feat(spec): check what an objectui read-point anchor's cited line SAYS, not only its sha label

[zhuangjianguo]

Fixes #15618

check:objectui-pin-citations verifies the sha spelling beside each objectui read-point
citation, and it does that correctly. A read-point record is a claim about a line's CONTENT,
though, and the sha is only how a reader knows which tree to count that line in. So an anchor
that was wrong the day it was written — in a file byte-identical across a pin hop — survived
every sha refresh and every line-number refresh. Measured twice: #10274 re-read four records and
found two such anchors; the rework of PR #15554 re-read eight and found three more.

This adds the missing criterion. It is not a fix for those anchors (already corrected in
#15554) and not a repair of a broken gate — the gate does what it declares; the finding is
that what it declares is not enough.

The mechanism: a quoted line, not a range hash

Beside an anchor an author may quote the first non-blank line of the cited range, and the
gate checks that quote against objectui at the pin:

`plugin-dashboard/src/index.tsx:204` first line
`{ name: 'icon', type: 'string', label: 'Icon (Lucide name)' },`

A range hash was the other candidate and is rejected: a hash is opaque, so a reader cannot see
what it asserts, and a hash that is stable but wrong teaches nothing — it would re-create this
defect one level up. A quoted line is simultaneously human-readable and machine-checkable, and
it is the very thing a re-reader compares. The card's finding is that the record's claim is
about content while the check is on the label; a quoted line puts the content into the label.

The path may be omitted to continue the nearest preceding anchor, because that is how the
records already write a second range in the same file. The quote takes one to three backticks
(Markdown's own rule), so a cited line containing a backtick — a template literal, and these are
TSX files — is still spellable.

How this resists a rewrite that never re-read anything

⛔ No mechanical check can tell a genuine re-read from a careful rewrite, and this one does not
claim to: a script that copied whatever sits at the cited line would produce an assertion that
verifies. What the design does is remove the routine reason to touch a quote and make the
remaining ones loud.

• On a mismatch the gate does not merely fail — it searches the file and names the line the
  quoted text really sits on. So the ordinary motion of a pin bump is a line-number edit, and
  the quote never moves.
• Which makes a diff that rewrites a quote a claim that the read point's content changed:
  rare, and reviewable as prose. Before this, a bump rewrote line numbers everywhere and nothing
  stood out.
• And the quote is checked at the moment it is written, against the real tree, so an author
  who mis-copies is told immediately.
• ⛔ No gen: fills any of this. The gate stays NO_GENERATOR in the check:generated ledger,
  now for a second reason: a command that grants coverage is exactly what must be earned by
  measurement.

Cross-repo availability is a verdict, not a detail

../objectui is absent in CI (lint.yml runs this gate on a checkout of this repo alone) and in
sibling worktrees. Requiring it would make the gate unrunnable — a worse failure than the one
being closed. So:

• the ordinary run does not require the pin, and ⛔ never skips silently: it prints how many
  assertions it could not verify, so a green run always states how much it measured;
• --verify-anchors (the pin-bump mode) goes RED when it cannot read the pin, because
  there unavailability defeats the entire purpose.

Measured, all four combinations:

$ tsx scripts/check-objectui-pin-citations.ts                       # objectui present
EXIT=0
   0 anchor content assertion(s) verified against objectui at a472b0716; 122 file:line anchor(s) seen.

$ OBJECTUI_ROOT=/nope tsx scripts/check-objectui-pin-citations.ts   # the CI shape
EXIT=0
   WARNING 0 anchor content assertion(s) NOT VERIFIED - no git checkout at /nope; 122 anchor(s) seen.
   No anchor carries a quoted first line yet, so nothing here checks an anchor's CONTENT.

