ci: run Claude Opus 5.5 with a supported CLI

[cquil11]

Restore Claude Opus 5.5 in the six Claude Code calls changed by #2805. Pin the action to v1.0.234 and install Claude Code 2.1.282 explicitly before each call, passing its executable path to the action.

The earlier sign-off failure used Claude Code 2.1.265. Anthropic's reported compatibility error for Opus 5.5 requires 2.1.280 or newer; the workflow previously installed the older stable channel. The explicit version check makes a CLI mismatch fail before invoking the model.

Validation:

• Sign-off verifier run 36160952462 passed against PR perf(amd): enable MiniMax-M3 ATOM FlyDSL decode / 启用 MiniMax-M3 ATOM FlyDSL 解码 #3388 using Claude Code 2.1.282 and claude-opus-5-5 (is_error: false).
• Zizmor passed on the final commit.
• actionlint passed for all five workflows after filtering the repository's existing $/.github/workflows/... reusable-workflow references.
• git diff --check passed.

[claude]

Findings marked 🟡 are optional suggestions and need no follow-up push.

Additional findings (outside the current diff — GitHub can't attach inline comments there):

• 🟡 docs/klaud.md — Maintainers reading docs/klaud.md (and docs/klaud_zh.md) after this merge see stale model and action-pin info that no longer matches the workflows. Line 11 still says the PR review uses claude-opus-5 and candidate execution uses claude-fable-5-1, and the table at line 121 still lists anthropics/claude-code-action as v1.0.218 / 0d0e0876d3ea, but this diff changes klaud-plan.yml and klaud-candidate.yml to claude-opus-5-5 and bumps the action pin to v1.0.234 / 9171db3e57d6 everywhere. Fix: update docs/klaud.md and its Chinese counterpart docs/klaud_zh.md (line 11 prose and the Action/Version/Commit table) to match the new model name and action pin, per AGENTS.md's rule to update the nearest English guide plus its Chinese counterpart when implementation changes.

  Why this was flagged

  AGENTS.md:8 states: repository workflows are authoritative and documentation disagreeing with implementation must be updated (nearest English guide plus its Chinese counterpart). This diff changes klaud-plan.yml's and klaud-candidate.yml's --model flag to 'claude-opus-5-5' and the anthropics/claude-code-action pin to 9171db3e57d6a3140a37ddc2ba92788584e0ead6 (v1.0.234) in all five workflow files. docs/klaud.md:11 and docs/klaud_zh.md:11 still say the PR review uses claude-opus-5 and candidate execution uses claude-fable-5-1; docs/klaud.md:121 and docs/klaud_zh.md:121 still list the action pin as v1.0.218 / 0d0e0876d3ea. A maintainer or contributor consulting these guides after merge gets incorrect model/version info, violating the AGENTS.md doc-sync rule; no safeguard catches doc drift automatically.

  Verification: normal (documentation regression the base does not have). CLAUDE.md points to AGENTS.md, whose rule at AGENTS.md:8 states: "Repository source, schemas, workflows, launchers, and collectors are authoritative. If documentation disagrees with implementation, follow the implementation and update the nearest English guide plus its Chinese counterpart." The diff changes the implementation but not…

[claude]

🔴 The new Install Claude Code step has no continue-on-error, so an npm hiccup or a CLI version mismatch now hard-fails the job even though the Claude action step right after it still has continue-on-error: true. In run-sweep.yml this is inside the 'setup' job (run-sweep.yml:269-276); every eval job is gated on needs.setup.result == 'success' (run-sweep.yml:860), so a flaky install now skips the whole sweep instead of degrading to the '["patchwork"]' fallback at run-sweep.yml:329. Same gap in klaud-plan.yml:69-76 (continue-on-error at klaud-plan.yml:80) and klaud-candidate.yml:65-71 (continue-on-error at klaud-candidate.yml:74). Fix: add continue-on-error: true to each new Install step so it fails no harder than the step it precedes. [also at: .github/workflows/klaud-plan.yml:75 - Maintainers now get a hard job failure (and skipped downstream steps) from a transient npm/CLI issue, where the base only failed the Claude call itself.]

Why this was flagged

Trigger: an npm registry blip or claude-code version drift makes the test "$("$claude_cli" --version)" = "2.1.282 (Claude Code)" assertion fail in the 'Install Claude Code 2.1.282' step (run-sweep.yml:269-276, klaud-plan.yml:69-76, klaud-candidate.yml:65-71). That step lacks continue-on-error, so GitHub Actions marks the whole job failed regardless of later steps' continue-on-error. In run-sweep.yml this step lives in 'setup'; 'Classify priority criteria' (run-sweep.yml:277-282) has continue-on-error: true and 'Normalize priority classification' (run-sweep.yml:319-331) already falls back to '["patchwork"]', proving the workflow was built to tolerate Claude-call failures. But the Install step now fails 'setup' outright, so every downstream eval job gated on needs.setup.result == 'success' (run-sweep.yml:860) is skipped. Before this diff there was no Install step, so only the continue-on-error-protected action could fail, and 'setup'…

Verification: normal (conflicts with stated purpose, partially): merging introduces a new hard-fail point that the base branch does not have, in code whose whole design is best-effort/non-blocking. Facts, run-sweep.yml: - The new step (269-276) Install Claude Code 2.1.282 has no continue-on-error; line 275 test "$("$claude_cli" --version)" = "2.1.282 (Claude Code)" (and the npm install at 273)…