refactor: unify the LLM retry classifier across client + judge retry

[drewstone]

Summary

Merge B of the agent-eval consolidation. callLlm and withJudgeRetry each carried a private retry-classification surface — two backoff functions, two retryable-status sets, two divergent transient-error pattern lists. One concern, three copies.

Collapse to one exported primitive in llm-client.ts:

• isTransientLlmError(err) — THE retry classifier. Inspects an error's name/message/code, honors a numeric HTTP status (LlmCallError or a foreign SDK error that duck-types .status), and recurses into error.cause — undici nests the real socket fault under .cause.
• backoffMs(attempt) — exported; withJudgeRetry's default backoff.

judge-retry.ts drops its ABORT_PATTERNS, RETRYABLE_HTTP_STATUS, DEFAULT_BACKOFF, and defaultIsRetryable in favor of the shared pair. callLlm's catch path routes through the same function.

Bug class fixed

Neither old classifier matched undici HTTP/2 transport faults — terminated, NGHTTP2_INTERNAL_ERROR, UND_ERR_*, other side closed. llm-client's regex was only fetch failed|ECONNRESET|ETIMEDOUT|EAI_AGAIN; judge-retry's list was broader but still HTTP/1-shaped. An HTTP/2 keep-alive connection dropping mid-response escaped the retry loop — surfacing as an uncaught rejection in the HTTP client and a silently non-retried trial failure in TCloud-backed judges. The unified pattern list covers them and the .cause chain is followed to the real fault.

Test plan

• callLlm retries an HTTP/2 terminated / NGHTTP2_INTERNAL_ERROR fault to recovery
• withJudgeRetry does the same via the shared classifier
• isTransientLlmError unit-tested: HTTP/2, undici .code, .cause chain, network/abort, retryable vs non-retryable status, deterministic failures (JSON parse, schema reject), self-referential cause chain
• pnpm typecheck — clean
• pnpm test — 1273 passed (131 files)
• pnpm exec biome check src — 0 errors
• pnpm build — green

[tangletools]

✅ No Blockers — 7351f0fd

Readiness 95/100 · Confidence 95/100 · 0 findings (none)

kimi-code: Correctness 95 · Security 95 · Testing 95 · Architecture 95

I read every changed file in full, traced callers/callees, ran the full test suite (1272 tests passed), and verified the build. The PR cleanly deduplicates retry logic between llm-client and judge-retry into a single exported classifier, adds HTTP/2 transport fault detection with cause-chain traversal, and covers the new paths with regression tests. No runtime defects were found.

No findings.

---

_{tangletools · 2026-05-21T21:16:07Z · trace}

[tangletools]

✅ Clean — 7351f0fd

I read every changed file in full, traced callers/callees, ran the full test suite (1272 tests passed), and verified the build. The PR cleanly deduplicates retry logic between llm-client and judge-retry into a single exported classifier, adds HTTP/2 transport fault detection with cause-chain traversal, and covers the new paths with regression tests. No runtime defects were found.

Full findings and scores: review summary

---

_{tangletools · 2026-05-21T21:16:07Z · trace}