[xmh1011]

Summary

• preserve the chunk-level exception cause when fact extraction fails
• classify retryable content extraction failures (rate limit, timeout, connection, 5xx) and re-raise them so worker retry handling can run
• keep the existing skip-and-empty behavior for deterministic/non-retryable extraction failures

Closes #1833.

Tests

• uv run ruff check hindsight_api/engine/retain/fact_extraction.py tests/test_fact_extraction_retry.py
• uv run pytest tests/test_fact_extraction_retry.py tests/test_experience_fact_type_passthrough.py -q
• ./scripts/hooks/lint.sh

[nicoloboschi]

the transient errors are already handled at llm level
see HINDSIGHT_API_LLM_MAX_RETRIES
https://hindsight.vectorize.io/developer/configuration

if you think we're missing some errors in there that should be retried, please adjust this pr or create a new one, thanks