feat: Log failed LLM calls to PromptLayer before re-raising

[adagradschool]

Summary

When .run() calls an LLM provider and the call throws (rate limit, auth, timeout, etc.), the SDK now catches the exception, logs it to PromptLayer with status=ERROR, error_type, and error_message, then re-raises. Also exposes these fields on log_request() for manual callers. Error categorization is duck-typed — no provider imports in runtime code.

Edge cases tested

• Tracking failure doesn't mask the original LLM exception
• Success path doesn't include error fields
• status_code=402 quota detection
• Generic exceptions with "quota"/"timeout" in message only match for known provider modules (not arbitrary errors)
• Real OpenAI/Anthropic exception classes (rate limit, auth, timeout, bad request)

Known concern

When PromptLayer API is down, the error-tracking call in the except block may add latency due to retries (via track_request/atrack_request retry logic) before the original exception propagates. Consider a no-retry or shorter-timeout path for error logging.

Test plan

• tests/test_error_categorization.py — 11 parametrized cases
• tests/test_error_logging.py — 5 cases (sync/async run + success path)
• Full suite passes (71 tests), ruff lint + format clean

🤖 Generated with Claude Code