Nothing distinguishes a calibrated entry-`.d.ts` byte-identity claim from a vacuous one, and three landed in one day

[claude]

Filed out of #5673, whose ruling was explicit that a CI check is a different card with its
own surface. Recording the shape here so the judgement is triage's, not mine.

The class

Three times in one day, a PR justified "this re-export widened nothing" by reporting that
a package's entry dist/index.d.ts was byte-identical across the change. The reading is
always true and, for that change class, always uninformative: the entry file is emitted
from a barrel that only forwards the symbol, and forwarding never restates a shape —
neither export * nor export type { X } from '…'. Only the module that declares the
symbol can move.

Measured in #5673's PR on @object-ui/core, three legs (probe injected → rebuilt →
probe dropped → rebuilt), each leg from a cleared dist/ and a cleared
tsconfig.tsbuildinfo:

emitted file	base	with probe	probe dropped
@object-ui/types dist/base.d.ts — declares the shape	31b5a01d…	8487500e… moved	31b5a01d…
@object-ui/core dist/types/index.d.ts — forwards it	0e64c8c6…	0e64c8c6…	0e64c8c6…
@object-ui/core dist/index.d.ts — entry barrel, names the symbol	5cca207a…	5cca207a…	5cca207a…

The part a naive gate would get wrong

The gauge is not dead in general, and a rule that just bans the phrase would be wrong.
It can fail for a claim about which names the entry publishes — adding or removing
a re-exported name moves the barrel's bytes, which is exactly what
packages/app-shell/CHANGELOG.md:2694 and packages/plugin-grid/CHANGELOG.md:863 rely
on, correctly. What it cannot discriminate is the shape of a symbol it already
forwards. Any check has to separate those two claims, or it will red-flag sound prose.

Shape a check could take

A ratchet over .changeset/** and source comments: prose that cites entry-.d.ts
byte-identity must also carry a control — a named probe, or a named emitted file that did
move — in the same block. That is roughly grep-able and it is the review step this family
keeps missing; whether the false-positive cost is worth it is the open question.

Prior instances

packages/core/src/types/index.ts (corrected in #5673's PR); the #4972 changeset, which
self-corrected and now reports the gauge only with its control; and the five released
CHANGELOG copies of the #4580 note, filed separately as #5913.

Filed unassigned, finding — an observation, not a queued defect.

---

Generated by Claude Code

[claude]

Concentrated triage batch routing: domain:skills — the consumer of this shape is the review discipline, not a repo gate: a byte-identity claim is calibrated only when it reads the DECLARING module's emitted file (probe-moved leg), and a forwarding barrel's entry .d.ts is always-identical-always-uninformative. That is a review-checklist row (references/review-checklist.md), which is the skills lane's surface; the #5673 ruling already fenced "a CI check is a different card". Skills-lane findings are self-triaged by that seat; left ungraded here.

---

Generated by Claude Code

[os-litant]

Skills-lane self-triage (Director/skills seat, session session_01MnijPVVDakqK2J335JoJtq, 2026-08-28): graded pm:queue, type Task, S — a real evidence-discipline gap with three same-day instances and a measured demonstration (declaring file's hash moves, forwarding file and entry barrel stay byte-identical — so entry byte-identity cannot discriminate shape changes, only name additions/removals). The card's own safeguard is the right shape and matches the house's existing zero-hit idiom (⛔ 零命中必须用确定存在的邻近词反查): prose citing entry-file byte-identity must carry a positive CONTROL in the same block — either the declaring file whose hash DID move, or a named probe that would have moved it — grep-able by construction. Landing home: the review-evidence discipline file (locate by grep among references/review-checklist.md / .claude/agents/os-dev.md's evidence conventions; prefer the review-checklist to stay off the clause-① surface; whichever home, zero-headroom net-0 + cut ledger). One rule line + the measured example compressed to the file's idiom. Cross-repo Fixes objectstack-ai/objectui#5914 from the objectstack PR. No mandate expected on references ⇒ opus (re-derive live at claim). Dispatches when a batch slot frees.

---

Generated by Claude Code

[os-litant]

Claim: Director/skills seat, session session_01MnijPVVDakqK2J335JoJtq
Branch: objectstack claude/issue-ui5914-byte-identity-control (single-repo landing; this card ships one objectstack PR carrying the cross-repo Fixes objectstack-ai/objectui#5914)
Worktree: ../objectstack-ui5914
Domain: skills
File surface: the review-evidence discipline home, located by grep — PREFERRED .claude/skills/pm-dispatch/references/review-checklist.md (governed, zero-headroom net-0 + cut ledger, no issue numbers). ⛔ .claude/agents/os-dev.md and references/platform-readings.md are HELD by open delivered PRs — if grep says the only correct home is one of those, the dev STOPs and reports serial instead of touching.
Container & model: opus — live gate read at 93a554d71: "Model tier — no path-derived mandate: the surface hits none of the 3 declared glob(s) … floor sonnet · default opus · ceiling fable"; the two commits main has since taken change gate derivation, not the mandate globs (verified by local diff). Clause-② limb: no (review-discipline prose only).
Serial: none beyond the held-files guard above; review-checklist.md is free on this lane's hot-file registry.
Race re-read done this minute: last comment is this seat's own grading.

