feat(knowledge): enable hybrid vector retrieval

[Ldsystem]

Summary

• replace the hashed 64-dimensional lexical vector with pinned FastEmbed BAAI/bge-small-en-v1.5 embeddings
• remove runtime package-manager shell-outs and declare the ks.py PEP 723 uv environment
• query knowledge_chunk_vec, validate index compatibility, and fuse FTS/vector candidates with deterministic RRF
• preserve truthful FTS fallback, exact-identifier strength, provenance, and the script/agent authority boundary

Validation

• model smoke: 384 dimensions; paraphrase outranks unrelated noise
• focused WOR-58 contract: 14 passed
• full Linux-equivalent suite: 578 passed
• macOS-portable suite: 577 passed, 1 deselected
• uv run scripts/ks.py --help: passed
• git diff --check: passed

Scope

This closes production gaps G1 provisioning, G2 representation, and G3 query wiring under the verified WOR-58 specification. BFS/backlink work and WOR-54 macOS sqlite-vec portability remain out of scope.

Refs WOR-58.