B-side LLM-only memory extractors - #6

Summary

• Replace B-side hardcoded/rule-based memory extraction with LLM-only semantic extraction.
• Keep existing public extractor class names and method signatures compatible with MemoryEvent -> MemoryCandidate.
• Keep local logic limited to prompt packaging, sensitive-data redaction, JSON-to-candidate conversion, validation, deduplication, and conflict annotation.
• Add demo and documentation for before/after dataset input-output display.

Verification

python -m pytest tests/test_extractors.py tests/test_llm_memory_extractor.py tests/test_environment_extractor.py tests/test_tool_extractor.py tests/test_workflow_extractor.py tests/test_ingestion_to_extractors.py -q

Result: 30 passed

python demo\b_llm_pipeline_demo.py

Demo prints raw MemoryEvent[], sanitized LLM input JSON, fake LLM JSON output, and final MemoryCandidate[].

Note

This PR supersedes the previous B LLM pipeline PR. The older PR should not be used for review.