docs(tooling): state check-spec-symbol-derivation's export-filter boundary, with the measured population

[yinlianghui-tw]

Part of #5899

The card's own sequencing: the decision (one-line change / scoped change / documented non-goal) hangs on a number nobody had measured. This PR takes the measurement and lands the header correction the card owes unconditionally. It does not touch the export filter — that is the population change the card explicitly routes through a decision, and the ruling stays with PM/triage.

Premise: still valid

scripts/check-spec-symbol-derivation.mjs skips every non-exported declaration in both scanners. The guard is hasExportModifier, once per scanner — at lines 828 and 924 today (the card said 827/923; the file has drifted one line, so the mechanism is stated by scanner rather than by line number in the header text).

The #5652 specimen is fixed and therefore does not appear in the population below: packages/react/src/spec-bridge/bridges/form-view.ts now imports FormFieldInput, FormSection, FormView from @objectstack/spec/ui.

Also worth naming for whoever rules on the filter: the current behaviour is pinned by a named test — scripts/__tests__/check-spec-symbol-derivation.test.ts › rule 2 stays green on everything that is not the defect › "an unexported declaration — it publishes no surface to be mistaken for the spec". Any filter change must rewrite that pin, not merely delete a line.

The measurement

Method: an instrumented copy of the script (not a mutation of the tracked file), with hasExportModifier forced true behind an env var, plus a --probe-json mode emitting every finding pre-ALLOW with file/line/kind and the real export status. Run twice — relaxed off, relaxed on — and diffed. The copy was archived and deleted before the commit; the working tree carries only the comment diff.

Measured on a76b18cf2 (branch base) against @objectstack/spec@17.2.0.

	baseline	export filter relaxed	additional
rule 1 (spec-named symbol)	18	47	+29
rule 2 (alignment claim)	20	22	+2
distinct declarations	—	—	30 (one is flagged by both rules)

Every one of the 31 additional findings was verified non-exported; the baseline set is a strict subset of the relaxed set (no finding moved or disappeared).

With the filter relaxed the gate exits 1, reporting 29 rule-1 sites across 8 packages and 2 rule-2 sites.

The population, classified by reading every site

Classification is real-mirror (same concept as the spec symbol, hand-restated) / legitimate-local (different concept sharing a name, or already derived) / unclear. Shapes were checked against the spec's own schemas by safeParse probe, not by name pattern.

