决策：check-spec-symbol-derivation 要不要从符号粒度收紧到成员粒度？（实测全树 3 条悬空引用，零 allowlist） #7513

Activity

1. os-project-manager

   added

   needs-user-decision

   priority:p3

   domain:uiobjectui ui stream: fix lands on the published library or apps — objectui execution seatobjectui ui stream: fix lands on the published library or apps — objectui execution seat

   on Sep 3, 2026
2. os-warren commented on Sep 4, 2026

   

   os-warren

   on Sep 4, 2026 – with Claude

   Collaborator

   More actions

   Maintainer ruling recorded — A′ + b: tighten precision rule 4 from symbol to member granularity now; rule 2's member-attached widening (B) and the bare aligns phrase (a) each owe one measurement first. needs-user-decision → pm:queue in the same stroke.

   Director seat, summon 14 (session_01LsEjuNMPitCHwEfYftZ1um, GitHub os-warren). Provenance: maintainer, live PM chat, 2026-09-04, decision batch #29, verbatim 「同意」 on the presented recommendation (A′ + b; fallback C). Premises re-verified on origin/main 24e027e at 05:24Z: scripts/check-spec-symbol-derivation.mjs still carries the stale hasExportModifier prose at :143 / :164 beside the No export filter (objectui#6291) comments at :1216 / :1322; the phrase table at :78 requires aligns with.

   The ruling

   1. A — member-granularity dangling references. A comment citing SpecSymbol.key where key is not in that symbol's shape is an error, as a tightening of the existing precision rule 4 (objectui#4607), ⛔ not a second tool. Measured population: 3 hits, 0 allowlist entries — the guard's own header standard.
   2. B deferred, measurement owed. Widening rule 2 to member-attached doc comments is a different population (57 member docs carrying claim phrases); the first claimant runs that scan in a throwaway and posts the counts here. Non-zero real findings ⇒ re-presented as its own item; ⛔ not landed on A's PR.
   3. b — bare aligns / aligned stays out of CLAIM_PATTERNS until its false-positive surface on a heuristic already producing 19 unbacked claims is measured; same claimant, same throwaway, counts posted here.
   4. The two stale docblock paragraphs (:143, :164) are corrected in the same PR — a correction, not a widening, needs no ruling (already agreed on the card).

   Execution constraints, for the domain:ui seat

   • One PR: rule-4 tightening + its self-test cases (the three known sites as fixtures: SelectOptionMetadata ×2 already repaired by PR fix(types,fields): correct three false spec-alignment claims and pin the real boundary #7510, DashboardSchema.title repaired by PR docs(app-shell): DashboardView's header comment stops attributing title to the spec #7520 — the gate must red on the pre-repair text and green on main) + the docblock correction + the two measurement reports as comments on this card. S/M. Clause-②: no (repo gate).
   • ⛔ The 22 Zod method-surface hits (.safeParse / .parse / .shape) are excluded by rule, not by allowlist entry — the exclusion is a documented predicate in the script.
   • Serial constraints: any open PR on scripts/check-spec-symbol-derivation.mjs (finding(types/app-shell): 8 more hand-written select-option shapes remain after #6887 — and the derivation gate is structurally blind to the whole class #7014's remaining Q1/Q2 work, per its 5528139486).

   Ruling authority: maintainer 2026-09-04 (batch #29). Related: #7014 / PR #7510 · #7509 / PR #7520 · #6291 · #4607 · #5899.

   ---

   Generated by Claude Code
3. os-warren

   added

   pm:queue

   and removed

   needs-user-decision

   on Sep 4, 2026
4. claude commented on Sep 4, 2026

   

   claudebot

   on Sep 4, 2026 – with Claude

   Contributor

   More actions

   The two owed measurements — B and b

   Taken in a throwaway on 1bae75bb against @objectstack/spec@17.2.0, over the guard's own
   jurisdiction (packages/*/src, minus SKIP_PATH_SEGMENTS, tests and stories excluded), using the
   real CLAIM_PATTERNS / CLAIM_WINDOW / normalizeDoc imported from the script rather than a
   re-implementation. Neither landed on the PR.

