docs(qa): ai checklist — apply the 18.2.0 sweep's 14 ledger findings (#25836) (#26099)

<!-- claude-projects-attribution: {"github_login":"os-elon-musk"} -->
_Requested by **Elon** · [project
thread](https://claude.ai/code/project/chan_01WedDzV26nniwxv7jC88B9D?thread=cmsg_01WedDzV26nniwxv7jC88B9DF1Sr5Z9ggF6iF1xM67fuqu)_
Fixes #25836
Clause-②: no
Metadata-protocol: none

Applies the 14 ai checklist-accuracy findings of the 18.2.0 release
sweep (wave anchor #25807) to
`docs/qa/platform-checklist/areas/ai.json`. Each revised item has its
`revision` bumped and a `history` row citing its run record: P0 #25812
or Tier 1 #25815, both measured at 6679644. Every line was re-read
against `main` at b4199c9 first. No clause is added or dropped, and no
`status` changes.

## Per line

| Item | Finding (record) | Applied as |
|---|---|---|
| `ai.mcp-stdio-fail-closed` | steps 2/9 need `--log-level info`;
`query_records` caps `limit` at 50 (#25812) | steps[2], [3], [9] and
acceptance[7].verify. `os start` still defaults to `warn`;
`DEFAULT_MAX_LIMIT` is still 50 |
| same | a0: pin checks the variable name only (#25815) |
`automated.ref` prose: pin layer and reach |
| same | a1/a6/n4: no API_KEY_REVOKED pin (#25815) | `automated.ref`
prose. Still true on main: no `packages/mcp` test names it |
| `ai.mcp-coded-refusals-and-runtime-exposure` | a1:
`mcp-coded-refusal.test.ts` not in the ref (#25815) | added to the ref,
plugin-unit layer |
| same | a2: no pin drives `query_records` across a restart (#25815) |
**No longer holds on main**:
`packages/qa/dogfood/test/mcp-stdio-runtime-object-exposure.test.ts`
does, on the stdio door at engine layer. Added to the ref with that
layer; the HTTP door across a restart stays a live leg |
| `ai.mcp-http-surface` | a3: 401 status only; a4: session bearer, no
initialize, 2 names; ref bare (#25815) | ref names every pin with its
layer and what it leaves unasserted. **Part no longer holds**: a0–a2 and
a5 do have pins on main (dispatcher units and Hono integrations in
`packages/runtime`, scope narrowing in `packages/mcp`). None scores
those clauses pass alone |
| `ai.mcp-oauth-private-host-transport` | a0–a4: unit-only pin scores no
clause; a4 predicate half only (#25815) | ref prose.
`mcp-oauth-plaintext-notice.test.ts` is added as plugin-unit
corroboration of a2 |
| `ai.mcp-validate-expression` | a2: no same-expression site flip;
a0/a1/a3: fake bridge (#25815) | acceptance[2].verify and ref prose |
| `ai.skill-instructions-mcp-prompts` | a3: unknown name only; n1: no
prompts/get-inactive pin (#25815) | acceptance[3].verify and ref prose |

Plus, per the dispatch order, one retired-pin edit:
`ai.console-ai-surface-gating` acceptance[5].verify told runners to run
two objectui tests "at the .objectui-sha revision". It now points at
their home in this tree, `ui/packages/plugin-chatbot/src/__tests__/`
(#26035, #25745).

No line described a product defect, so no card was filed.

## Verification

- `pnpm check:platform-checklist`: OK (16 areas, 329 items).
- `node scripts/pm/dispatch-gates.mjs --commands` for this file list:
all 13 commands exit 0. `check:doc-formula-expressions` exited 3
(prerequisite) until `@objectstack/lint` and its deps were built, then
0.
- Every file newly named in an `automated.ref` passes at this head:
- mcp: `mcp-coded-refusal`, `mcp-http-tools.scopes`: 2 files, 52 tests.
- runtime: `http-dispatcher.mcp`, `http-dispatcher.mcp-oauth`,
`mcp-anonymous-first.hono.integration`,
`mcp-skill-method-not-allowed.hono.integration`: 4 files, 52 tests.
  - plugin-auth: `mcp-oauth-plaintext-notice`: 1 file, 26 tests.
  - dogfood: `mcp-stdio-runtime-object-exposure`: 1 file, 10 tests.
- Governed surface: none. `docs/qa/**` is not on the governed register.
Not Tier H, no contract review.

`skip-changeset`: docs only.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01GJ41Um5Ueq9ESR3Nj5D7qW

Co-authored-by: Claude <noreply@anthropic.com>

Author: objectstack-fleet[bot]