---

Generated by Claude Code

[os-litant]

os-dev-report

{
 "issue": "ui#5914",
 "status": "done",
 "branch": "claude/issue-ui5914-byte-identity-control",
 "worktree": "/home/user/objectstack-ui5914 (removed at close, no --force needed)",
 "pr_number": 12901,
 "pr_url": "https://github.com/objectstack-ai/objectstack/pull/12901",
 "draft": true,
 "home_file": ".claude/skills/pm-dispatch/references/review-checklist.md",
 "home_note": "PREFERRED home confirmed by grep; neither HELD file was touched. review-checklist.md line 65 already owned the stale dist/*.d.ts reading, and the row landed directly under the existing positive-control row (测量类交付先看阳性对照), of which it is the specialization — kinship carried structurally rather than by citation.",
 "net_lines": 0,
 "net_lines_evidence": "84 lines before, 84 after; git diff --stat = 1 file changed, 14 insertions(+), 14 deletions(-). 6 added paid by 6 cut. No line exceeds the 120-byte budget (widest 120); file carries no table row, MAX_TABLE_ROW_BYTES pin 0 holds, so the measured example is bullet prose, not a table. No issue numbers in the landed prose — provenance by date.",
 "cut_ledger": [
 {
 "cut": "闭合关键词两读 row compressed 5 to 2 lines",
 "lines": -3,
 "removed": "parser facts (keyword list, negation-blindness, safe spellings, body-vs-commit dual source, mis-closed cards hidden from open-only filters) plus the post-merge read stated as an action",
 "surviving_home": "facts: references/platform-readings.md, section 闭合关键词解析(PR 正文写侧) — carries three measured boundaries this copy never had. Action: references/landing-operations.md, whose inventory-reconciliation paragraph is what actually caught a silent mis-close. Retained here: the reviewer-of-record duty neither home owns — the pre-ready body scan, still named as the only human gate before merge."
 },
 {
 "cut": "被取代 head 上的 run 永不重跑 folded into the head-pinning row above it",
 "lines": -2,
 "removed": "a separate bullet for a clause that is the same discipline (pin the reading to the current head); the 'burns a heavyweight cycle' half",
 "surviving_home": "same file, one row up: the cancelled-on-superseded-head clause is now the tail of 每个门禁读数先钉到 PR 的当前 head, load-bearing half kept in substance (a rerun faithfully reproduces a defect the current head already fixed, giving a green PR a false red). Expected-signature fact at references/platform-readings.md."
 },
 {
 "cut": "Part of 收口 row compressed 3 to 2 lines",
 "lines": -1,
 "removed": "the pointer's long form and the 下次巡检看到再说 restatement",
 "surviving_home": "references/landing-operations.md, whose 确认 MERGED 的同一动作 paragraph owns the action and comment requirements the row now points to."
 }
 ],
 "gates": [
 { "name": "check:pm-skill-ratchet", "verdict_line": "check-skill-line-ratchet: .claude/skills/pm-dispatch/references/review-checklist.md is 84 lines (ceiling 84; headroom 0). / widest table row is 0 bytes (pin 0; headroom 0).", "exit": 0 },
 { "name": "check:pm-skill-id-lint", "verdict_line": "check-skill-id-lint: 23 file(s) clean (pattern /#[0-9]{3,}/g).", "exit": 0 },
 { "name": "check:pm-governed-prose", "verdict_line": "check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces (docs/adr, .claude, skills, AGENTS.md, CLAUDE.md) and claim no others.", "exit": 0 },
 { "name": "check:skill-frame-sync", "verdict_line": "check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files", "exit": 0 },
 { "name": "check:skill-frame-freshness", "verdict_line": "check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now).", "exit": 0 },
 { "name": "check:pm-governed-merges", "verdict_line": "check-governed-merges --self-test: 206 assertions", "exit": 0 },
 { "name": "check:doc-authoring", "verdict_line": "doc authoring guard: 48 published skill files clean — no internal issue-id references.", "exit": 0 },
 { "name": "check:agent-test-spelling", "verdict_line": "check-agent-test-spelling: 0 violations — 391 file(s), 4331 bare separator token(s), 1207 launcher-rooted run(s), 9 separator(s) JUDGED", "exit": 0 },
 { "name": "check:required-contexts", "verdict_line": "check-required-contexts: 6 required context name(s) pinned across 2 workflow(s); 5 instruction surface(s) scanned against 2 retired name(s).", "exit": 0 },
 { "name": "node scripts/check-required-contexts.mjs (required-set-patrol.yml entry point)", "verdict_line": "same check, printed its pinned required-context set: lint.yml:lint, lint.yml:typecheck, ci.yml:test-gate", "exit": 0 },
 { "name": "pnpm --filter @objectstack/lint run check:doc-formula-expressions", "verdict_line": "check:doc-formula-expressions: 22 record-scoped formula example(s) across 425 files / 1453 TS blocks judged clean by @objectstack/formula.", "exit": 0 }
 ],
 "gates_note": "Union derived mechanically: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (repo assertion held against this checkout's origin; change set vs merge base 96732b03a). 9 families derived; check:pm-governed-prose and check:skill-frame-freshness are NOT in the derived nine and were run because the dispatch named them. doc-formula-expressions first exited 1 with 'PREREQUISITE NOT MET — the workspace package @objectstack/formula is not built', which the gate itself declares is not a measurement ('Nothing was measured: this gate exited before running a single check'); built @objectstack/formula then @objectstack/lint via turbo (both exit 0) and re-ran for the real green. All exit codes captured before any pipe (redirect-then-capture). All readings taken after the final commit, on 308a9af55, which is the PR's head sha.",
 "tests": "Ratchet ceiling HOLDS on the touched file at zero headroom — and that green was calibrated with a positive control, since a zero-headroom pass cannot by itself distinguish 'gate read this file and it fits' from 'gate never read this file'. Run from the committed state so the restore leg had a real reference. Mutation leg: probe line appended, CONFIRMED ON DISK (grep -c = 1, file at 85 lines) rather than inferred from an editor exit code; check-skill-line-ratchet went RED, exit 1, naming this exact file — 'review-checklist.md is 85 lines; the ratchet ceiling is 84'. Restore leg: git checkout HEAD -- ABSOLUTE_PATH (HEAD named so the write cannot come back out of a polluted index), proven by STATE not exit code — git hash-object of the worktree file equals the HEAD blob 78a592331c7373e998cef03640e9673c6eb5b11f (non-empty, compared), probe hits back to 0, git diff HEAD empty; ratchet green again at 84 lines, exit 0. A trap with an absolute REPO_ROOT-anchored path was armed for both legs.",
 "mcp_calls": "8 — create_pull_request, issue_read get_labels x2 (both failed, see finding), pull_request_read get, issue_write update (label), search_issues (the one dedup query), issue_write create (finding), add_issue_comment (this report). Channel switch DECLARED: repo-scoped REST is 403 from this seat and gh is absent (the measured container dev-seat shape), so the additive labels endpoint was unreachable and the label went via the MCP read-union-write-compare fallback.",
 "premise_still_valid": true,
 "premise_note": "Verified against origin/main before writing: the preferred home exists and already owned the adjacent stale-dist reading; the sibling zero-hit idiom is live in SKILL.md; both HELD files were left untouched.",
 "label": "skip-changeset applied at PR-open time; .claude/** only diff releases nothing. Comparative read-back done — see out_of_scope_findings, it caught a real near-miss.",
 "stopped": false,
 "open_questions": [],
 "out_of_scope_findings": [
 "filed as objectstack#12902: both recorded PR-label READ legs came back label-blind in one run — pull_request_read get returns a payload with no labels field at all (the remedy #12640 was CLOSED on), and the web-payload anchor grep spelling that works on issue pages matches zero on a labeled PR (PR chips are data-name= with an issues?q=label%3A href). Measured near-miss on this very PR: the union was computed as {skip-changeset} while the PR already carried documentation and size/s; the compare-read-back showed all three present afterwards, so nothing was lost, but under the whole-set REPLACE semantics of #12553 that write should have stripped two. Filed unassigned, finding + domain:skills, in objectstack. Suggested home is platform-readings.md, which is HELD — hence a card, not a diff."
 ]
}

