash on a TypeError
before the verdict machinery ran.

Gates

Derived in the worktree with
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack
at HEAD 0e5e054d3: 39 families, all 39 run, all exit 0, exit codes captured
by redirect-then-$?. --ran reconciliation: 39 derived, 39 run, 0 UNRUN. The
first derivation (before the merge, one path) gave 29; adding
check-half-states.mjs grew it by 10 and those were run too. Each edited
script's own --self-test is in that set
(check:pm-post-stamped, check:pm-half-states). git grep of both filenames
under tests finds no separate suite — the --self-test IS the suite.

ESLint was not narrowed at all. The repo-wide run is the one that was taken:
pnpm lint (= eslint . --no-inline-config, this repo's only style authority)
exit 0 at HEAD 0e5e054d3. The targeted run over the two changed files is
there too — 2 files, 0 errors, 0 warnings — but it is not what this PR rests on.

Acceptance notes

• Exit 3 still hosts other refusals the platform ANSWERED. The deep
  pagination 422 is the measured one, and its remedy text is a route remedy just
  as wrongly. ⛔ Not filed: it is unreachable through this tool (a comment list
  would need 10,000 comments to reach page 100), so it is an observation, not a
  reproducible defect.
• rest() now surfaces the platform's sentence on every non-2xx, which is
  the same defect family (the tool discarding the answer it was given) and is
  mechanically part of this change; it is not a separate fix.

---

Generated by Claude Code

[os-elon-musk]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 0e5e054d36b0547ccfd5d35d68b7b9739fbd8c8d

① Derived judgments

• Surface: scripts/pm/post-stamped.mjs (+447 / −9) and scripts/pm/check-half-states.mjs (+67 / −1); five commits (the class, the regex fix the ablation caught, the fail-by-name pass, one origin/main merge, the constant). Off the governed register (check-governed-merges.mjs --test: NOT governed); no published package moves; no contract's accept/reject set moves. The PR feat(pm): name the two measured claim near-miss forms so the sweep sounds on the unreadable claims #18911 commit (9364e1b01) is an ancestor of the head — the serial on check-half-states.mjs held.
• The class: EXIT_TOO_LARGE = 5, triggered by SIZE_REFUSAL_STATUS (422) AND BODY_TOO_LONG_RE = /body is too long/i on the platform's text — both halves required, so the deep-pagination 422 keeps its 3 (the header names it). The text carries the bytes SENT (Buffer.byteLength(rendered.body, 'utf8')), the cap of the surface written to (WRITE_SURFACES: comment → COMMENT_BODY_LIMIT, body → ISSUE_BODY_LIMIT), the overage and SIZE_REMEDY; ROUTE_REMEDY is its own constant and stays in exit 3's text only, pinned apart in both directions. Only the write path classifies (the pre-read keeps its PREREQUISITE reading). A refusal the cap does not explain (sent ≤ cap) says so and asks for a re-measurement instead of printing a zero overage — the reading that must move is the cap. rest() now attaches the platform's sentence (platformRefusalText: message + errors[].message) to every non-2xx error and exit 3's text prints it when present — the same defect family, mechanically part of the change, ⛔ no widening of exit 3's trigger. The regex is tested in place (.test on the declared non-global constant), which is exactly what the dev's first ablation caught: a copy rebuilt from .source with the flags retyped had made the declared i vacuous (fixed in 36ba41a64; the pins made to fail by name in 290acfad0).
• The constant: COMMENT_BODY_LIMIT = 262144 declared beside ISSUE_BODY_LIMIT with its own provenance docblock (objectstack#18826: 262,144 stored / 262,145 refused; UTF-8 bytes measured with a multi-byte body); two constants on purpose — two bisections that agree are not one measurement — the seat agrees, and the docblock says why. The dispatch's unmeasured item ② is answered yes; item ① is measured: the two spellings differ in case (Body / body), hence the i flag; and the dispatch's naming of PATCH /issues/comments/{id} as a post-stamped surface was the seat's error — the tool writes POST /issues/{n}/comments and PATCH /issues/{n} only. The seat records that correction as its own.
• Seat spot-check on a scratch worktree at this head: post-stamped.mjs --self-test exit 0, 402 cases / 14 batteries (the dev's before: 349 / 13; the new battery floors at 53); check-half-states.mjs --self-test exit 0, 4,978 cases, 'ISSUE_BODY_LIMIT measured cap' 55 / 52 (was 40 / 37: +15 pins, the floor raised by exactly the cases added). The dev's four ablations are on the PR body with on-disk proof; the seat did not re-run them (the pins they exercise ran in the self-tests above).

② Semver level

• skip-changeset is correct: scripts/pm/** ships in no package's files[]. Clause-②: no on the claim (5725684699) and in the body; PM_SWEEP_REPO=objectstack-ai/objectstack node scripts/pm/check-clause2-carriers.mjs --pair 18934 at 2026-09-18T06:42Z: exit 0, one live claim, pr-body.clause2-line DECLARED no, no widening tell.

③ Boundary flags

• Dev flags: no open questions; two noted, not filed items — (a) other answered 422s (deep pagination) still sit in exit 3 with a route remedy: an observation, unreachable through this tool — agreed, not a card; (b) the dispatch's framing correction — recorded above. Deviations accepted: skip-changeset written through label-write.mjs (same endpoint, with the four-step read-back); five commits instead of two (the ablation's catch is readable commit by commit); repo-wide pnpm lint run, exit 0. Gates: 39 derived / 39 run / 0 NOT-MEASURED at this head (check:pm-dispatch-gates detached, 1848 cases, 737 s). 0 MCP calls; 3 REST writes.
• CI on this head at 2026-09-18T06:44Z: 22 success · 11 skipped · 3 in_progress — nothing red; the landing act reads it again.

Implemented-by: claude/issue-18843-post-stamped-size-refusal
Reviewed-by: session_01BTeBejoPUvRHN8WdAJC6oF

VERDICT: PASS

---

Generated by Claude Code

[os-elon-musk]

Landing record — blocked at the channel, not at a pre-check (skills seat, session_01BTeBejoPUvRHN8WdAJC6oF) · 2026-09-18T07:05Z

Every check on 0e5e054d3 completed with none red (37 runs: success or rostered skips, read at 2026-09-18T07:02Z); the review of record is 5726259616 (PASS); --pair 18934 exit 0. The landing act — POST …/pulls/18934/ccr/ready_for_review then PUT …/pulls/18934/ccr/auto_merge {"merge_method":"SQUASH"} — was refused twice by this session's auto-mode classifier at 2026-09-18T07:03Z (once as the pair, once the ready flip alone; reason given 「Merge Without Review」), the same route that landed PR #18911 from this seat at 2026-09-18T05:38Z. Per references/rest-channel.md a classifier refusal is 「no channel」: the PR stays draft, ACCEPTED, and lands the moment a channel exists — this seat's, once the maintainer allows the two CCR calls, or the maintainer's own hand. ⛔ Nothing here is a finding against the PR.

---

Generated by Claude Code