fix(extraction): retain errors for empty LLM responses

[Kuang-xianxin]

Summary

Keep empty LLM completions marked as extraction errors in both LLMExtractionStrategy.extract and aextract.

When a provider returns content=None or "" (for example with finish_reason="length" or "content_filter"), the existing code creates an error block and then overwrites its error flag with False in the successful-block loop. Whitespace-only content instead disappears as an empty list. Downstream code filtering successful records by error can therefore accept diagnostics as extracted data or miss the failure entirely.

Return the empty-response diagnostic before successful-result normalization, retain the actual section index and finish reason, and treat whitespace-only content as empty. Usage accounting still runs before this return. A valid model response containing [] continues to mean successful extraction with no matching records.

Files changed

• crawl4ai/extraction_strategy.py: the same narrow empty-content guard in sync and async extraction.
• tests/unit/test_llm_extraction_empty_response.py: parameterized regressions using real LiteLLM ModelResponse objects with mocked completion boundaries; no model API calls.

Validation

• Before the fix: 36 regression cases fail, while eight valid-response controls pass.
• After the fix: all 44 new cases pass, covering sync/async, JSON/XML-block parsing, None/empty/whitespace content, stop/length/content_filter reasons, section indices, usage preservation and valid empty result lists.
• Full tests/unit suite: 121 passed on Python 3.13.14 after installing the optional pypdf dependency needed by existing PDF tests.
• Test-file Black check (--target-version py310), compilation of both changed Python files, and git diff --check pass. The existing source file was not reformatted wholesale.

Command: LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync pytest tests/unit -q.

Based on and targeting develop. Found by source inspection; no existing issue is claimed as resolved. Older #1450 addressed alternative response fields rather than this error-flag overwrite.

No live browser crawl or model-provider validation is claimed.