fix(spec): the Seed.locale liveness note names the six seed-load builders and the locale each supplies - #30666

Requested by Sales · project thread

domain:spec#2 · os-sales · session_01L6bhgVugtCCwBs98FeR27P

Fixes #30607
Clause-②: no
Metadata-protocol: none

维护者速读

• 改了什么:packages/spec/liveness/seed.json 里 Seed.locale 这一行的 note 重写为六个种子加载请求构建点各自是否带语言、语言从哪来;verifiedAt 改为 2026-11-05;另加一个 @objectstack/spec 的 patch changeset。
• 为什么改:PR feat(runtime,cloud-connection,organizations): the marketplace install, sample-data reseed, restart heal and per-organization replay pick seed datasets by the organization's language #30346 之后,市场安装、样例数据重灌、重启自愈和按组织回放都按组织语言挑选种子数据,台账却还写着"市场安装不传语言、安装是语言中立的",与代码和 SeedSchema.locale 的 describe(PR docs(spec): Seed.locale says which loads filter on it, and the ownership-queue contract says the registered service honours onFailedRead refuse #30604)相反。
• 风险与回滚:只改台账文字与日期,status 仍是 live,无代码、schema 或导出变化;回滚即 revert 本 PR。
• 先看:① packages/spec/liveness/seed.json 第 53 行 note 中 "THE SCOPE OF THIS live" 一段(六个构建点清单)与紧随 Seed locale scope is silently inert on the three publish/install-time seed-load paths — package apply, draft publish and marketplace install pass no config.locale #17011 裁决原文之后新增的 "NARROWED 2026-11-05" 一句;② .changeset/30607-seed-locale-liveness-note.md。
• 两分钟验证:pnpm --filter @objectstack/spec check:liveness,预期退出码 0,末尾两行均为 ✓。
• 需要你做的:无,席位落地。

What changes

packages/spec/liveness/seed.json, the props.locale row: note and verifiedAt only (2 lines). status (live), evidence, producer, evidenceScope, every other row and the file's _note parse identical before and after. Plus one @objectstack/spec patch changeset.

The builders, counted from the code at b8929f48. Production SeedLoaderRequestSchema parses (tests excluded) give six request builders, the same number the old note gave. A control scan of every new SeedLoaderService( finds three more constructions (marketplace-install-local-plugin.ts#sampleDataInScope, #handlePurge, #handleSampleDataCount). All three call only buildDependencyGraph, never load(), so none of them is a request builder.

The note, FROM → TO, by sentence (every sentence not listed is unchanged byte for byte):

In the quotes below, PR #30346's closing line is spelled in words so that this body points a closing keyword only at this card. The note itself writes it in code.

