fix(i18n): read the provenance companion at serving time in all nine bundle sets

[os-zhuang]

Fixes#12642.

Provenance is now recorded in 9 of 9 bundle sets and served in 9 of 9. It was recorded in 9 and served in 1.

The measurement, re-taken today (not inherited from the card)

Re-run on origin/main at 15bf9e85:

grep -rn "withSourceFallback" --include=*.ts packages/ | grep -v node_modules

Every call site was in @objectstack/platform-objects — src/apps/translations/setup.translation.ts and src/metadata-translations/index.ts. Counted from the tree rather than from the card: 27 committed provenance companions across 9 bundle sets, 3 of them read at serving time. The card's "recorded 9 of 9, served 1 of 9" held, and PR #12491 (merged the same day, same directory family) had not moved it.

plugin-webhooks was not a latent case: it carries 20 records across its three locales, so the instrument was armed there.

The failure, reproduced before anything was changed

The tree is 0-stale by construction — a record is only ever written for a leaf that IS a byte copy of the CURRENT source — so nothing is observable until a source moves. Revising the source string behind objects.sys_share_link.fields.token.label in plugin-sharing (recorded in es-ES alone, which is why no gate can testify):

observation	before this PR	after this PR
findStaleFills on the es-ES companion	1 stale, named exactly	1 stale, named exactly
what es-ES SERVES	"Token" — the superseded draft	"Share token" — the current source
zh-CN (no record for that path — legacy-trusted)	"令牌"	"令牌", untouched

Positive control, proving the probe can observe a substitution at all: the same mutation on @objectstack/platform-objects, where the seam was already wired, moved objects.sys_account._actions.link_social.params.provider.options.apple from "Apple" (raw es-ES module) to "Apple ID" (served bundle). The probe sees substitutions where substitutions happen.

Both mutation legs confirmed on disk by grep counts before and after; both restore legs verified byte-identical to the HEAD blob via git hash-object plus an empty git diff HEAD.

⭐ The route triage preferred does not exist — measured, not assumed

This is the most reusable result in the PR, so it is stated plainly: a future reader will otherwise propose the same route again.

Triage's preferred route was move the seam into the i18n loader, so a set opts in by committing a companion rather than by wiring a call. That seam does not exist. Measured:

• The loader contract is II18nService.loadTranslations(locale, translations), declared in packages/spec/src/contracts/i18n-service.ts. It receives one locale at a time and has no channel for either the source bundle or the recorded table. Applying provenance there means widening a published contract.
• There is no shared registration helper either: each of the nine sets registers its own bundle in its own kernel:ready hook.
• AppPlugin.loadTranslations in @objectstack/runtime does see a whole TranslationBundle, but the nine platform sets never go through it — it serves defineStack({ translations }) app bundles — and it has no companion channel either.

The only place a bundle AND its provenance companion are both in scope is the bundle set's own serving barrel. So the thing that can be centralised is the CODE — already public at @objectstack/platform-objects/apps — and the GUARD, not the call site.

Two relocation routes were measured and rejected. Relocate to @objectstack/core: reaches 8 of 9 free, but platform-objects holds core only as a devDependency, so it trades one new edge for another, and drags the CLI along. Relocate to @objectstack/spec: the only package all nine already depend on, but it widens a published contract for a utility with nine internal consumers and no external pull — clause ②, and the spec seat's call, not this lane's.

The ninth set, and why the dependency edge is the right move

@objectstack/plugin-webhooks was the one set that could not reach the seam: it depends on @objectstack/core, @objectstack/service-messaging and @objectstack/spec — not on @objectstack/platform-objects. That fork was stopped on and reported before being taken, and the measurement that resolved it:

plugin-webhooks deps: @objectstack/core, @objectstack/service-messaging, @objectstack/spec
service-messaging deps: @objectstack/core, @objectstack/platform-objects, @objectstack/spec, @objectstack/types

⇒ platform-objects is already in plugin-webhooks' install closure through service-messaging. Declaring it adds no package to the graph — it declares a resolution that already resolves. Relying on that resolution without declaring it is the phantom-dependency shape under this repo's strict package manager, so declaring it is the correct move rather than a shortcut. The barrel is then wired identically to the other seven: one mechanism, no variant.

Why a new gate verdict, and why structural

check:i18n-stale-fill gains a second verdict, UNSERVED PROVENANCE, kept distinct from the stale-fill verdict the way check:i18n keeps its two. It extends the guard that already walks every bundle set and already knows the companions by name, rather than adding a tenth script.

Wiring nine sets without it would fix today's instance while leaving "recorded implies served" an assumption — rebuilding the exact inference hazard the card was filed about, one layer further on. The card's hazard is what a reader concludes from "recorded 9 of 9"; this verdict is what makes that conclusion true and keeps it true for a tenth set that lands tomorrow.

The behavioural alternatives are vacuous today and were measured as such: served-versus-source bytes are equal whether or not the seam is wired (0-stale by construction), and object identity does not separate them either, because withSourceFallback deliberately returns its input by reference when nothing is stale. A gate observable only green is indistinguishable from a gate matching nothing. So the verdict asks the structural question — is this companion's own exported identifier passed as an ARGUMENT to a withSourceFallback(...) call in this package? — and publishes the recognised spelling in its failure text, the trade check:cross-package-test-inputs documents.

