[thekevinbot]

Fixes #932

Docs-only. The code half of the post-#834 cleanup — just preflight reading the deleted conventions.yml — landed separately as #983, and this PR deliberately does not touch it.

Item 1 — retire the pre-#834 workflow names

conventions.yml, python-test.yml, ts-test.yml, rust-test.yml are gone; the docs still routed readers to them. Renamed to their surviving homes across AGENTS.md, all three named reference docs (testing-conventions-ci.md, testing-gates.md, e2e-attestation.md), plus internals/distcheck/README.md, the justfile recipe comments and a testing-conventions.toml comment — same stale names, and a contributor reading just --list hits those first.

Added the caller-job table to testing-conventions-ci.md, so "which workflow runs this gate" is answerable without grepping .github/workflows. The eight caller job ids survived the split unchanged, which is why check-run names did too — worth stating, since it's the non-obvious part.

Historical mentions of conventions.yml are kept where they're describing what #861 changed.

Items 2 + 3 — merge gating, rewritten against what actually runs

Staler than the issue assumed. AGENTS.md was already off Mergify (#950), but it taught pre-sleep: 30 / minimum-checks: 0 / timeout: 45 as the gate's tuning knobs — and pr-monitor@v1 has none of those inputs. github-token is the entire surface; the rest were removed upstream when the action stopped guessing. Ours sat inert until 3d248a0 dropped them as such. So the documented failure modes described machinery that was not running.

What v1 actually does: willfire evaluates the workflow files against the PR's base, changed files and head to predict the exact run set, and the gate is a set comparison — yellow while a predicted run is missing or unfinished, red on a bad conclusion, red on a run nothing predicted.

Item 3's premise inverted, and I documented the reversal rather than the request. The issue asked to write down "a path-filtered workflow still in flight does not hold the gate". Under prediction the opposite holds: it was predicted, so it is required, and it holds the gate pending. Zero-check-merge is still the policy, but it is now a derived verdict (the gate reads the same files GitHub does) rather than the accepted blind spot it was under check-counting.

Item 3 also said to cross-reference #862. That's closed not_planned, so pointing at it would send readers to a dead end — documented instead that prediction subsumes most of what it would have checked.

Note on size: the rewrite pushed AGENTS.md past the 40k limit its own line 13 declares. Split the operational half into agents/reference/pr-monitor.md, following the convention the file already uses for four other references. AGENTS.md ends at 36,479 — smaller than the 36,884 it started at.

Item 4 — READMEs

Recorded in AGENTS.md and agents/reference/changelog-migrations.md that a package README needs a changelog fragment: READMEs ship in all three published artifacts, which is why release-ci.yml keeps them while excluding fragment dirs, and why _is_exempt omits them. No code change — the gate is correct as written. #834's spec row amended in a comment.

Item 5 — #885

Amended in a comment: the publisher is docs-release.yml, not docs-deploy.yml. The underlying fix already landed in #967, and I verified the asymmetry is fully closed — neither docs workflow lists MIGRATIONS.md, and docs-ci.yml now carries its own self-path.

Why comments and not body edits for items 4 and 5: the GitHub MCP issue_read returns bodies HTML-entity-escaped and lossy — it silently drops HTML-ish text, including the VitePress include directive in #885. Writing a fetched body back would have destroyed content in a closed issue.

Rebased onto #983

#983 merged while this was open and touched two of the same files. Resolution:

• internals/checks/README.md and the preflight recipe comment — fix(checks): derive preflight's gate matrix from every workflow that calls testing-conventions #983 made the same correction with better knowledge (it documents --conventions, and the new internals-checks-integration job). Took its version wholesale; this PR no longer touches that README.
• The justfile's other four stale names (python-test.yml / ts-test.yml in the distcheck recipes, conventions.yml in e2e-attest-python and test-packaging) did not conflict and are still needed, so they are restored in 5bc9a39.
• That commit also names --conventions in AGENTS.md's preflight line, since the line enumerates the other two options.

Verified the resulting claim rather than asserting it — dirsql-checks preflight --dry-run on the rebased tree derives exactly 40 gate pairs across 8 scan roots in 6 workflows, which is what both AGENTS.md and the justfile now say (48 total lines, less the 4 uv-sync + 4 declared-deps backstop pairs documented separately).

Follow-up filed, deliberately not fixed here

#982 — the docs still name a single-file e2e-attestation.json in ~8 places, but attestations are per-branch receipts under e2e-attestations/<branch>.json. AGENTS.md contradicts itself on this. It's a different drift (the receipts change, not the #834 split), so folding it in would have widened this PR.

E2E Verification

N/A - docs only

No source under packages/** or plugins/** changed. The internals/distcheck/README.md edit sits outside any --scope, so no attestation stales.

Changelog / Migrations

• Changelog fragment added for each changed package — not required: no packages/<pkg>/ or plugins/<pkg>/ source touched; every commit carries a skip-changelog: reason
• Migration fragment added — not required: documentation only, no behavior change

No PARITY.md change — no SDK public API touched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KcNqe44kvgP5mBRr9LjQcE