Summary

Fixes #95 — Memory write pipeline: add deduplication for episodic/event_log and expiry cleanup for foresight.

Problem

1. Duplicate records: When the same MemCell is processed more than once, both episodic_memory and event_log collections accumulate duplicate entries with the same parent_id. This degrades retrieval ranking quality.

2. Stale foresight: ForesightRecord has a validity window (start_time / end_time), but expired records are never deleted — they just accumulate dead data across MongoDB, Elasticsearch, and Milvus.

Changes

1. Delete-before-insert dedup in save_memory_docs()

In src/biz_layer/mem_memorize.py, before inserting new docs:

• Episodic memory: delete existing records with matching parent_id from MongoDB, ES, and Milvus
• Event log: same delete-before-insert by parent_id across all three stores
• Dedup is best-effort: failures are logged as warnings but do not block the insert

2. Foresight expiry cleanup

New cleanup_expired_foresights() function that:

• Queries MongoDB for ForesightRecords where end_time < today
• Deletes those records from MongoDB, Elasticsearch, and Milvus
• Returns the count of deleted records
• Can be invoked periodically (e.g., via cron/scheduler)

3. New delete_by_parent_id on EpisodicMemoryRawRepository

Added the missing method (EventLogRecordRawRepository already had this).

4. Tests

tests/test_write_pipeline_dedup.py covers:

• Episodic dedup deletes old records before insert
• Dedup failure does not block insert (best-effort)
• Event log dedup deletes old records before insert
• Foresight cleanup removes expired records from all stores
• Cleanup returns 0 when no expired records exist