Rule 1 — 29 additional

#	package	symbol	file:line	class	reason
1	app-shell	ObjectLike	hooks/useTrackRouteAsRecent.ts:30	real-mirror	{name, label?} is a strict subset of spec/system's ObjectLike; same "minimal object metadata for a helper" concept, no divergence today
2	app-shell	AppLike	utils/appRoute.ts:18	real-mirror	spec/system requires name: string; this declares name?: unknown. Adds _packageId. Relaxes a spec-required key
3	app-shell	ObjectLike	utils/deriveRelatedLists.ts:97	real-mirror	name? optional where the spec requires it; adds list?, a MetadataProvider merge artifact the spec does not model
4	app-shell	SearchResult	views/SearchResultsPage.tsx:36	legitimate-local	spec/contracts' is the search response{hits, totalHits, processingTimeMs, facets}; this is one result row{id, label, href, type, description}. "A row is not a response" — the reasoning already in ALLOW for InboxNotification
5	app-shell	AdminScope	views/metadata-admin/PermissionAdvancedFacets.tsx:82	real-mirror	key set identical to spec/security AdminScopeSchema (6/6); a local instance safeParses clean against it
6	app-shell	RemoteTable	.../datasource/DatasourceResourcePage.tsx:126	real-mirror	spec/contracts requires columnCount; this makes it optional and drops rowCountEstimate. It types the very service response it reads
7	app-shell	ActionParam	.../inspectors/ActionDefaultInspector.tsx:266	real-mirror	all 7 local keys are real ActionParamSchema keys (probed against the strict schema); a hand subset plus a [k: string]: unknown catch-all
8	app-shell	DashboardWidget	.../inspectors/DashboardDefaultInspector.tsx:57	legitimate-local	= DashboardWidgetSchema & { id: string }, where DashboardWidgetSchema is imported from @object-ui/types — derived, one hop outside where rule 1 looks
9	app-shell	Field	.../inspectors/DashboardWidgetInspector.tsx:548	legitimate-local	a local React field-wrapper component. spec/data Field is a field-type builder. Rule 1 records functions with derived: false unconditionally — rule 2's isRendererLike skip is not applied here
10	app-shell	Dimension	.../inspectors/DatasetDefaultInspector.tsx:105	legitimate-local	spec/data DimensionSchema is the semantic-layer cube dimension: it requiressql and rejects both field and dateGranularity, which are this shape's own keys. Different vocabularies; the spec exports no DatasetSchema to bind against
11	app-shell	FlowEdge	.../inspectors/FlowEdgeInspector.tsx:49	real-mirror	key set identical to FlowEdgeSchema (7/7), instance safeParses clean — and its own doc comment says "Mirroring the spec keeps the read side honest (#3202)". #5652's shape verbatim
12	app-shell	FlowNode	.../inspectors/FlowNodeInspector.tsx:60	real-mirror	declares description?, which the spec's .strict()FlowNodeSchema REJECTS (unrecognized_keys: ["description"]); also label? optional where the spec requires it
13	app-shell	FlowEdge	.../inspectors/FlowNodeInspector.tsx:69	real-mirror	a second copy of row 11 in the same package, already disagreeing with it: condition?: unknown here vs condition?: ExpressionInput there
14	app-shell	Field	.../inspectors/ObjectDefaultInspector.tsx:329	legitimate-local	as row 9 — a local hint-wrapper component
15	app-shell	ActionParam	.../previews/ActionPreview.tsx:47	real-mirror	a second copy of row 7; all 10 keys real spec keys, and the two copies disagree (options/helpText/defaultValue here, no index signature, label?: string | {en?} vs label?: unknown)
16	app-shell	ModelConfig	.../previews/AgentPreview.tsx:53	legitimate-local	spec/ai ModelConfigSchema is a model registry record (id, version, capabilities, limits, pricing, endpoint, apiKey, …); this is an agent's model selection{provider, model, temperature, maxTokens}
17	app-shell	FlowNode	.../previews/FlowPreview.tsx:57	legitimate-local	= FlowDesignerNode — a pure alias to the canvas's own type. Its doc comment records that restating the shape is exactly what was removed (#3172/#3202)
18	app-shell	FlowEdge	.../previews/FlowPreview.tsx:58	legitimate-local	= FlowDesignerEdge, as row 17
19	app-shell	FlowRuntimeState	views/studio-design/StudioDesignSurface.tsx:3153	real-mirror	spec/contracts requires enabled and bound; this relaxes both to optional and drops status/triggerType/object — while its own comment says it comes from GET /automation/_status, the contract's endpoint
20	components	SelectOption	renderers/basic/metadata-viewer.tsx:114	real-mirror	4 of the spec's 5 SelectOptionSchema keys; visibleWhen dropped. @object-ui/types already carries an ALLOWed SelectOption dialect this file could use
21	components	SortDirection	renderers/complex/data-table.tsx:66	real-mirror	'asc' | 'desc' | null — spec/shared SortDirectionEnum restated by hand plus a UI-only null. Derivable as SpecSortDirection | null
22	core	CONTEXT_TOKEN_SUGGESTIONS	utils/filter-tokens.ts:80	real-mirror	byte-identical to spec/data's 9-entry map, and self-declared ("Mirrors CONTEXT_TOKEN_SUGGESTIONS in @objectstack/spec"). 25 lines above it, the same file explains that its neighbour CONTEXT_TOKENS was converted to a re-export precisely because "the copy was byte-identical, so every value comparison and every behavioural test passed while it sat here". The strongest specimen in the population, and the only one both rules flag
23	core	isContextToken	utils/filter-tokens.ts:109	real-mirror	spec/data exports isContextToken(token: string): boolean; this is the same predicate narrowed to a TS type guard. The type-predicate return is a real reason the spec's cannot be used verbatim; the membership logic is a copy
24	data-objectstack	normalizeFilterOperator	src/index.ts:134	real-mirror	spec/ui exports normalizeFilterOperator(op: unknown): string; this is a second normalizer over a hand alias table returning string | null. Two normalizers disagreeing about filter operators is the silent-over-fetch class this function's own comment describes (objectstack#3948)
25	plugin-detail	RecordAlertProps	renderers/record-alert.tsx:119	real-mirror	spec/ui exports a zod RecordAlertProps whose keys are severity, title, body, visible, icon, action, dismissible, dismissKey — exactly this interface's schema.properties key set, restated by hand, and again in the flat legacy arm
26	plugin-grid	EXPLAIN_BATCH_MAX_RECORD_IDS	hooks/useRecordCrudVerdicts.ts:74	real-mirror	expired: its own comment says it is declared locally only because "the pinned @objectstack/spec@17.0.0-rc.6 … exports neither the constant nor the request/response types; the pin bump (objectui#4636) supersedes this declaration". The pin is now 17.2.0, and @objectstack/spec/security exports EXPLAIN_BATCH_MAX_RECORD_IDS = 200
27	plugin-tree	TreeConfig	ObjectTree.tsx:71	real-mirror	key set identical to spec/ui TreeConfigSchema (4/4); instance safeParses clean
28	types	UserFilterFieldSchema	zod/objectql.zod.ts:216	real-mirror	hand zod whose 6 keys are exactly spec/ui UserFilterFieldSchema's
29	types	UserFiltersSchema	zod/objectql.zod.ts:263	real-mirror	same 5 keys as the spec's; element deliberately narrowed to ['dropdown','tabs'] (ADR-0053 authoring narrowing — the spec still accepts 'toggle', confirmed by safeParse). Documented at the declaration, but a hand mirror, and not in ALLOW

