fix(error_classifier): stop empty-response advisories from triggering compression

[HexLab98]

Summary

• Provider empty-response advisories (e.g. nano-gpt / OpenRouter) say "very low max_tokens" as a possible cause. That literal substring used to match _CONTEXT_OVERFLOW_PATTERNS, so Hermes misclassified the error as context overflow, entered the compression loop, and ended in "Cannot compress further" on otherwise healthy sessions.
• Tighten overflow matching to specific `max_tokens exceeded` phrasing and classify empty-response advisories as retryable `server_error` without compression.

Test plan

• `scripts/run_tests.sh tests/agent/test_error_classifier.py -q` (199 passed)
• Reproduce with a custom endpoint that returns the empty-response advisory and confirm Hermes retries instead of compressing.

[alt-glitch]

This was generated by AI during triage.
Related to the prior empty-response handling work in #5278 and #4552, but this patch fixes the distinct false context-overflow/compression classification.