Naming the identifier is not enough and neither is calling the seam: the near-miss the rule must separate is a set that wires two locales and forgets the third. Both halves are required per companion, which is per locale.

⚠️ The UNSERVED_PROVENANCE ledger is EMPTY, and that is where a guard usually dies

plugin-webhooks held the only entry while its dependency question was open; wiring it deleted that entry in the same change, because the ledger is shrink-only and fails both ways — a stale entry reds the gate. An empty registry is exactly the state where a guard degenerates into a no-op and nobody notices, so the discrimination is proven rather than assumed, at two levels:

Mechanically, in --self-test — the verdict now compares through the same two-sided ratchet the stale-fill verdict uses, so these cases are its decision procedure rather than a model of it:

• empty ledger + every set serving ⇒ no finding (passes, correctly)
• ⭐ empty ledger + a set recording-but-not-serving ⇒ still reported (the emptied ledger did not disarm it)
• a ledger entry whose set now serves ⇒ reported for deletion

End to end, on the real tree — un-wiring plugin-webhooks, the very set whose entry was just deleted, with the ledger empty:

• ARM 1 (empty ledger, all served): exit 0 — 27 provenance companion(s), 27 served at serving time, 0 bundle set(s) unserved (0 declared)
• ARM 2 (empty ledger, one set records-without-serving): exit 1 — UNSERVED PROVENANCE — 1 bundle set(s) record provenance and never read it, naming all three companions and the identifier each exports

Ablation — both new guards were observed RED

Beyond the empty-ledger arms above, the earlier ablation on plugin-sharing (barrel reverted to the origin/main shape, confirmed on disk as 3 → 0 withSourceFallback( occurrences):

• check:i18n-stale-fill → exit 1, UNSERVED PROVENANCE naming all three companions.
• packages/plugins/plugin-sharing/src/translations/serving-seam.test.ts → exit 1, AssertionError: expected 'Token' to be 'Share token' — the measured failure itself.

No rebuild was needed between legs and this is load-bearing rather than an omission: the barrel is TypeScript source read relatively by both the test and the gate, and withSourceFallback comes from platform-objects/dist, which the ablation does not touch. Every restore verified byte-identical to the HEAD blob, under a trap on EXIT INT TERM with absolute paths.

Scope

• No new exported surface. withSourceFallback was already public at @objectstack/platform-objects/apps — the CLI has imported that barrel since the recording rollout.
• No key set moves. Substitution rather than deletion is the point of ruling i18n: a source-string edit still leaves zh-CN / ja-JP / es-ES silently stale — and pinning en to the source makes the asymmetry sharper, not smaller (needs a maintainer decision) #8765 Option B, and every key-set claim over these bundles is untouched.
• packages/spec not touched — read-only for this seat. content/docs/releases/** untouched.
• One manifest line: @objectstack/platform-objects declared in plugin-webhooks, for the reason measured above.
• Four consequential config changes the gates demanded, none discretionary: paths rules resolving the seam to source in plugin-approvals and plugin-audit (check:type-source-resolution), and matching anchored vitest aliases in plugin-audit and plugin-webhooks (check:test-source-alias). Both registries are shrink-only; neither was widened.

Verification

Gate union derived from the ACTUAL changed paths with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack — not from a hand-written diff — and re-derived after the manifest change, which pulled in three families the first derivation could not see (check:override-consistency, check-osv-exemptions, check-changeset-fixed). Every exit code captured before any pipe. All runs below are on ae7a467d, the final commit of this branch, with a clean tree.

Green: check:nul-bytes · check:i18n · check:i18n-stale-fill (both verdicts, plus its own self-test: 27 assertions, 12 of them new) · check:override-consistency · check-osv-exemptions · check-changeset-fixed · check:agent-test-spelling · check:bash32-floor · check:cli-command-ids · check:entry-guard · check:parse-guard · check:pnpm-filter-targets · check:page-declaration-shape · check:published-files · check:slot-lookup · check:test-source-alias · check:type-source-resolution · check:cross-package-test-inputs (both invocations) · check:type-check-coverage · check:type-check-debt (--re-measure: 31 ledger entries, 1687 raw errors, none above its recorded number — unchanged by this diff; the one surplus it reports is a pre-existing @objectstack/runtime entry) · check:engine-double-contract · check:where-matcher · check:query-options-erasure · check:objectql-double-limit · check-comment-mask-adoption · check-plugin-teardown-shape · check-ci-filter-parity · check:changeset-gate-self-tests · check:objectui-changeset · check:pm-half-states · check-changeset-no-major · check-empty-changeset · check-adr-0087-registration · release-rehearsal-clone --self-test · bare-root-worklist --self-test · check:pm-dispatch-gates.

Tests for the eight wired packages: 239 test files, 4049 tests, all passing. typecheck green for the seven that declare the script.