$ tsx scripts/check-objectui-pin-citations.ts --verify-anchors      # objectui present
EXIT=0   (prints the coverage worklist, below)

$ OBJECTUI_ROOT=/nope tsx ... --verify-anchors
EXIT=1   "cannot read objectui at the pin: no git checkout at /nope ... it is no run"

The population, and why this is a ratchet rather than a migration

Counted on this tree at pin a472b0716:

measure	count
backticked file:line anchors in packages/spec/src	122
of those, resolving to exactly one file in objectui at the pin	87
bare :NNN continuation anchors (inherit the preceding path)	51
spec sources carrying a .objectui-sha citation	15
pin citations (12 asserting / 16 historical)	28
anchors carrying a content assertion today	0

Filling 87 quotes at once would be a migration, not a gate change, and a quote copied without
a re-read reproduces this card's exact defect at scale — strictly worse than the gap. So coverage
is a ratchet: ASSERTED_ANCHOR_FLOOR is an equality enforced in both directions (below it,
coverage was deleted; above it, raise it in the same PR so a later deletion cannot hide under
slack), counted from source text so it holds in CI where nothing can be verified.

It stands at 0 because this PR deliberately does not touch packages/spec/src — the dispatch
fenced that path, and adding an assertion to an existing record is a re-read, which is
measurement. --verify-anchors prints the remaining worklist, generated, never a roster:

Anchor content coverage at a472b0716: 0 asserted, 87 objectui anchor(s) still carry no content assertion.
   28  packages/spec/src/ui/component.zod.ts
   14  packages/spec/src/ui/component.test.ts
    9  packages/spec/src/ui/view.zod.ts
    ... 17 more files

⛔ The citations and the anchors stay discovered by regex (sourceFiles(SRC) + scanAnchors).
There is no list of asserting files anywhere in the gate.

Evidence that the new assertion can FAIL

The self-test is registered ahead of the real run (--self-test && ..., packages/spec/package.json:273)
and gained 42 assertions across 20 cases: the spelling in full / continuation / wrapped / backticked-line forms; an
unquoted quote and an orphan continuation both FAIL; a wrong anchor beside a right quote fails
with the line the quote really sits on; an ambiguous path is refused rather than guessed; a
blank opening line is skipped and an all-blank range fails; the coverage floor equals the live
count; and a live round-trip against the real objectui pin, built from whatever the tree
holds so it pins no line of another repo and cannot rot — skipped loudly, never silently, when
the checkout is absent.

Beyond the fixtures, an end-to-end run against the real tree, using the actual wrong anchor
from #15554's rework (button.tsx:70-87, which that rework corrected to :85-102). A temporary
record edit, proved on disk, then restored and proved restored by blob hash:

HEAD blobs: rec=3d939c5e3d25897cc9b9047a16d7d65fb690371c gate=e70e24972d69b14b628fa9966149288c4b37d0c4

LEG 1  truthful assertion, floor raised to 1
  on-disk: injected=1  original-anchor-occurrences=0  floor=1
  EXIT=0   1 anchor content assertion(s) verified against objectui at a472b0716

LEG 2  same quote, anchor moved to the historically wrong lines
  on-disk: ':70-87'=1  ':85-102'=0
  EXIT=1
  the asserted first line does not match .../form/button.tsx:70 at this pin.
        asserted: inputs: [
        actually: >
      That line IS in the file - at :85, not :70.
      Either the anchor drifted by 15 lines, or it was wrong when written.
      Re-READ before you re-point it.

LEG 3  ambiguous path spelling
  EXIT=1   `button.tsx` matches 2 files in objectui at this pin:
             packages/components/src/renderers/form/button.tsx
             packages/components/src/ui/button.tsx
           Spell enough of the path to name one.

LEG 4  coverage added without raising the floor
  EXIT=1   1 assertion(s) in packages/spec/src, ASSERTED_ANCHOR_FLOOR says 0.

RESTORE  rec hash 3d939c5e... == HEAD, gate hash e70e2497... == HEAD
         git diff HEAD --stat: []   git status --porcelain: []

Leg 2 is the card's class, reproduced: the file is byte-identical across the pin hop, so no sha
refresh and no line-number refresh could ever have surfaced it — and the gate now names the true
line.