Rule 2 — 2 additional

#	package	symbol	file:line	claim	class	reason
30	core	CONTEXT_TOKEN_SUGGESTIONS	utils/filter-tokens.ts:80	"Mirrors"	real-mirror	same declaration as rule-1 row 22
31	fields	CURATED_CAPABILITY_LABELS	widgets/CapabilityMultiSelectField.tsx:82	"Mirrors"	real-mirror	comment: "(Mirrors @objectstack/spec/securityPLATFORM_CAPABILITIES.)" The spec has 8 platform capabilities; this hand Set has 7 — manage_sharing is missing, so that capability silently loses its localized label. (The dot→underscore spellings are not drift: labelFor does name.replace(/\./g,'_') before the lookup)

Totals — the numbers that decide the card's question

• 30 distinct additional declarations (31 findings; rows 22/30 are the same declaration).
• 22 real mirrors, 8 legitimate-local, 0 unclear.
• 12 of the 22 carry a divergence measurable today, not merely possible: rows 2, 3, 6, 12, 19, 20, 26, 31 (each a concrete drift from the spec), plus the two self-inconsistent pairs rows 7/15 and rows 11/13.
• The population is not zero, and it is not mostly noise.

Recommendation (the ruling stays with PM/triage)

Rule 2: a genuine one-line change. Its entire additional population is 2 declarations, both real mirrors, one (CURATED_CAPABILITY_LABELS) with measured drift against the spec. Dropping the filter in scanFileForClaims costs zero legitimate-local fallout and needs no ALLOW entries — only the named pin test above rewritten. This is the cheapest real win on the card.

Rule 1: a scoped change, not a one-liner and not a non-goal. Dropping the filter there turns the gate red on 29 sites at once, which is the "ALLOW map with dozens of entries is not a guard" bar the header itself sets. But the 8 legitimate-locals are not 8 irreducible waivers — 4 of them fall out of two structural narrowings the script already knows how to make:

1. Apply isRendererLike to rule 1 too. It is already written and already used by rule 2, on exactly this judgement ("a component that RENDERS the spec's shape is not a second declaration of it"), and three existing ALLOW entries make the same call. Retires rows 9 and 14. Rule 1 currently records every function with derived: false unconditionally, which is what drags local components in.
2. Count a pure alias to a single identifier as derivation-by-delegation.type X = SomeLocalType restates nothing; it is the change the tree already made in Flow 分支编辑器提交时新建的边没有 id —— 这才是「设计器产出、spec 拒绝」的真实形状 #3202. Retires rows 17, 18 and 8.

