chore(deferrals): re-decide both expired rows, and bump the ratchet they were holding

[tomtom215]

What and why

cargo xtask deferrals --check is a weekly scheduled gate, not a PR gate (ADR-0010, deliberately — an expiry should page the schedule, not block unrelated work). Nobody acted on it. Two rows passed their review-by dates, and three published claims had rotted behind them. Both are re-decided against re-fetched evidence rather than re-dated blind, which is what ADR-0010 requires.

rust-sdk-902-offer-clock (review-by 2026-08-10). The offer is open with zero replies at 68 days — no comment, no assignee, no linked PR, labelled P2/enhancement. But R9 has not fired: its trigger is two substantive offers unanswered for 60+ days, and the same day's other filing was answered — rust-sdk#903 closed and fixed by merged #905 in nine days. So the count stands at one and M4's DoD does not re-scope. Outcome recorded in the risk register and the engagement doc; not posted upstream, because this repo's own pattern (issue #9) makes posting a maintainer action.

That PR was also not "maintainer-authored" as the engagement doc claimed — GitHub shows the author carrying the Contributor badge, approved and merged by a Member. Still a successful engagement, but the distinction is precisely what R9 measures: maintainers triaged and merged a community fix promptly; they did not write it.

suite-0-2-0-stable-pin-bump (review-by 2026-08-15). Half of it was buildable, so it is built. The row and register row 2.4 both said the npm alpha dist-tag "has been quiet since 2026-07-01" — alpha.10 published 2026-07-27, the day after that was written, and alpha.11 on 2026-08-07. That is ADR-0010's own founding example recurring verbatim. The stable pin genuinely cannot move (latest is still 0.1.16, unchanged since 2026-03-30) and re-dates; the draft ratchet moves here with a BLESS=1 re-measurement in the same commit.

The bump paid for itself immediately

No pre-existing check changed status. The entire delta is 36 new wire-schema-valid checks, which validate every message against the negotiated revision's JSON schema. Thirty-two pass. The four that fail are all on the 2025-11-25 leg — resources-{list,read-text,read-binary,templates-read} — each for must have required property 'cacheScope' and 'ttlMs'.

That is CACH-001: the single clause this workspace's registry had already flagged against the legacy server, while the official runner scored both servers an indistinguishable 23/23. The runner found it independently, six weeks later. The standing finding "the runner cannot distinguish the two servers" is recorded as superseded rather than deleted — a negative result about an instrument expires when the instrument improves. Legs now read 37 passing / 4 failing (legacy) and 41 / 0 (stateless); baseline totals move 46 → 78 passing, 0 → 4 failing, 46 → 82 checks.

The asymmetry is recorded because it is instructive: tools/list and prompts/list pass only because rmcp's #[tool_handler]/#[prompt_handler] expansions attach caching hints unconditionally, while resources/* go through this workspace's revision-aware cached(), which correctly withholds them at 2025-11-25. The honest implementation is the one the new check fails.

The gate caught me, and found older residue

The first push failed CI: draft-coverage — CHANGELOG.md quotes a verdict of 123 pass, 1 fail that no committed report produced. The entry had quoted register row 1.5i in good faith. Those figures are the pre-ADR-0012 counts — 58 + 65 = 123 and 59 + 65 = 124 exactly, folding not-observed into pass, the very accounting ADR-0012 removed. The 2026-08-17 sweep that corrected this arithmetic fixed the CLAIM_FILES the gate parses and stopped exactly at that boundary; register row 1.5i, 03-conformance-strategy.md and 06-roadmap.md are outside it and still carried the inflated pair. All three corrected against the committed goldens, second commit. The lesson kept is the boundary, not the arithmetic.

Deliberately not done

Moving deferrals --check into xtask ci. ADR-0010 excludes it on purpose and that reasoning holds. The real gap is the last mile: a red weekly job reaches a human only by an easily-missed email, and the job runs spec-drift too, so a network flake looks the same from outside as an expired row. Fixing that needs issues: write on a workflow that is deliberately contents: read, plus a new pinned action — a security-surface change with no actionlint available here to validate it. Opened as ledger row expired-deferral-notification rather than shipped untested as a rider. Second new row, draft-suite-pin-currency, covers the blind spot that let the ratchet go six weeks stale: the weekly alpha job runs at the registry's revision, so it cannot see draft scenario churn.

Checklist

• cargo xtask ci passes locally — reports every local gate ran and passed: 12 steps, 85 suites, 1,757 tests, 0 failed, 0 SKIPPED, no warnings or errors outside the 3 known-benign license-not-encountered notes. MSRV clippy (1.88.0) and cargo-deny both ran (advisories ok, bans ok, licenses ok, sources ok).
• SPDX header on every new file — no new files; all changes are edits to existing tracked files.
• No source or registry file exceeds 500 lines — every source and registry file (182) is within 500 lines.
• New public items have rustdoc — no public API changed; the only code change is one const string, DRAFT_SUITE_VERSION.
• New code has tests — no new code paths. The behaviour change is the pin, and it is covered by the ratchet itself: cargo xtask draft-readiness re-run after the bless reports matches the committed baseline with a clean working tree, so the recorded baseline is self-consistent at the new pin.
• Golden reports regenerated deliberately — none needed, and none changed. The committed captures under corpus/draft/captured/ are not auto-regenerated (re-copying a tap is a deliberate act, per corpus/README.md), so no .jsonl and no golden moved. conformance/draft-readiness.json was re-blessed and its diff reviewed check by check.
• ADR added/updated — none needed. ADR-0010's mechanism is being used, not revised; ADR-0012's rule is being applied to docs that had escaped it. The one genuine policy change — decoupling the two suite pins' triggers — is a refinement of the version policy in 03-conformance-strategy.md, stated there with its reasoning.
• Plan documents updated — deferrals.json (both rows re-decided, two rows added), 01-ecosystem-context.md (rows 1.5i and 2.4), 03-conformance-strategy.md (pin triggers decoupled; runner-vs-registry paragraph), 06-roadmap.md, 07-ecosystem-engagement.md (rows 1 and 9), 08-risk-register.md (R9), corpus/README.md, crates/mcp-everything-server/README.md, and a dated note in scheduled.yml.

Other gates run

Gate	Result
cargo xtask deferrals --check	7 rows, none expired — the gate this PR exists to satisfy
cargo xtask draft-readiness	matches the committed baseline at the new 0.2.0-alpha.11 pin; working tree clean
cargo xtask conformance	green both legs — zero unexplained divergence, coverage manifest in sync
cargo xtask spec-drift (all features)	412 quotes across 2 revisions, 0 drifted
cargo xtask draft-coverage --check	5 captures evidence 109 of 124 judgeable clauses; every prose claim agrees
cargo xtask mutants	no mutants in scope — a true empty, not a pass. The diff is xtask/** (excluded by .cargo/mutants.toml), Markdown and JSON; no shipped-crate src/ changed. Confirmed with cargo mutants --in-diff --list.

Every figure above was re-measured against the committed data before posting.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SjVp39nWRhAQeYjSGbQWoZ

---

Generated by Claude Code