docs: add runner port-forward staleness keepalive (stop-gap) guide

[weicao]

Summary

New methodology doc addon-runner-portforward-staleness-keepalive-guide.md covering workstation-side kubectl port-forward silent TCP staleness (process alive, stream dead) and a lightweight /version probe + auto-restart stop-gap monitor. The doc explicitly frames the keepalive as transitional; preferred structural fix is addon-runner-incluster-vcluster-access-pattern-guide.md.

Body (generic methodology, version-agnostic, no engine binding):

• §1: failure shape (PID alive but TCP stream dead, kubectl returns EOF / Client.Timeout); runner cannot distinguish from remote outage
• §2: 6 hard rules — monitor only probes + restarts (no proxy, no business-request retry); INTERVAL and FAIL_THRESHOLD separately tunable, both non-extreme; precise pkill -f pattern that won't kill other port-forwards on workstation; monitor liveness logged; framed as runner-harness only (NOT product / addon / KB / vcluster fix evidence); when in-cluster runner pattern is feasible, prefer it and retire keepalive
• §3: 5-point PR review checklist
• §4: 3 anti-pattern vs correct-pattern pairs
• §5: relation to addon-runner-incluster-vcluster-access-pattern-guide.md, addon-runner-openapi-schema-fetch-brittleness-guide.md, addon-test-runner-cadence-discipline-guide.md, addon-evidence-discipline-guide.md

Appendix A is OceanBase enterprise addon case: (A.1) N=3 attempt RUN_ID pitr-runtime-runner-hardening-N3-... hit T1 1800s budget overflow due to workstation pf-staleness; remote was healthy throughout. (A.2) keepalive landing in RUN_ID pitr-runtime-pf-keepalive-N3-...: monitor PID 87376 alive 47m21s, 3 auto-restarts (05:33:26Z / 05:34:13Z / 05:38:51Z), T1 PASSed where prior attempt failed. (A.3) subsequent migration to in-cluster pattern retired the monitor. Explicit boundary: keepalive 1-sample landing not extrapolated to permanent immunity.

SKILL-INDEX.md updated: added entry under ### 5. 改造 runner / 工具链.

Test plan

• Manual: body explicitly frames stop-gap, prefers structural fix
• Manual: appendix has explicit "1-sample, not extrapolated" boundary
• Manual: cross-doc references resolve

🤖 Generated with Claude Code

[weicao]

Blocking for merge:

1. PR body still contains 🤖 Generated with [Claude Code].... Public PR body must not include AI/tool attribution.
2. The new guide intro has only 4 standard fields. Please add > **Affected by version skew**: ... after Applies to KB version.
3. This PR links to docs that are not on main yet: addon-runner-incluster-vcluster-access-pattern-guide.md (docs: add runner in-cluster vcluster access pattern guide #124) and addon-runner-openapi-schema-fetch-brittleness-guide.md (docs: add runner OpenAPI schema fetch brittleness + --validate=false guide #125). Please merge/rebase in dependency order, or remove/defer those links until targets exist.

The stop-gap framing is important and should stay: port-forward keepalive is transitional; new/long-running test paths should prefer host-runner + in-cluster vcluster API access.