feat(retrieval): add BGE-M3 sparse document search

[juemimgcd]

Summary

• add nullable BGE-M3 sparse vectors and a scoped pgvector retrieval path for document chunks
• generate dense and sparse embeddings in one model pass, then fuse dense/sparse/keyword scores with evaluated weights
• keep the feature off by default and fall back to the existing dense + keyword path until every active chunk in an owner/KB scope is backfilled

Evaluation

SciFact, 300 queries / 4,000 documents, at k=10:

Metric	Baseline	Sparse fusion	Change
Recall	0.836778	0.846278	+0.009500
MRR	0.658104	0.686276	+0.028172
NDCG	0.696016	0.718469	+0.022453

The selected fusion weights are dense 0.41, BGE-M3 sparse 0.54, and keyword 0.05.

Validation

• Ruff passed for all changed Python files
• existing non-integration suite: 345 passed, 2 deselected, 8 subtests passed
• Docker Compose configuration rendered successfully
• real BGE-M3 sparse-head output matched the reference score
• PostgreSQL/pgvector smoke test covered sparsevec(250002), HNSW sparsevec_ip_ops, insert, and inner-product retrieval

Rollout and limits

• apply the Alembic migration, enable MEMORY_AGENT_EMBEDDING_SPARSE_ENABLED, then run the existing document embedding backfill
• sparse retrieval is skipped for an entire owner/KB scope while any active chunk lacks a sparse vector
• the offline benchmark used TF-IDF as the PostgreSQL keyword proxy; live PostgreSQL retrieval p95 still needs production-like measurement before enabling broadly