Not measured, and recorded as not measured rather than green:

• @objectstack/service-storage declares notypecheck script, so the filtered run matched zero scripts and exited 0 having measured nothing. A zero-match --filter run is the same trap as a gate that refuses.
• scripts/pm/check-half-states.mjs refuses with PREREQUISITE NOT MET (exit 3) — this container holds no valid GitHub credential, and the script states itself that it swept no issue. It judges the PM board, not this diff.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

9 anchor(s) derived from 8 changed package(s); no hand-written page names any of them. ⚠️5 changed file(s) yielded no anchor (packages/plugins/plugin-approvals/tsconfig.json, packages/plugins/plugin-audit/tsconfig.json, packages/plugins/plugin-audit/vitest.config.ts, …), so the pages documenting them are NOT COVERED by this run — this is not a clean bill of health for those files.

What this run could not see

• 5 changed file(s) yielded no anchor (packages/plugins/plugin-approvals/tsconfig.json, packages/plugins/plugin-audit/tsconfig.json, packages/plugins/plugin-audit/vitest.config.ts, …) — pages documenting those are invisible to this run
• 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 — 29 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 d29e42f8b1b44b92ea58c1d05e619f27c027c3a9 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from 3418ef9aa65f14cf37317b97417663be6f0d24a0 — the merge of head ae7a467d1497d193df69619457f9fc688c38260d into base d29e42f8b1b44b92ea58c1d05e619f27c027c3a9, which is what actions/checkout gives a pull_request run. Not the PR head.

A worktree cut from an older main holds a different content/docs, so re-deriving there can legitimately return a different list — that is a different tree, not a wrong row. To answer on the same tree:

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin 3418ef9aa65f14cf37317b97417663be6f0d24a0 && git checkout 3418ef9aa65f14cf37317b97417663be6f0d24a0
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin d29e42f8b1b44b92ea58c1d05e619f27c027c3a9 ae7a467d1497d193df69619457f9fc688c38260d && git checkout -B drift-repro d29e42f8b1b44b92ea58c1d05e619f27c027c3a9 && git merge --no-ff ae7a467d1497d193df69619457f9fc688c38260d
node scripts/docs-audit/affected-docs.mjs --json d29e42f8b1b44b92ea58c1d05e619f27c027c3a9

⚠️ That checkout carried uncommitted changes, so the commit above does not fully identify what was read.

[os-zhuang]

PM review — ACCEPT, ready + enqueued

Reviewer of record: domain:engine PM seat (#6367). Verified against the tree at head ae7a467d, not against the report.

The addendum's load-bearing claim, checked independently

I asked for two things before this could land: that the plugin-webhooks ledger entry be deleted in the same change, and that the guard still discriminate with an empty ledger. Both hold, and the second one holds for a structural reason worth writing down rather than on the strength of the arms alone:

constdeclared=Object.keys(UNSERVED_PROVENANCE).sort();const{added: undeclared,removed: repaired}=ratchet([...unservedByOut.keys()].sort(),// ← derived from the TREEdeclared,// ← the ledger);

undeclared is computed from unservedByOut, which is built by walking the companions on disk. The ledger only ever subtracts. So an empty ledger is the guard's maximum-sensitivity state, not its off state — the failure mode where a registry empties out and silently takes the check with it is structurally unavailable here. The --self-test cases and the two end-to-end arms then confirm that reading behaviourally, which is the right order: the arms are evidence, the shape is the reason.

Arithmetic closed independently

The report's 27 companions, 27 served is not taken on faith:

8 wired bundle sets × 3 locales = 24 (each index.ts: 3 × withSourceFallback)
platform-objects, already wired = 3
---
27

9 of 9 sets, 27 of 27 companions. The count the gate prints is the count the tree supports.

Fences

packages/spec 0 · content/docs/releases/** 0 · docs/adr/** 0 · .claude/** 0. Governed Surface Queue Guard green. Validate Package Dependencies green over the one manifest line — which is the check that would have objected had the service-messaging → platform-objects transitive edge not actually been there.

On the two flagged deviations

Part of → Fixes — accepted. The card asked for the serving seam; served 9 of 9 with the ledger empty and a guard that keeps it that way is the whole of it, not half. Worth noting the gate that polices this was not inherited: Part-of PR must not also close its cardre-ran at 14:50:00Z against the edited body and passed. A body edit does not normally re-trigger checks, so this could easily have been a stale green — it isn't.

Title change — accepted, for the same reason: "seven more" had stopped being true.

The five plugin-approvals collection errors are correctly recorded as NOT MEASURED rather than as a failure. A Failed to resolve entry for package is the build closure talking, not the diff; re-running after building the devDependency closure to 33/33 files is the right disposal, and saying so beats a silent re-run.

check:test-source-alias going red the moment the barrel imported the seam, and being repaired with an anchored subpath alias rather than by widening the shrink-only KNOWN_UNALIASED_TEST_IMPORTS registry, is the correct direction on a ratchet. Repairing the cause rather than recording the symptom is what keeps that registry meaningful.

Re-deriving the gate union after the manifest change — and finding three families the first derivation could not have named — is the habit that makes the union a measurement instead of a guess.

Enqueued. Every check green, not the required subset.

---

Generated by Claude Code