docs: add chaos auto-rejoin race detection guide + MySQL MGR case

[weicao]

Summary

• Add docs/addon-chaos-rejoin-race-detection-guide.md — engine-neutral soak design that catches HA engine auto-rejoin race (5-element design + 7-signal detection + layer classification + patch validation plan).
• Add docs/cases/mysql/mysql-mgr-auto-rejoin-race-case.md — engine-specific MySQL MGR case backing the guide: two independent observations (William chaos.sh 2026-05-15T04:49Z + Henry 4h soak 2026-05-15T18:43Z), both matching mgr-server/internal/syncer/JoinCurrentMemberToCluster once-and-return pattern. Trigger conditions + patch hypothesis documented.

Why now

• 22 valid chaos C2 attempts (batches 1-6, fresh + C1→C2 sequence, varied conditions, single-row writers) recovered ≤30s but never hit the race — strong negative evidence that the race requires sustained-load + long-duration to expose.
• 4h soak with 5 parallel writers, alternating chaos every 5 min, hit the race at cycle 36-37 (~3h in), with engine log pattern identical to William's earlier independent observation.
• Document the soak design so any addon team can detect this class of race without having to rediscover it.

Test plan

• Both observations have evidence packages with full pod-side engine logs, KB controller logs, kill-log, 3-pod independent MGR view snapshots. shas listed in the case appendix.
• Engine-neutral guide checks the index pattern (Hard Rules / When To Invoke / Mechanism / Soak Design / Detection / Classification / Escalation / Validation Plan / Related Docs).
• Case appendix keeps engine specifics out of the main guide.
• Tracked as task docs(probe-script-zombie): cross-link mariadb negative case #27 in #mysql; will update case material when William's patch (JoinCurrentMemberToCluster bounded retry + reconnect) clears patch-version 4h soak validation.

🤖 Generated with Claude Code

[weicao]

Review note: useful guide and case split. One scope correction before I would call it fully clean: docs/addon-chaos-rejoin-race-detection-guide.md:128 says the cluster surface phase eventually returns Running, but both MySQL evidence variants we are citing had the cluster stay Updating beyond the normal recovery window. Please reword this detection indicator to cover both outcomes, e.g. "the surface phase may return Running and mask the issue, or may remain Updating; either way pod-local view / role-label evidence is authoritative." That keeps the generic method aligned with the case evidence.

[weicao]

Re-reviewed the phase wording fix. Detection indicator now covers both Running-mask and Updating-stuck variants and keeps the evidence anchor on pod-local view + role label, which matches the MySQL case evidence. GitHub will not let me approve a PR under the same account, but from my side this PR is OK.

[weicao]

Final docs-gate check: useful topic, but not mergeable yet.

Blockers I see:

1. Public GitHub hygiene: PR body still has the generated-by footer, and both commits have Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>. Please rewrite/squash the branch so origin/main..HEAD commit bodies are clean, and remove the PR body footer.
2. Missing index entries: both new files need docs/SKILL-INDEX.md entries (docs/addon-chaos-rejoin-race-detection-guide.md and docs/cases/mysql/mysql-mgr-auto-rejoin-race-case.md).
3. Missing standard intro metadata: the guide and the MySQL case need the standard intro block fields (Audience, Status, Applies to, Applies to KB version, Affected by version skew; case files should also state scope / generalization boundary clearly).
4. Public-doc residue in the case: remove local absolute evidence paths (/Users/wei/...) and Slock msg-id references (msg 61756a30, msg f5289913). Keep public artifact names, sha, timestamps, and technical evidence summaries instead.
5. Evidence boundary: the guide currently recommends N=3 patch validation, while the case says patch validation is still pending. That is fine, but please make the case status explicit as "candidate mechanism / patch validation pending", not settled or release-ready.

No technical objection to the direction; this is a hygiene / public-doc boundary pass before merge.