Record the object-literal code-helper blindness with its measured blast radius (no widening)

[os-elon]

Fixes#13131

This round is scoped to measurement, per triage's grading comment: 「本轮范围只到测量 …… 先量出加宽后的爆炸半径，⛔ 不先加宽。数没量出来之前，"该不该加宽"不是一个可裁的问题。」

⛔ No shape is widened. ⛔ No verdict row is added. The gate's verdict on the tree is unchanged: 22 classified sites, 0 unresolved, exit 0 — byte-identical to main's verdict.

What lands

One file: scripts/check-dispatcher-error-vocabulary.mjs.

• OBJECT_LITERAL_CODE_HELPER_BLINDNESS — a frozen declaration carrying the corpora, the method and the measured numbers, following the KEBAB_DIAGNOSTIC_VOCABULARY precedent in the same file.
• A --self-test block pinning both blindnesses, each against a same-genre positive control, so a later widening of either half cannot land silently.
• A line in the run's printed bounds, so a partial gate cannot read as a complete one.

The card's four-corpus table, reproduced through the real deriveSites and the real findViolations

#	helper body	code	vocabulary gate	casing gate
A (control)	(e as any).code = code	HELPER_SCREAMING	1 site, shape codehelper	0
B	return { severity: 'error', code, message }	HELPER_SCREAMING	0 sites, 0 unresolved	0
C	same, shorthand	helper_lowercase	0 sites, 0 unresolved	0
D	return { severity: 'error', code: code, message }	HELPER_SCREAMING	0 sites, 0 unresolved	0

A reproduces exactly, so B/C/D's zeros are readings rather than a dead instrument.

The blast radius — the two counts, reported separately

Predicate: a code property of an object literal whose value is an identifier that is a parameter of the enclosing function. Evaluated on the real TypeScript AST over packages/** non-test source (2082 files, 297 registered codes), with call-site arguments reduced by this gate's own literal grammar (^[A-Za-z][A-Za-z0-9_]*$, so a hyphen does not reduce) and checked against the registered vocabulary and the declaration table.

predicate matches 13 helpers (12 named, 1 anonymous)
(a) call sites newly reached 106 - 76 reduce to a literal, 30 do not
(b) NEW verdict rows needed 29 - 0 of them already carry a row
(c) undischargeable unresolved 4

Split by whether enclosingDeclaration can see the declaration at all:

bucket	helpers	(a) call sites	(b) new rows	(c) unresolved
reachable today (position blindness only)	9	81	29	1
blocked by the class-method blindness	4	25	0	3

⚠️(b) is not the whole cost, and reading it as the whole cost is the trap. All 29 new rows are lowercase FieldErrorCode diagnostics (required, invalid_type, min_value, …) in four files — ADR-0112 D6 genre, cheap paperwork. The 4 unresolved findings are the expensive half: an unresolved entry is pushed unconditionally and no declaration row discharges it (see reconcile), so each is a RED gate with no verdict available. That includes Parser#error, whose 16 call sites all pass kebab literals this gate's grammar refuses to reduce.

A second, independent blindness — the card's own live instance needs both

enclosingDeclaration's DECL_HEADER_RE recognises function NAME(, constructor( and const|let|var NAME = ( — and no class method. Measured with a same-genre control:

#	declaration form	body	sites
K (control)	free function	e.code = code	1, codehelper
I	class method	e.code = code	0, and 0 unresolved
J	class method	{ severity, code, message }	0, and 0 unresolved

Parser#error in packages/sdui-parser/src/parse.ts is a class method and stamps through an object literal, so widening the position alone would not reach it. Filed separately as #13226. parseParamNames mis-parsing two-modifier TypeScript parameter properties is #13227 — same failure class, third mechanism.

Triage's p2 basis — verified, and refined

The grade rests on there being no visible victim today. Verified across the whole repo, not relayed: of the 33 SCREAMING_SNAKE values the widening would reach, 33 are already registered, so the count that are both ADR-0112 D1 shaped and unregistered is 0. check:error-code-casing's findViolations returns 0 on every file involved. The p2 grade stands.

Two refinements to the card's reasoning, neither of which moves the grade:

1. Kebab-case is not a full exclusion in the helper position. Measured: the A-control spelling with a kebab call-site literal yields 0 sites but 1 unresolved finding. Kebab keeps the values out of the site population; it does not keep the helper out of the finding population.
2. The reason Parser#error is silent today is the blindness itself (both halves), not only its casing.

Verification — all on the final commit 5f70b8bfb9

Exit codes captured before any pipe, and each verdict quoted from the gate's own line.

check-dispatcher-error-vocabulary --self-test exit=0 "8 shapes + 175 assertions OK"
check-dispatcher-error-vocabulary exit=0 "OK - 22 unregistered code-stamping site(s), all classified"
check:self-test-wired --self-test / (scan) exit=0/0 "every one of the 148 script(s) CI runs ... has that self-test run by CI"
check:nul-bytes exit=0

Derived family, node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (13 matched + 2 convention-triggered for editing a gate script) — all exit=0:
check:agent-test-spelling, check:bash32-floor, check:cli-command-ids, check:cross-package-test-inputs, check:dispatcher-error-vocabulary, check:entry-guard, check:parse-guard, check:pnpm-filter-targets, check:watch-hint-literal, check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, pm/bare-root-worklist.mjs --self-test, check:pm-dispatch-gates.

check-test-completeness.mjs exits 3 with PREREQUISITE NOT MET — its own text says this branch is NOT MEASURED, not a red (#13110).

The pins are shown capable of failing

A green self-test over a healthy tree proves nothing, so both halves were ablated. Each mutation was proved on disk by an anchored grep -c on the injected text plus a blob-hash change; each restore leg used git checkout HEAD -- ABSOLUTE_PATH under an EXIT INT TERM trap and was proved by hash equality with the HEAD blob and an empty git diff HEAD, with the self-test green again afterwards.

• Mutation A — widen codehelper with an object-literal shorthand anchor: self-test exit 1, 4 assertions fire, naming the measured numbers and the decision.
• Mutation B — give DECL_HEADER_RE a class-method header: self-test exit 1, the class-method assertions fire, including the mechanism-level one on enclosingDeclaration.

The counting instrument was falsified before its output was read: 5/5 corpora (A/B/C/D plus the class-method case) and 8/8 negative controls — call-argument list, array element, template interpolation, destructured local, destructured param, module constant, property-access value, object pattern in an assignment. An earlier regex instrument was discarded because it matched ${code} in a template and f(a, code, b) in an argument list, then lost a true positive when a bracket inside a regex literal unbalanced its scan.

Bounds this measurement does not have

helperCodesFor scans only the declaring file, so every count is over in-file call sites. An exported helper called from another package contributes 0 — sendError in packages/types/src/response-envelope.ts has 0 in-file calls and is one of the four unresolved. Counting cross-file callers needs resolution this source scan does not have: that number is unmeasured, and (a) is a lower bound on it, never an upper one. Measuring it would need a real module graph or a type-aware pass.

Scope

packages/sdui-parser/** was read as evidence only and is untouched here (held by #13154). No changeset: this edits a CI-internal gate script and releases nothing, which is the skip-changeset case lint.yml prescribes; the label is applied on this PR.

The widening question is now decidable, and by triage's own framing it belongs on its own card together with the domain:cli verdict rows. Not filed here — that is the PM's call, and it is raised in the dispatch report.

---

Generated by Claude Code