Let Claude gather CI review context and notify Feishu best-effort

[RyanLee-Dev]

Summary

Simplify the CI review and Feishu notification workflow and hand more of the work to the model.

• Remove the github-script evidence step. Claude now finds the merged PR, reads the diff, inspects job results (including failed logs) and reads the repository rules itself with gh, git and Read/Grep/Glob. Non-PR pushes to main are skipped by the model.
• Check out the triggering commit (workflow_run.head_sha, depth 2) instead of the latest main.
• Rewrite the prompt in Chinese and shorten it; layout and retry decisions are left to the model.
• Raise --max-turns from 12 to 100; step timeout 25 min, job timeout 30 min.
• Make the notification best effort with continue-on-error: true: a failed review or delivery no longer fails the job.

gh needs no extra secret: claude-code-action exports its github_token input as GH_TOKEN to Claude (src/entrypoints/run.ts, base-action/src/parse-sdk-options.ts at the pinned commit). allowed_non_write_users: '*' stays because the merging actor may lack write permission.

Motivation: run 36815477330 failed with error_max_turns under the 12-turn limit.

Validation

• actionlint .github/workflows/ci-review.yml passes.
• Not yet exercised on a real run; workflow_run only takes effect once this reaches main. gh availability on Blacksmith runners is unverified.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}