docs(skill): add release-gate reference

[robotlearning123]

What

Adds skill/agent-ready/references/release-gate.md: a methodology reference for a release-verification layer above unit / integration / BDT / E2E tests, linked from SKILL.md under a new "Beyond the 9 Areas" section (the 9-area table and checker are unchanged).

Covers: diff-based risk tiers (T0–T3), a small bounded mission pack with personas and oracles, evidence capture, a deterministic lane that stays blocking, the ship / investigate / block verdict with an explicit per-mission roll-up, the one-command release:ready pattern, and common mistakes (false-green scripts, advisory treated as approval).

Tests

• New test/skill-references.test.ts: doc exists, is linked from SKILL.md, and keeps the tiers, verdicts, non-blocking investigate and the critical-failure → block roll-up. The content test fails against the first draft of the doc (which had a contradictory verdict rule) and passes now.
• npm run check exit 0; npm test 60/60 pass.

Review

Independent review found a contradiction in the first draft (advisory never fails the build vs. block fails the check); fixed with an explicit roll-up table and precedence rule, then re-reviewed: PASS.

🤖 Generated with Claude Code

[robotlearning123]

Loop review: verified by execution (review-only)

Scope: docs-only skill reference PR. Verified at head 7fefc45 (worktree git worktree add ... origin/docs/release-gate-reference-20260924 --detach), base e704b49 == origin/main. This account is not the PR author and did not modify the branch.

1. CI statusCheckRollup at head — 8/9 SUCCESS

claude-review        FAILURE   (Claude Code Review)
Check Repository     SUCCESS
Lint & Format        SUCCESS
Validate PR          SUCCESS
Type Check           SUCCESS
Scan Agent Readiness SUCCESS
Test (Node 20.x)     SUCCESS
Test (Node 22.x)     SUCCESS
Build                SUCCESS

The only red check is claude-review, and its failure is the review bot's own harness error, not a finding about this diff:

##[error]Claude result reported subtype success with is_error:true (run did not complete successfully)
##[error]Action failed with error: Claude execution failed: result is_error:true

Pre-existing and repo-wide: gh run list --workflow="Claude Code Review" shows 10 failures in the last 12 runs across unrelated branches (chore/merge-satellites, fix/agent-ready-pr71-baseline-20260922, chore/agent-native-baseline-20260924, feat/repo-infra-dogfood, and the four 2026-09-28 loop/* PRs #89-#93 — all after this PR was opened). The first green run is #78 (2026-09-25), which re-routed the bot to GLM-5.3; this PR predates that fix. No push-event runs exist for main (workflow is pull_request-only), so cross-PR reproduction is the base-ref evidence: the PR does not make the baseline worse.

2. Docs oracle (broken-reference check) — PASS

$ grep -n "release-gate" skill/agent-ready/SKILL.md
40:| Release Gate | `references/release-gate.md` | ... |
$ ls -l skill/agent-ready/references/release-gate.md
-rw-rw-r-- 1 robot robot 9062 ... release-gate.md

Target exists at the referenced relative path, same form as the other rows in the references table.

3. Narrow test suite (test/skill-references.test.ts, added by this PR) — 3/3 PASS

$ npx tsx --test test/skill-references.test.ts
# tests 3
# pass 3
# fail 0

Tests use real oracles (existsSync, exact sentence/table-row string matches), not assert-true.

4. Supersession / overlap check

None. Sibling #76 (agent-native baseline) touches only Makefile — zero file overlap with this PR. No other open PR adds a release-gate doc.

5. Independent review (grok-4.7, full diff pasted) — FAIL with 2 doc-consistency findings

1. (major) The roll-up table (release-gate.md lines 122-124) routes critical mission fail to block but non-critical fail / required mission not run to investigate, yet nothing operationally marks individual missions as critical/non-critical/required — the pack section only says "cover critical paths only". An evidenced pack-mission failure can be read as either block or investigate.
2. (minor) "First match wins" leaves edge results unmapped: a pass without evidence matches no row (row 3 requires evidence); a critical fail without evidence is covered only by the trailing prose sentence, not by a table row.

The reviewer confirmed the SKILL.md link is correct and the tests are real oracles.

Verdict

Checks/tests: green at head; the single failure is a pre-existing repo-wide bot-harness failure (see above), so this PR is not worse than base. The two review findings above are docs-consistency gaps for the PR author (or a follow-up PR) to resolve — they are advisory, not execution failures, and this review lane cannot rewrite a foreign branch.

[robotlearning123]

Independent review (grok lane, execution-based; writer was a different model) at head 7fefc45.

Verdict: FIX-FIRST (1 major, 1 minor; no scope creep, no broken references)

Findings

1. major — skill/agent-ready/references/release-gate.md:120-126 (roll-up table gap): a required mission with result pass but a missing evidence pointer matches no roll-up row: the investigate row (line 123) covers only blocked / non-critical fail / required-not-run, and the ship row (line 124) requires pass with evidence; the precedence prose (line 126) remaps only an evidence-less fail. A literal top-to-bottom matcher on {critical: true, result: pass, evidence: false} returns UNMATCHED, yet release:ready must emit one of ship/investigate/block — an undefined case risks falling through to exit 0, i.e. the false-green anti-pattern the same doc warns against. Fix: add an explicit row/rule (e.g. evidence-less pass -> investigate) so the three-way verdict is total.
2. minor — test/skill-references.test.ts:31: the content test locks the block row and the investigate-non-blocking phrase but does not guard the evidence-less-pass case, so finding 1 could be (re)introduced without a test failure.

Checked, does not hold

• Diff is exactly the 3 stated files; 9-area table and checker untouched; no scope creep.
• SKILL.md:40 links references/release-gate.md; target exists in the same PR. Not on main — not superseded; no overlap with chore: agent-native baseline (AGENTS.md, make check, CI, dev env) #76 (Makefile only).
• Makefile fence uses a real tab (make release-ready reproduced, exit 0); prettier clean; lane semantics internally consistent (deterministic lane always blocking; investigate never fails build; no auto-merge).

Machine gates

• tsx --test test/skill-references.test.ts at this head: 3 passed / 0 failed.
• CI at head: Test (Node 20/22), Lint & Format, Type Check, Build, Scan Agent Readiness all SUCCESS. claude-review FAILURE is a pre-existing bot-infra failure (fails on all recent branches; being re-routed in ci: route PR review bot to GLM-5.3 via z.ai #78), not caused by this diff.

Review-only comment — merge decision stays with the owner.