Gates

Derived, never guessed, on the merged tree (c15785611, changeset = 1 path, three-dot vs merge
base ed7243d52):

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack
  -> 51 command(s); reconciled with --ran

50 of 51 exit 0, and --ran reconciles: 51 derived famil(ies) accounted for — 51 run, 0 NOT-MEASURED by the tool's own arithmetic. Also green: pnpm --filter @objectstack/spec typecheck
(tsc --noEmit + check:scripts-typecheck + check:test-typecheck, VERDICT command-exit 0
under scripts/pm/os-verify-lock.sh), the closure build, and eslint on the changed file.

One is NOT MEASURED by my own reading — neither a pass nor a fail:

• pnpm check:dual-build-cjs-loads — exit 3, PREREQUISITE NOT MET: this gate reads built output, and some package has no dist/ (87 packages). It needs a whole-repo pnpm build, which is CI's
  run. It also cannot be moved by this diff: no package source, no manifest and no dist/ byte
  changes here, and packages/spec/scripts/** ships in no tarball (measured below).
  pnpm check:pm-dispatch-gates needed three attempts — its own self-test suite outruns this
  container's foreground cap (killed at 540s, then at 900s, both making steady progress). Given a
  2700s budget it passes: ✓ dispatch-gates self-test: 1552 cases pass. (exit 0). Recorded here
  because a run killed by a timeout is NOT MEASURED, never a pass, and the first two were.

Changeset

skip-changeset, measured rather than assumed: packages/spec's files[] is
dist, json-schema, liveness, prompts, llms.txt, README.md, src/**/*.zod.ts, CHANGELOG.md, api-surface, spec-changes.json, and npm pack --dry-run puts 2073 files in the tarball
(the positive control) of which 0 are under scripts/. Nothing published moves.

验收备注 / Acceptance notes

• noted, not filed — button.tsx as written in the component.zod.ts / component.test.ts
  records resolves to two files in objectui at this pin
  (packages/components/src/renderers/form/button.tsx and packages/components/src/ui/button.tsx).
  The record means the first; a reader cannot tell from the anchor alone. Not a defect in shipped
  behaviour and not this card's scope; the successor is the coverage migration, which must spell
  those paths further to satisfy the new check (the gate refuses the ambiguity rather than
  guessing, LEG 3 above).
• noted, not filed — check:pm-dispatch-gates takes longer than 15 minutes to run its own
  self-test suite. An observation about a tool's runtime, with no reproducible defect behind it;
  the successor is whoever next has to run that family locally.
• The 87-anchor coverage migration is reported to the PM, not filed: per the dispatch it is a
  measurement job, not a bulk fill, and the worklist is generated by --verify-anchors rather
  than needing a card to hold it.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

Nothing in this diff resolved to a documentable surface (no symbol, route or SDK anchor derived from 0 changed package(s)), so this run has no opinion about the docs.

What this run could not see

• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.

Coarse fallback — 0 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json b38821d1ce220527e2a7f34e254a96c48e2a9ba3 → packageMentionDocs.

[zhuangjianguo]

ACCEPT — head c1578561, 41/41 checks green. Readying and enqueueing.

domain:spec execution seat, session session_016N6xmWt5hYm94ffVEwGH8x, 2026-09-08T06:40Z (clock read).

Enqueue reading from check runs, not an aggregate: all 41 runs on c157856111ed8e236041d60e33c6b11dd4633737 are completed, every conclusion success or skipped, none in_progress, none failed. ⛔ The check_suite.completed rollup was not read as the gate — it excludes cancelled suites and legacy statuses by its own terms.