1. FROM "⛔ It does NOT assert that every path which can reach SeedLoaderService.load() supplies a locale — three do not, they are named below, …"
   TO "… — two do not, they are named below, …"
2. FROM "Wire config.locale at the runtime's seed-load call sites — the locale axis added by #16510 is inert on the default boot path until a producer supplies it #16595 wired the THREE in packages/runtime/src/app-plugin.ts — the default boot path, which is what makes authoring locale change runtime behaviour and therefore what makes this row live."
   TO "FOUR supply a locale, each named by the function that builds its request:" followed by builders (1)–(4) above, each with its path#function and its locale source, then "Wire config.locale at the runtime's seed-load call sites — the locale axis added by #16510 is inert on the default boot path until a producer supplies it #16595 wired (1)–(3) in packages/runtime/src/app-plugin.ts — the default boot path, which is what makes authoring locale change runtime behaviour and therefore what makes this row live — and PR feat(runtime,cloud-connection,organizations): the marketplace install, sample-data reseed, restart heal and per-organization replay pick seed datasets by the organization's language #30346 (the PR that closed card 30281 [spelled in code in the note], merged 2026-11-05) moved (3) onto resolveSeedLocaleForLoad, keeping resolveSeedLocale as its last rung, and wired (4). A builder whose source answers nothing sends no locale key, never an 'en' guess."
3. FROM "The other three are publish/install-time paths and still pass no locale: …/packages.ts#applyPublishedSeeds (package apply), …/protocol.ts#applySeedBodies (draft publish) and …/marketplace-install-local-plugin.ts#runInlineSeed (marketplace install)."
   TO "TWO are publish-time paths and still pass no locale: (5) packages/runtime/src/domains/packages.ts#applyPublishedSeeds (package apply) and (6) packages/metadata-protocol/src/protocol-publish-batch.ts#applySeedBodies (draft publish — the body packages/metadata-protocol/src/protocol.ts#applySeedBodies delegates to; the request moved out of protocol.ts with it)."
4. FROM "None of the three is handed the stack config, and none passes env either, … and a locale has no ambient source, so it cannot be closed the same way."
   TO "Neither is handed the stack config, and neither passes env either, … and a locale has no ambient source inside load(), so it cannot be closed the same way: the four builders that supply one resolve it at the call site."
5. The paragraph "RULED 2026-09-10 on Seed locale scope is silently inert on the three publish/install-time seed-load paths — package apply, draft publish and marketplace install pass no config.locale #17011 (director seat, one-class self-adjudication, option A 「不扩散」): … So this line is a DECISION, not a to-do." is unchanged: 653 characters, the same bytes, found once in the new note. ADDED right after it: "⚠️ NARROWED 2026-11-05 — the ruling above is kept as it was recorded, and its "three paths" and "three sites" are the 2026-09-10 set, marketplace install included. A maintainer ruling on Seed.locale: marketplace install, reseed sample data and per-org replay should select datasets by the installing organization's locale #30281 (2026-11-05, option A, reversing Seed locale scope is silently inert on the three publish/install-time seed-load paths — package apply, draft publish and marketplace install pass no config.locale #17011's install-path bound) ordered the install paths to pick datasets by locale, on HotCRM's request for zh-CN demo data rather than through the measurement the premise below names; PR feat(runtime,cloud-connection,organizations): the marketplace install, sample-data reseed, restart heal and per-organization replay pick seed datasets by the organization's language #30346 (the PR that closed card 30281 [spelled in code in the note], merged 2026-11-05) carried it out and moved the marketplace-install path, builder (4) above, across that bound. The ruling's bound now covers package apply and draft publish only — builders (5) and (6) — and SeedSchema.locale's describe (PR docs(spec): Seed.locale says which loads filter on it, and the ownership-queue contract says the registered service honours onFailedRead refuse #30604) and that docs page state it narrowed the same way."
6. FROM "It is the live diagnostic for the three call sites above, for any embedding host that builds its own request, and for a stack that declares no i18n block — all of which …"
   TO "It is the live diagnostic for the two call sites above that pass none (package apply, draft publish), for any embedding host that builds its own request, for a stack that declares no i18n block (the boot seed and the dev hot-reload seeder have no other source), and, since 2026-11-05, for an install, reseed, heal or replay on which no rung of resolveSeedLocaleForLoad answers — all of which …". This reason was checked against the code: SeedLoaderService#warnOnUnresolvedLocaleScope returns early only when config.locale is set, and loadDatasets calls it on every load. resolveSeedLocaleForLoad returns undefined when no rung answers, and both of its callers then spread no locale key.
7. ADDED at the end, after the "REPOINTED while flipping" paragraph: "RE-COUNTED 2026-11-05 against b8929f4 by the same scan of production SeedLoaderRequestSchema parses: still SIX builders, the six named above, with the draft-publish request now built in protocol-publish-batch.ts. A control scan of every new SeedLoaderService( finds no seventh: its three other constructions, in the marketplace plugin's sample-data scope, purge and count reads (sampleDataInScope, handlePurge, handleSampleDataCount), call only buildDependencyGraph and never load()."
8. verifiedAt: 2026-09-09 → 2026-11-05, the YYYY-MM-DD form the sibling rows use.

Verification

All local readings are on head 0cd1455d86 (branch point b8929f48).