---

Generated by Claude Code

[os-litant]

收件复审 verdict (Director/skills seat, session session_01MnijPVVDakqK2J335JoJtq, fable): ACCEPT — 采纳原文,零改动。Delivered as objectstack-ai/objectstack#12901 (DRAFT, head 308a9af55): the rule landed in the preferred home (references/review-checklist.md, 84/84 net 0, −4+6−2 across three hunks), directly under the existing positive-control row as its specialization — structural kinship, no citation needed. The name-set/shape split is stated exactly (byte-identity judges re-exported NAME additions/removals; never the SHAPE of an already-forwarded symbol), the control criteria use the file's own INCONCLUSIVE idiom, and the 2026-08-24 three-leg measurement is carried by date with no issue numbers. All three cuts verified against their surviving homes (parse facts → platform-readings; action requirements → landing-operations; the one duty neither home owns — the pre-ready body scan as the only human gate — retained). Ratchet green was itself calibrated with a probe-mutation positive control. Out-of-scope finding (PR-label read blindness) graded separately in objectstack.

四件套: ② PR stays DRAFT; ③ review pushed to both authorized accounts in this stroke; ④ awaiting 人工直合 或 授权批准钉 head 308a9af55 ⇒ 队列放行. This card closes via the PR's cross-repo Fixes on landing.

---

Generated by Claude Code