spec(liveness): bring the plugin manifest into the ledger's governed universe

packages/spec/liveness/README.md

@@ -704,7 +704,7 @@ The governed set is `GOVERNED` at the top of `check-liveness.mts`. To add a type
  RecordDetailView had been gating the History tab on it the whole time (#2707).
 4. Add the type to `GOVERNED`; confirm the gate is green.
 
-## Current state — 30 governed types (complete registry coverage)
+## Current state — 31 governed types (complete registry coverage)
 
 > **This heading is now checked** (#7257). `check:liveness` reconciles the table
 > against `GOVERNED` in both directions — a governed type with no row fails, a row
@@ -816,6 +816,7 @@ marker where the Notes cell goes, never a guess at what belongs there.
 | validation | seeded 2026-08-01 (#4488). The ADR-0020 carrier: the evaluator honors active/events/priority/severity/type/condition/message (the zod header's "only reads type/condition/…" prose is STALE — trust the ledger). Dead 3 = label/description/tags, declared governance metadata, kept unmarked. Union walk boundary recorded: only base + `script` keys walked; per-variant keys are governed by the evaluator's tests, not ledger rows. **No longer a registered metadata kind** — #4509 retired it under ADR-0088 (a standalone rule had no object-binding key and every variant is `.strict()`, so it bound to nothing and gated no write; a state machine authored that way saved cleanly and did nothing). The rule VOCABULARY is untouched and fully live via `object.validations[]`, so the ledger keeps governing it through the gate's spec-only override, alongside `webhook` and `query`. The contrast with the two bridges in the same batch is the point: enforce-or-remove picked ENFORCE where the feature existed and only the wiring was missing, and REMOVE where the shape itself could not carry the feature |
 | api | seeded 2026-08-04 (#5271, part of #5206; PR #5312) — **not a metadata type until that same change made it one**, which is the row's point: governance and registration landed together, the treatment `datasource` did not get (#4487) and paid for with six inert keys found by hand. What #5206 measured before the fix: `api` was in neither `DEFAULT_METADATA_TYPE_REGISTRY` nor `BUILTIN_METADATA_TYPE_SCHEMAS`, so `saveMetaItem`'s `resolveOverlaySchema('api', …)` → `getMetadataTypeSchema('api')` returned `undefined` and took its own documented branch — an unregistered type is stored **unvalidated** — while `getMetaTypes()` could not enumerate the type at all, so Studio rendered neither list nor form. That issue names the shape precisely and it is the inverse of this ledger's usual one: **enforced but undeclared** (the matcher was already indexing these entries, #5089), where `dead` is declared-but-unenforced. The seeding pass classified 27 keys — live 25 / planned 2 / dead 0 — each cited `file:line` at the consumer layer that reads it: the MATCHER (`packages/metadata/src/endpoint-matcher.ts`) indexes `name`/`path`/`method`; the EXECUTOR (`packages/runtime/src/endpoint-executor.ts`) dispatches on `type` and reads `target`/`objectParams`; the POLICY chain (`packages/runtime/src/endpoint-policy.ts` + `security/inbound-rate-limit.ts`) enforces `authRequired`/`rateLimit`/`cacheTtl`; the MAPPING layer (`packages/runtime/src/api-mapping.ts`) applies `inputMapping`/`outputMapping`; and OpenAPI enrichment (`packages/rest/src/openapi-endpoints.ts`) emits `summary`/`description`. Timing was the reason it was cheap: #5040's E-series had built every one of those consumers and all of it was on main, so each key had a real evidence path rather than a promise. **Planned 2 = `inputMapping.transform` + `outputMapping.transform`, and `planned` rather than `dead` is load-bearing**: `dead` here means parsed with no consumer — a silent no-op — and these are the opposite, parsed and then LOUDLY REFUSED at publish (`endpoint-publish-gate.ts` mappingGate) and again at runtime, because no transformation-function registry exists anywhere in the platform. An author who writes one is told so and told what to do instead, so there is nothing for enforce-or-remove to chase; they stay in the vocabulary because admitting them needs a function registry **and** a sandbox ruling (#5040 §3.4), which is a design decision, not a key to quietly delete. Zero dead |
