fix(harness): classify undici stream-timeout terminations as transient network failures

[LEFTEQ]

Fixes #1037.

Problem

When a provider/proxy response stream stalls past undici's headers/body timeout, the fetch is killed with TypeError: terminated (cause BodyTimeoutError / HeadersTimeoutError, codes UND_ERR_BODY_TIMEOUT / UND_ERR_HEADERS_TIMEOUT). The semantic-error catalog recognizes neither the codes nor the message, so classifyModelCallError falls through to "recoverable" — and in task mode the harness fails the run outright with zero retries. A transient transport stall becomes a terminal run failure (hit in production behind a LiteLLM proxy; details in the issue).

Fix

• Add UND_ERR_BODY_TIMEOUT and UND_ERR_HEADERS_TIMEOUT to NETWORK_ERROR_CODES in semantic-errors/rules/system.ts — the coded rule wins whenever the cause chain survives.
• Add "terminated" to the exact-equality message fallback, same stripped-cause rationale as "fetch failed" (undici's Fetch.terminate rejects with a TypeError whose message is exactly terminated). Equality, never containment.

Both signals now summarize as network-request-failed (tag transient) → classifyModelCallError returns "retry" and runModelCallWithRetries handles it like any other transient network failure.

Tests

• semantic-errors.test.ts: coded shape (terminated + UND_ERR_BODY_TIMEOUT cause) and bare TypeError: terminated both classify as network-request-failed.
• model-call-error.test.ts: the body-timeout shape classifies "retry" alongside the existing structural-signal cases.

pnpm vitest run on both files: 71 passed.

Note on commit signature: this commit carries a DCO sign-off but is not yet GPG/SSH-signed — happy to re-sign and force-push once my signing key is set up, or feel free to squash-merge.

[vercel]

Someone is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.