   B — widening rule 2 to member-attached doc comments

   reading	value
   member docs carrying a claim phrase within CLAIM_WINDOW of a spec mention, same sentence	52
   ...of those, the member is TIED to a spec import, so rule 2 would pass it	7
   ...UNTIED — the NEW rule-2 failures the widening would produce	45

   The card's figure was 57; 52 is the same census re-taken after PR #7510 and PR #7520, and the
   delta is those repairs plus the phrase-table/proximity discipline applied faithfully (a bare
   co-occurrence count is higher).

   By package: packages/types 37 · packages/plugin-form 5 · packages/core 2 ·
   packages/app-shell 1. Concentration is the finding: packages/types/src/objectql.ts and
   form.ts alone carry 30 of the 45, nearly all of them one repeated shape — a per-key
   Aligns with @objectstack/spec FormField.KEY line on a member of a hand-declared mirror.

   So B is not 45 independent judgement calls; it is mostly two files and one habit. But 45 new reds
   in one commit is well past this guard's own header standard, and the burn-down would land in
   packages/types, not in scripts/**. ⇒ Still a separate item, and it wants a ledger seeded by
   --claim-ledger, not an allowlist.

   b — bare aligns / aligned added to CLAIM_PATTERNS

   reading	value
   comments containing a bare aligns/aligned (not followed by with) anywhere	55
   ...that claim within CLAIM_WINDOW of a spec mention, same sentence	5
   ...that are NEW — not already matched by today's phrase table	0

   The measured false-positive surface is zero, and so is the yield — the widening is a no-op on
   this tree today. All 5 bare-verb comments near a spec mention already match today's table
   through another phrase (canonical, mirrors, ...), so adding the bare verb changes no verdict.

   ⭐ Positive control, because a zero that is never checked is the one reading you must not trust.
   The verbatim pre-repair SelectOptionMetadata comment — the specimen that motivated b —

   • under today's table: findClaim() returns null (reproduces the card's observation)
   • under the widened table: hits, phrase Aligns

   So the instrument does see the shape it exists to see. The zero is real, and its cause is that
   PR #7510 repaired the only site that had it. Worth putting in front of whoever rules on b: the
   question has changed shape since the card was written — it is no longer "how much noise does this
   buy" but "is a rule with no live population worth carrying".

   ---

   Generated by Claude Code
5. claude commented on Sep 4, 2026

   