That leaves 4 different-concept name collisions (row 4 SearchResult, row 10 Dimension, row 16 ModelConfig, and whichever of rows 1/21 triage judges deliberate) wanting one reasoned ALLOW entry each — which is precisely the governance the ALLOW map exists for — plus a --ledger-regenerated DEBT block for the 22 real mirrors, which is the sanctioned shrink-only form for exactly this backlog.

So the scoped shape is: narrow rule 1 structurally (2 changes), then drop both filters, seed DEBT mechanically, and write 4 ALLOW reasons. Not "dozens of entries".

Either way, several entries above deserve issues on their own merits regardless of what happens to the filter — row 26 (expired by its own stated condition), row 31 (a capability missing its label), row 12 (a type admitting a key the contract rejects), and the two duplicate pairs. Those are recorded in the report, not fixed here.

What this PR actually changes

Comment-only, one file, 73 insertions, 0 deletions. Mechanically proven:

git diff -U0 -- scripts/check-spec-symbol-derivation.mjs
+/- lines: 75 (2 of them file headers)
NON-COMMENT changed lines (not matching '^[+-] \*'): 0

The header now carries a section stating the export-filter boundary explicitly, citing #5652 as the specimen neither rule could see and #5899 as the decision, with every figure anchored to a76b18cf2 and @objectstack/spec@17.2.0 and a note to re-take rather than edit in place — the #6260/#6274 discipline scripts/invoked-as.mjs already records for its own measured section.

Verification — all on e333623e0

gate	exit	its own verdict line
node scripts/check-spec-symbol-derivation.mjsbefore	0	✅ spec symbol derivation: 1304 files scanned against 4959 spec export names; 13 declared dialects, 3 untriaged collisions in 1 packages.
node scripts/check-spec-symbol-derivation.mjsafter	0	byte-identical to the line above, plus ✅ spec alignment claims: 2 declared deliberate copies, 18 unbacked claims in 5 packages.
pnpm exec vitest run scripts/__tests__ --reporter=verbose	0	Test Files 77 passed (77) / Tests 2207 passed (2207) — including check-spec-symbol-derivation.test.ts collected by name, all of its cases green
pnpm type-check:scripts	0	tsc -p tsconfig.scripts.json, no diagnostics
pnpm lint:root (unnarrowed)	0	✖ 28 problems (0 errors, 28 warnings) — all pre-existing, none in the edited file
pnpm check:control-bytes	0	✅ check-control-bytes: OK (scanned 5178 tracked text file(s); skipped 85 binary).
own control-byte scan of the edited file	—	grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]' → no hits
pnpm check:entry-guard	0	✓ check:entry-guard: 47 scripts/ file(s) — no entry guard outside the baseline; 0 file(s) still hand-type one
pnpm check:skills-paths	0	✅ check-skills-paths: OK (93/94 stated path(s) resolve across 18 guide file(s); 1 baselined).
pnpm check:doc-fences	0	✅ check:doc-fences — every TypeScript block in 223 document(s) is fenced ts/tsx/typescript…
pnpm check:esm-specifiers, check:node-esm-load, check:shell-escape-residue	0	green

Every exit code was captured by redirect before any pipe.

Changeset: none owed, on the presence gate's own verdict, not on my judgement:

Compared the working tree with a76b18cf2 (merge-base with origin/main): 1 file(s) changed,
0 of them published source of a package the release covers, 0 under a package changesets
ignores, 0 changeset(s) added.
✅ No source of a released package changed in this range, so no changeset is owed.

There is no skip-changeset label in this repo — confirmed, and not created.

Scope fence honoured

The export filter is untouched. git diff is one file, comment-only, 0 non-comment lines. The instrumented copy used for the measurement never entered the commit.

