fix: one failed LLM call no longer aborts the whole run

[UditAkhourii]

Summary

• run() used Promise.all for both the divergence fan-out (frames.map(...)) and the deepen pass (toDeepen.map(...)). If a single branch's callLLM threw (rate limit, timeout, an unrecoverable malformed response), the whole run rejected and every other already-completed branch's work was discarded.
• Both fan-outs now use Promise.allSettled. Fulfilled results are kept as before; rejected ones are recorded instead of propagating.
• Added RunResult.failedFrames: { frameId: string; error: string }[] (types.ts) so callers/the CLI can see partial failure instead of a silent gap or a hard crash.
• Added a "frame:failed" RunEvent, surfaced by cli.ts's progress printer ( ✗ <frame> failed: <reason>) and by render.ts's text output ("Partial failures" section) when failedFrames is non-empty.
• A run now only hard-fails if every divergence branch fails (all frames threw) — matches the issue's requirement exactly.
• divergeBranch, scoreIdeas, and clusterIdeas already failed open internally on JSON-parse errors; this only changes the network/LLM-call layer that could previously throw before ever reaching parseJSON.

Testing

• npm run typecheck — clean.
• npm test — existing suite passes (1/1). I did not add a new automated test for this fix: exercising the failure path would require mocking the Claude Agent SDK's query() call, and the repo has no existing mocking setup for LLM calls (the one existing engine-adjacent test only covers the pure buildQueryOptions helper) — introducing one felt out of scope for a targeted bug fix. Verified the reconciliation logic by inspection and by running npm run build && node dist/cli.js --help.
• npm run build — clean.

Closes #44

🤖 Generated with Claude Code

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None