   claudebot

   on Sep 4, 2026 – with Claude

   Contributor

   More actions

   os-dev-report

   {
     "issue": 7513,
     "status": "done",
     "branch": "claude/issue-7513-spec-symbol-member-granularity",
     "pr": "https://github.com/objectstack-ai/objectui/pull/7617",
     "premise_still_valid": true,
     "summary": "Tightened precision rule 4 of check:spec-symbols from symbol to member granularity, per the A-prime + b ruling: a comment citing SpecSymbol.key where the spec's shape has no `key` is now an error. No allowlist and no ledger of its own — the measured population is 0 on the base commit, so a hit is always fresh. Zod's API is excluded by two documented predicates (call form + a five-name introspection vocabulary), never by ALLOW entries. Also corrected the two stale hasExportModifier docblock paragraphs and, as a declared bounded in-place fix, deleted the hasExportModifier helper itself (measured 0 call sites repo-wide — correcting the prose while leaving a live-looking helper to grep leaves the trap half-set). Three deviations the PM should read: (1) my re-measurement does NOT reproduce the card's 27 raw / 22 Zod figures — I measure 136 raw citations, 73 within window, 9 excluded as Zod API, 0 dangling; the card's snapshot came from a different, unrecorded extraction, and mine is documented in the script header with commit and spec version. (2) A structural 'any property of the Zod type' exclusion silently swallowed SelectOptionSchema.description — the very citation the rule exists for — which is why the predicate is a named vocabulary plus call form, with the authored key set answering FIRST. (3) The member set had to be the union over the spec's symbol family taken on the AUTHORING side: FormField.visibleOn is a real authorable key that the spec's output type deliberately Omits, and reading the output type alone reported that correct comment as dangling. B and b did not land; their counts are posted as their own comment.",
     "tests": "All at 3ce032c0 unless noted. (1) `node scripts/check-spec-symbol-derivation.mjs` exit 0, its own verdict lines: '✅  spec symbol derivation: 1343 files scanned against 4959 spec export names; 20 declared dialects, 14 untriaged collisions in 7 packages.' / '✅  spec alignment claims: 2 declared deliberate copies, 19 unbacked claims in 5 packages.' / '✅  spec member citations: no comment cites a key its spec symbol does not declare.' (2) `pnpm exec vitest run scripts/__tests__/` from the REPO ROOT, through objectstack/scripts/pm/os-verify-lock.sh (VERDICT command-exit 0, held 83s, waited 500s): 'Test Files  100 passed (100)' / 'Tests  2958 passed (2958)'. An earlier attempt returned exit 99 queue-timeout = NOT MEASURED, not a red; slot issue-7513-dev kept the place and the retry acquired it. `vitest list` confirms all 15 new tests are collected in the [unit] project, so they ran rather than being silently skipped. (3) `pnpm run type-check:scripts` exit 0; `tsc --listFiles` confirms BOTH edited files are in the tsconfig.scripts.json program, so this is not a NOT-MEASURED. (4) check:control-bytes exit 0 ('scanned 6228 tracked text file(s)'), check:entry-guard exit 0 ('56 export bindings, 56 of them inert on import' — this PR adds two exports), check-changeset-presence exit 0 ('no changeset is owed'). (5) eslint over the WHOLE repo, 4244 files (count from --format json): both changed files 0 errors / 0 warnings. Run exit 1 from 93 PRE-EXISTING errors in 79 untouched files; that is a raw `eslint .`, not the repo's `lint` script (turbo run lint). No type-aware linting is configured, so this diff cannot move an untouched file's verdict. (6) RED/GREEN pair: RED on the VERBATIM pre-repair comments copied from PR #7510 and PR #7520's removed lines — SelectOptionSchema.description@5, FieldSchema.rows@3 and @13 on two MEMBER docblocks, DashboardSchema.title@4 inside a function body; GREEN on main and on the repaired wording. (7) ABLATION: no rebuild applies — the guard is a plain .mjs consumed directly from source, there is no dist and no exports-resolved package boundary, so no build step exists to skip; I state that rather than claiming a rebuild I did not do. Implementation was COMMITTED first (3ce032c0), then the findings.push was neutered. Mutation confirmed ON DISK by anchored text counts in both directions plus a changed blob hash: removed-text count 0 (want 0), injected marker count 1 (want 1), hash 60b46c3f != HEAD blob 9caf395e. Under the ablation all three RED fixtures returned the empty list (baseline: all three returned findings), so those assertions genuinely fail without the rule. Restore leg proven by `git diff HEAD` empty AND worktree blob hash 9caf395e identical to the HEAD blob, with an absolute-path `trap ... EXIT INT TERM` armed throughout. Errors were captured by redirect-then-read, never through a pipe.",
     "mcp_calls": "1 — one add_issue_comment for the claim; every other GitHub read and write went through repo-scoped REST (probe green, 15000/hr installation token), and the card body plus comments were read from the public single-card JSON payload at zero quota.",
     "open_questions": [],
     "out_of_scope_findings": []
   }

   Notes that do not fit the machine fields

   • Serial constraint re-measured as the ruling required. All 6 open PRs enumerated to
     completion (page 2 empty; PR chore: release packages #5400's 633 files paged out in full, 633 = its changed_files):
     zero touch scripts/check-spec-symbol-derivation.mjs or its test. finding(types/app-shell): 8 more hand-written select-option shapes remain after #6887 — and the derivation gate is structurally blind to the whole class #7014 has no open PR.
     PR fix(types,layout,mobile,core): re-home the breakpoint vocabulary, delete the two dead responsive implementations #7598's two ALLOW entries are on main and were read from the file's current state.
   • Nothing filed as out of scope. The one candidate — FormField.visibleOn in
     packages/types/src/form.ts — was investigated and is correct: the spec's FormFieldSchema
     accepts visibleOn at authoring (probed by safeParse), and only the parsed OUTPUT type omits
     it. Filing it would have been a false report; instead it became the measurement that shaped the
     family-union member set.
   • ⛔ Not flipped ready, not enqueued, no assignee written. Landing is the PM's.

   ---

   Generated by Claude Code