fix(error-classifier): classify HTTP 408 as retryable timeout, not a non-retryable 4xx

[hashbender]

Problem

On a long session against GitHub Copilot, the chat turn would occasionally end
as a blank/empty assistant message (no content) with a visible
Error: HTTP 408: Timed out reading request body. Try again, or use a smaller request size. The turn aborted instead of recovering.

Root cause

agent/error_classifier.py::_classify_by_status had no branch for HTTP 408.
It fell through to the generic tail:

# Other 4xx — non-retryable
if 400 <= status_code < 500:
    return result_fn(FailoverReason.format_error, retryable=False, should_fallback=True)

But 408 is a timeout, not a permanent client error. GitHub Copilot returns
408 {'code': 'user_request_timeout', 'message': 'Timed out reading request body. ... use a smaller request size.'} when a large prompt is slow to read.
Classified as non-retryable format_error, the loop took the client-error abort
path and persisted an empty assistant turn (the blank bubble).

Fix

Add an explicit 408 branch classifying all 408s as a transient timeout
(retryable=True, not should_compress). The loop's existing transport
retry + eager-fallback-after-2-attempts path then recovers the turn.

Why NOT route it to compression (payload_too_large)

The 408 message literally says "use a smaller request size", so compressing like
a 413 is tempting — but field data argued against it. On a long copilot/opus-4.8
session the 408 was probabilistic jitter well below the hard prompt ceiling:
the same ~785k-token request that 408'd once succeeded on the very next attempt
at ~786k. Retrying the same body usually works, so auto-compaction on a 408
would silently delete conversation history for a merely transient timeout.
Genuine over-window prompts already surface as 413 / context_overflow and keep
their own compression path; a user can compact a 408-prone long session
deliberately via /compress.

Testing

• Classification (tests/agent/test_error_classifier.py::Test408RequestTimeout):
  oversized-body 408 and plain 408 both → timeout / retryable / not-compress,
  plus a test_408_never_auto_compresses guard.
• Loop-level (tests/run_agent/test_408_request_timeout_loop.py): a 408 in
  the real run_conversation loop retries into a real assistant turn (no blank
  turn), never calls _compress_context, and resends the same request body.
• Mutation-verified: deleting the 408 branch makes both suites fail with
  format_error/non-retryable (reproducing the blank-turn regression),
  confirming the tests guard the symptom rather than being tautological.

Full suite green on latest main: tests/agent/test_error_classifier.py +
tests/run_agent/test_408_request_timeout_loop.py (188 passed), and the sibling
tests/run_agent/test_413_compression.py unaffected.

---

Mirror-of: NousResearch#56932
NousResearch#56932