 | capability | seeded 2026-08-08 (#5961; PR #6540) — `CapabilityDeclarationSchema`, the DECLARATION side of ADR-0066 D1's three-way separation: packages DEFINE a capability, permission sets GRANT it via `systemPermissions`, resources REQUIRE it via `requiredPermissions`. **The gate's 12 and the seeding PR's 5 are the same measurement at two granularities** — PR #6540 call-graph-closed **5 authorable properties**, every one to a real reader in `packages/plugins/plugin-security/src/bootstrap-declared-capabilities.ts` (the one consumer that turns a declaration into a `sys_capability` row), all `live`, with no `PENDING_GOVERNANCE` debt recorded; the other 7 are the ADR-0010 protection-envelope keys the gate auto-classifies `live` and which carry `null` verdicts in the file, exactly as on `permission`/`position`. The same worked example as `api` above and PR #6540 says so in those words — **enforced but undeclared**, the mirror of the hole #5271 closed. What #5961 measured: absent from `DEFAULT_METADATA_TYPE_REGISTRY`, `BUILTIN_METADATA_TYPE_SCHEMAS` and `HAND_CRAFTED_SCHEMAS`, so `isRuntimeCreateAllowed()` took its no-static-entry fallback (permanently true) and `saveMetaItem` its no-schema branch — `PUT /api/v1/meta/capability/:name` accepted **arbitrary JSON** onto an authorization surface whose names `systemPermissions`/`requiredPermissions` resolve by string, while `/meta/types` synthesised a false `allowRuntimeCreate: true` descriptor Studio drew a raw-JSON create form from. #5870 did not open that path (the write gate reads the registry, not the item store); it only made the type visible in `getMetaTypes()`, and both the issue and this row say so to stop the next reader filing it as a regression. Landed as ruling A on ADR-0066 D1's own authority: `allowRuntimeCreate: false` **and** `allowOrgOverride: false`, the second self-judged inside the ruling's rationale and flagged for veto — a tenant overlay of a package declaration would lift `scope` from `org` to `platform`, which is the one field on this type that is an escalation rather than display. Its reverse verification is worth copying: deleting the registry entry gave 7 red / 3 green and measured something **sharper than predicted** — a garbage payload turned 422 rather than resolving, i.e. the schema binding is a real second line of defence behind the registry row, not a restatement of it; deleting the schema binding alone gave exactly 3 red. `packageId` is the one key that reads oddly: deliberately a FALLBACK, not the primary, since #5870 added `capabilities` to the ObjectQL stamped-collection list so `_packageId` now reaches a declaration and wins — it stays `live` because the fallback branch still decides materialization for any declaration arriving unstamped. Zero dead |
+| manifest | seeded 2026-08-23 (#10728) — **not a metadata type and not a stack collection either**, which is the row's whole point. `ManifestSchema` (`packages/spec/src/kernel/manifest.zod.ts:132`) is what an author writes as `objectstack.config.ts` or a packaged manifest; it is parsed at `packages/objectql/src/registry.ts:2950` and by `os plugin build`, and it sat outside the ratchet's universe entirely — `GOVERNED` listed no `plugin`/`manifest`/`package`, `SPEC_ONLY_SCHEMAS` covered only webhook/query/validation/qa, `PENDING_GOVERNANCE` was empty so the gate reported itself **complete**, and `liveness/` had no file for it. A ratchet extended only to unregistered KINDS would not have reached it either: the retired-key entry `17.kernel__Manifest__loading.ts` records that `PLURAL_TO_SINGULAR` has no `packages`/`plugins` entry, so a manifest is never walked as a stack collection member. That blind spot was paid for twice, by hand and after the fact — `loading`'s ten inert keys (#4914, one of them `sandboxing`, which isolated nothing while looking like isolation) and the `contributes` census (#10627), which found exactly ONE reader of the 11-member block monorepo-wide. Dead 21 = the ten dead `contributes` members (`kinds` is the sole live one — `engine.ts:4504` → `registerKind`), the five `capabilities.*` and two `configuration.*` keys (all three containers have ZERO reads of the container itself, so no key beneath one can be read), plus `extensions`, `integrity`, `runtime`, and the tombstoned `loading` whose row must stay because `retiredKey()` keeps the key in the walked shape. **Two of those are security-shaped and neither is retired here**: `integrity` declares per-file digests the runtime is documented to re-verify at unpack while nothing computes or checks them, and `runtime` — the ADR-0025 §3.6 trust tier — is read only by two CLI lines that ECHO the value, with no `runtime === 'sandbox'` branch anywhere, even though `loading`'s own tombstone redirects upgrading authors to it as something "which [is] enforced". ⚠️ The `contributes` dispositions are recorded as **PENDING** on purpose: #10724 (tombstone the nine mechanically-dead members) is open and blocked on the cloud census, and #10726 (`contributes.routes`, ruled B) routes its removal through that card — nothing has merged, so no row here says retired. ⚠️ Cloud is unmeasured throughout, carried unchanged from #10627 (that repo could not be attached to the measuring session), so every `dead` above is an objectstack+objectui verdict and an enforce-or-remove worklist row, not a licence to delete |
 
 The `dead` set across types is the enforce-or-remove worklist (ADR-0049); every
 misleading entry carries `authorWarn` so authors hear about it at compile time