oracle: the nine llm cells the mock answers are recorded as <old id>@fixture; golden/1.5.5 retires their bare-id SKIP rows (conformance-gaps D (a)); pin busbar-release #372

[MattJackson]

DRAFT. Owner rule 22 (named conformance gaps = 0); coordinator ruling 2026-10-14, conformance-gaps D, option (a). Stacked on #926 (rule 19), and it also waits on #861 and #893: the generator at the new pin emits their 7 cells too, so cells --check needs them.

Why the ids change. busbar's golden/1.5.5/ledger.tsv held each of the nine ids as a needs_fixture SKIP row, which is a recorded fact:

• golden::add never fills such a row. oracle-rerecord run 38521235302 added 0 cells.
• golden::guard refuses a changed input for such an id.

busbar-release #372 therefore emits the recordable cells as NEW ids <old id>@fixture. This follows the @hooks-webrequest and @1.5-line precedents.

This branch

• cells.json: the nine definitions as oidf rig: FAPI 2.0 subject, provisioned PS256 clients, IdP stub, plan overrides (depends on #153) #369 wrote them, under the @fixture ids (2566 cells).
  • stream_upstream_error x6: fresh.
  • ok_citation and ok_cachepoint_document: no needs_fixture.
  • ok_tool_use_tokens: mock_control tool-use-tokens.
• golden/1.5.5: the nine bare-id SKIP rows leave ledger.tsv, each through a cited, add-only retired.tsv row that names its replacement. They carried no bytes and no PASS, so meta recorded stays 1066. No other golden file changes.
• owed-baseline.txt owes the nine @fixture ids.
• accepted-gaps.json drops stream-upstream-error, gemini-tool-use-token-term, bedrock-cachepoint-before-document and responses-url-citation.
• testing/llm-conformance/named-gaps.json strikes its three accepted entries, and expected goes from 16 to 0.
• accepted-differences.json S-15 and S-16: the cells regexes rename the ids only. Counts, classes, premises and signoffs are unchanged.

Pin and goldens (done).

• oracle-rust.pin moves to busbar-release main 4c47cad0385d (token-sealed A: the kernel member crates' minting tests move into the kernel (GATE-GREEN) #373 and construction: token-sealed and its three mints struck from the standing reds (GATE-GREEN) #374, which carry oidf rig: FAPI 2.0 subject, provisioned PS256 clients, IdP stub, plan overrides (depends on #153) #369 and token-sealed B: the kernel's test-only mint for root, llm and core-admin tests (GATE-GREEN) #372).
  • The x86_64 and aarch64 digests are the dev-bin .sha256 assets.
  • Every .sig (including macos-aarch64) verifies Ed25519 over <asset>\n<sha256>\n against the release key.
• cells --check with that engine finds the nine @fixture cells byte-equal to the generator. The only drift is the 7 cells that row 155 oracle: billing|budget|downgrade cell, its 1.5.5 golden, and the oracle pin at busbar-release db64d9218d8b #861 and row 155 oracle: llm stream_upstream_cut cells and their 1.5.5 goldens (stacked on L1#2 #891 and on #861) #893 add.
• golden-guard --base origin/predev with that engine reports that add-only holds.
• The goldens come from oracle-rerecord run https://github.com/GetBusbar/busbar-release/actions/runs/38549469444, recorded from the published 1.5.5:
  • golden-add added 9, changed 0 and removed 0.
  • ledger.tsv is +9, meta recorded goes from 1066 to 1075, and retired.tsv is untouched by the run.
  • The files are copied from the artifact unchanged.

Money. The one ledger x rate card divergence is ok_tool_use_tokens@fixture: 1.5.5 billed 250 cents, and the right figure is 290. It is registered in the separate $ draft #1023, stacked on this PR. The other eight equal ledger x rate card.