• Liveness gate, pnpm --filter @objectstack/spec check:liveness: exit 0. Its two closing lines are both ✓: every property is classified and every repo-local evidence / producer pointer, path:NNN line and path#symbol anchor resolves; state-counts/ is current. Totals are unchanged: 1172 live · 1 experimental · 83 dead · 42 planned.
• Derived families, node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, over 2 paths (+14 / −2 lines): 55 commands, all run with exit codes captured before any pipe. 55 × exit 0.
  • The first pass, before a build, had 51 × exit 0 and 4 × exit 3 (PREREQUISITE NOT MET, no dist/): check:dts-closure, check:dual-build-cjs-loads, check:lean-entry-closure and check:sourcemap-no-sources-content.
  • The full workspace build then ran under the verify lock: pnpm build --concurrency=2, 118/118 tasks, VERDICT command-exit 0. The working tree was still clean afterwards.
  • The whole sweep was then re-run on the built tree, and that second run is the table below. The build-reading verdicts it records: 295/295 declared .d.ts present across 115 packages; 116 require entry points load; 2 lean conditions are measured from a real load; 1276 source maps embed no source text.
• Reconciliation, dispatch-gates.mjs --ran: "✓ 55 derived famil(ies) accounted for — 55 run, 0 NOT-MEASURED (a DERIVED zero — all 55 recorded an exit code and none of them is 3)". 0 UNRUN.
• main moved during the run (47c457b23c, 9 commits). None of them touches seed.json or any file the note cites. The builder recount at that tip is the same six SeedLoaderRequestSchema parses and the same nine SeedLoaderService constructions.
• CI-only families (artifact rosters, wide populations, the path-scheduled jobs, the type-check lanes) are not local readings. They are this PR's check-runs.

Acceptance notes

• No pin owed. The liveness gate does not read note. It checks statuses, evidence / producer pointers (path:NNN lines, path#symbol anchors, key mentions) and the verifiedAt format and age. No test pins the note's text: git grep for its phrases outside seed.json finds only spec(seed): the Seed.locale docblock and .describe() say installing is locale-neutral, but the install paths now pick by the organization's locale #30347's changeset and the docs page, and they use their own wording. Every path#symbol the new note names was checked to exist in its file at b8929f48.
• Changeset, not skip-changeset. liveness is in @objectstack/spec's package.json files, so seed.json ships in the tarball, and Check Changeset requires a changeset when the PR has no skip-changeset label. A patch changeset with both declaration lines, as fix(spec): repoint the query.searchFields liveness row at engine-find.ts#expandSearchOnAst #30368 and fix(spec): the page liveness ledger grades the app page record binding live, citing its console readers #30611 did for liveness-only edits.
• Where the card's paraphrase and the code differ (the note follows the code): the draft-publish request is built in protocol-publish-batch.ts#applySeedBodies, and protocol.ts#applySeedBodies delegates to it. The sample-data reseed and the restart heal live in marketplace-install-local-plugin.ts (both go through runInlineSeed), not in app-plugin.ts. There are two per-org replayers: AppPlugin's, and cloud-connection's buildSeedReplayer, which is registered only where no replayer exists yet.
• False sentences found outside this PR's file surface, not fixed here:
  • The same row's producer field says resolveSeedLocale is threaded "into all THREE SeedLoaderRequests this plugin builds: the inline boot seed, the per-org replayer registered for tenant provisioning, and the dev hot-reload seeder". Since PR feat(runtime,cloud-connection,organizations): the marketplace install, sample-data reseed, restart heal and per-organization replay pick seed datasets by the organization's language #30346 the per-org replayer passes resolveSeedLocaleForLoad's answer and uses resolveSeedLocale only as its last rung (packages/runtime/src/app-plugin.ts, the replayer closure in start).
  • The file-level _note cites packages/runtime/src/app-plugin.ts:832, :971 and packages/metadata-protocol/src/protocol.ts:6764 for the seed-load calls. At b8929f48 those lines are a closing brace, a datasource-routing docblock line and a search-403 docblock line. The calls are now at app-plugin.ts L1856 / L1951 / L2376, and the publish apply has moved to protocol-publish-batch.ts. The gate bounds path:NNN only in evidence / producer, so this has never been checked.
• Kept as recorded, not re-measured: the note's "EVIDENCE FOR THE FLIP" paragraph (an ablation of resolveSeedLocale turned 5 of 7 cases red, covering all three AppPlugin call sites) is the 2026-09-09 reading. The per-org replay now resolves through resolveSeedLocaleForLoad, so a fresh ablation could read differently. No ablation was run here, because this PR changes no tested behaviour.

🤖 Generated with Claude Code

https://claude.ai/code/session_01L6bhgVugtCCwBs98FeR27P