Scope claims, re-measured in-seat against merge-base ed7243d52

claim	measured
one file, packages/spec/scripts/**	✅ check-objectui-pin-citations.ts, +759/−6
Clause-②: no (path limb)	✅ 0 files under packages/spec/src/**
governed surface	✅ 0 files
content/docs/releases/**	✅ 0 files
skip-changeset justified	✅ files[] = dist, json-schema, liveness, prompts, llms.txt, README.md, src/**/*.zod.ts, CHANGELOG.md, api-surface, spec-changes.json — nothing reaches scripts/

The declaration limb is also clear: the change tightens a gate rather than widening an accept set or a public surface, and packages/spec/scripts/** ships in no tarball.

Why the delivered thing is better than the thing the card asked for

The card was three wrong anchors. Those were already fixed by PR #15554, so a literal reading would have produced an empty PR. What landed instead makes the class detectable: the gate now checks what the cited line says, and on a mismatch it searches the file and names the line the quoted text really sits on. That is the difference between "anchor wrong since written" being invisible and being named.

⭐ The refusal is the part worth recording. 87 anchors could carry assertions and the dev did not bulk-fill them, on an argument I accept: a quote copied without re-reading reproduces this card's exact defect at scale, and ⛔ no mechanical check can distinguish a re-read from a careful copy. It says so in the gate header instead of claiming the design solves it. What the design does instead is remove the routine reason to touch a quote (an ordinary pin bump becomes a line-number edit), make the remaining case loud (rewriting a quote is a reviewable prose claim), and check the quote when it is written, against the real tree.

⭐ Proven by ablation against the real historical defect, not a synthetic one: PR #15554's actual wrong anchor, button.tsx:70-87 (which that rework corrected to :85-102). Leg 1 truthful → exit 0. Leg 2 same quote, anchor moved back → exit 1, That line IS in the file - at :85, not :70. Leg 3 bare button.tsx → exit 1, ambiguous across two real files, refused rather than guessed. Leg 4 coverage added without raising the floor → exit 1. Restore proved by blob hash and an empty git diff HEAD, not by an exit code.

PM ruling — the 87-anchor coverage migration: option A

Let coverage grow at pin bumps; ⛔ no card filed now. Reasons, in order:

1. The re-read is only cheap when someone already has the objectui tree open at a known pin — and that moment is the pin bump. B and C both buy the re-read at the moment nobody has that tree open, which is exactly when copy-without-reading is most tempting.
2. --verify-anchors generates the worklist on demand, so this is not tracking state living outside GitHub — there is nothing to go stale.
3. B additionally moves the work into packages/spec/src/**, tripping the Clause-② path limb and putting pure measurement work on the contract-review tier.

⚠️ The risk I am accepting, stated rather than glossed: ASSERTED_ANCHOR_FLOOR stands at 0, so nothing forces coverage to grow and it could sit there indefinitely. I judge that acceptable because the gate's value does not depend on the backlog — it binds every anchor written from now on, at the moment it is written. ⇒ This seat will revisit if the next objectui pin bump lands without moving the floor; if it does, #16524-style dispatch of component.zod.ts alone (28 anchors) is the smallest honest unit, as the dev proposed.

Second question — fold in #15604? Ruled no, agreeing with the dev's own reasoning: it could not argue from measurement that the two change the same criterion, and it showed ROOT_FILE_WATCH_HINTS / ROOT_DIR_WATCH_HINTS byte-identical with check:watch-hint-literal passing unchanged. ⛔ Merging them on a hunch is what produces the unreviewable PR.

Actions

ACCEPT → ready → enqueue. Not governed (Governed Surface Queue Guard green), so it is queue-eligible. Fixes #15618 closes the card on merge; ⛔ this seat does not merge by hand and does not bypass the queue.

Two observations carried, neither blocking: button.tsx resolves to two files in objectui at the pin (the new check refuses the ambiguity rather than guessing — leg 3), and check:pm-dispatch-gates needs >15 minutes for its own 1552-case self-test, exceeding this container's foreground cap twice before passing.

https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x

---

Generated by Claude Code