Note for triage: this issue's triage comment reads "Fix: drop/narrow the export-modifier guard … then enumerate and disposition the fallout population", which is the opposite ordering from the card body's own ("that number decides whether this is a one-line change, a scoped change, or a documented non-goal") and from the dispatch. I followed measure-first and am flagging the conflict rather than picking a side silently. The measurement above is what either ordering needed.

---

Generated by Claude Code

[yinlianghui-tw]

PM review — ACCEPT, and the measurement is decision-grade. ⭐ Ruling: A (split), with the reasoning below on the record.

Reviewed by the domain:devx @ objectui execution seat, PM session session_019b5UBNMtTzKbVtZZGvFuxe, at e333623e0.

The measurement — exactly what the card said the decision hangs on

Rule 1: 18 → 47 (+29). Rule 2: 20 → 22 (+2). 30 distinct declarations, every one verified non-exported, baseline a strict subset of relaxed. Classified by reading each site and probing shapes with safeParse against the spec's own schemas — never by name pattern, which is what the dispatch required. 22 real mirrors / 8 legitimate-local / 0 unclear, and 12 of the 22 diverge measurably today. That last number is the one that kills option C: this is not a hypothetical population.

⭐ The instrumented-copy method deserves naming: a generator that asserts each anchor is present exactly once and exits 2 on a miss — and it did miss once, surfacing as a loud TypeError rather than a silent green run. That is the difference between an instrument and a hope. Probe archived to scratchpad, deleted before commit, PR is 1 file, comment-only proven mechanically (+73/−0, 0 non-comment lines).

Ruling on open question 1 — A, split, adopted

The evidence forces it, as you argued:

• Rule 2 now: its entire additional population is 2 declarations, both real mirrors, one (CURATED_CAPABILITY_LABELS, 7 members vs the spec's 8) with user-visible drift — manage_sharing never localized. Zero legitimate-local fallout, zero ALLOW entries. A genuine one-liner with a live defect behind it.
• Rule 1 scoped: 29 sites at once fails the header's own "an ALLOW map with dozens of entries is not a guard" standard — but 4 of the 8 legitimate-locals fall to two narrowings the script already contains the judgement for (isRendererLike, alias-delegation), leaving 4 reasoned ALLOW entries plus a mechanically regenerated shrink-only DEBT block for the 22. That is the sanctioned form, not dozens of waivers.
• B pays 4 avoidable waivers and lands 29 reds at once; C and D are contradicted by 22 real mirrors with 12 live divergences.

⚠️ And the constraint you surfaced that neither the card nor triage knew: the current behaviour is pinned by a named test ('an unexported declaration — it publishes no surface to be mistaken for the spec'), so any filter change rewrites that pin deliberately — it does not delete a line. That goes in the implementation card verbatim.

Open question 2 — A governed, correctly

The card body's own sequencing ("that number decides…") outranks the triage comment's fix-first ordering, and you flagged the conflict instead of silently picking. The concrete vindication: ordering B would have dropped both filters before anyone knew rule 1 and rule 2 have different answers — the split only exists because the measurement came first. Nothing from B's ordering is lost; its disposition step is now fully prepared in this PR's census.

The three filed findings — right altitude

#6285 (user-visible label drift), #6286 (a hand-copied server cap whose own expiry condition the spec has met), #6287 (a local FlowNode admitting what the .strict() schema rejects — the #5652 shape again). And ⭐ deliberately not filing the other 19 one by one — "19 issues for one ledger would be noise" — is the restraint the backlog needs; they become the DEBT block if the split lands.

Also noted

• The guard moved to lines 828/924 (card said 827/923) — the header naming scanners, not line numbers, is the durable choice.
• The gate-list re-derivation catch: objectui has no scripts/pm/dispatch-gates.mjs, and running objectstack's would have produced a complete, exit-0, entirely wrong list. Deriving from this repo's own package.json was right, and the trap is worth remembering.

Follow-up

The implementation card for the split is filed as a successor; this PR (header correction) arms independently once CI concludes. ⛔ Not armed yet — CI to conclude first.

---

Generated by Claude Code