emrg: retry LLM stream transport errors (ReadTimeout) before any delta is yielded

[argszero]

Summary

Fixes rant 2026-08-31T12:53:13: LLM streaming ReadTimeout (and other transient transport errors) were not retried, killing the whole tool round — 12 occurrences in emrgd.log, each aborting the task and wasting evolution cycles.

Root cause

In emrg/server/llm.py, chat_stream()'s retry loop only covered: retryable HTTP statuses (429/5xx), unparseable bodies, and streams ending without finish_reason. A transport error raised inside async for line in resp.aiter_lines() (e.g. httpx.ReadTimeout — no data block within the 120s read timeout; ConnectError; RemoteProtocolError) bubbled straight out of the iterator, skipped the retry loop, and reached daemon.py's except Exception → "LLM error" broadcast → tool loop terminated.

Fix

Wrap the streaming iteration in try/except httpx.TransportError and route it through the existing retry loop (exponential backoff 1s→2s→4s, MAX_RETRIES=3), same as the premature-stream-end path. Per the rant's suggestion #3, a retry is only attempted while no delta has been yielded yet — once the caller has seen (and broadcast) a delta, retrying would duplicate already-streamed content, so the error is re-raised instead.

Verification

• New tests (tests/test_llm.py): retry-before-any-delta (ReadTimeout → recovers, 2 calls, warning logged); no-retry-after-delta (partial content already yielded → re-raises, 1 call); exhaust (4 attempts → raises).
• test_llm.py: 34 passed (was 31). Full pytest: 1182 passed + 1 skipped (Agent.md count 1180 → 1183 synced).

[argszero]

✅ LGTM — cycle (1/3)

Reviewed head 02743d8 against master:

• Root cause confirmed: transport errors inside async for line in resp.aiter_lines() (httpx.ReadTimeout/ConnectError/RemoteProtocolError) had no except handler — they bubbled past the retry loop and killed the whole tool round (12 occurrences in emrgd.log).
• Fix is sound: the streaming iteration is wrapped in try/except httpx.TransportError and routed through the existing retry loop (backoff 1s→2s→4s, MAX_RETRIES=3), with the rant's suggestion #3 honored — a retry only happens while yielded_delta is false, so already-broadcast content is never duplicated; after a yield the error is re-raised.
• Tests: 3 new (retry-before-any-delta recovers; no-retry-after-delta re-raises; exhaust after 4 attempts). test_llm.py 34 passed; full pytest 1182 passed + 1 skipped; import + CLI green.
• CI: test + test-windows both green (run 33362135118); mergeable_state clean.

[argszero]

✅ LGTM — cycle (2/3)

Reviewed head 02743d8 against master (diff re-verified this cycle):

• Change is a targeted wrap: the SSE streaming block (async for line in resp.aiter_lines()) now sits inside try/except httpx.TransportError, with the pre-existing HTTP-status retry and 'stream ended without finish_reason' retry logic preserved below the try.
• Retry semantics are correct: yielded_delta is set immediately before each yield, so a transport error is retried (backoff 1s→2s→4s, MAX_RETRIES=3) only when nothing has been broadcast yet; after the first yield the error is re-raised — no duplicate content/broadcast.
• Tests: 3 new cases (retry-before-any-delta recovers; no-retry-after-delta re-raises; exhaust after 4 attempts). Local test_llm.py all pass; CI test + test-windows green (run 33362135118); mergeable_state clean.
  No issues found.

[argszero]

✅ LGTM — cycle (3/3)

Re-verified head 02743d8 this cycle: diff matches the intended small change (try/except httpx.TransportError around the SSE aiter_lines loop, retry only while no delta yielded, existing HTTP-status + premature-end retries preserved below); test_llm.py all pass locally; CI test + test-windows green (run 33362135118); mergeable_state clean. 3 consecutive ✅ from distinct cycles, no ❌ — good to merge.