Retrieval: dense brute-force + hard filter + artifact dedupe

[thorwhalen]

Refs #1.

ir.retrieve.search(corpus, query, *, k=10, filter=None, surfaces=None):

• embed query (input_type='query').
• hard metadata filter via vd.filters.matches_filter over record metadata (ownership / name / tags) — pre-filter the candidate set.
• dense: brute-force cosine over filtered vectors (numpy), top-k.
• dedupe to artifact level (best surface per artifact).
• Seams (next): hybrid BM25 + vd.reciprocal_rank_fusion (ir_02: matters for short identifier-heavy capability text), and ef.with_reranker.

Return SearchHits + egress helpers.

[thorwhalen]

Decisions (done in PR #14):

• Dense brute-force is the v1 default: exact cosine over an L2-normalized matrix, loaded from the store and cached in-process. Correct and instant at ir's sizes (≤ ~50k vectors).
• Hard metadata filter pre-applied via vd.filters.matches_filter over record metadata (ownership / name / tags), narrowing candidates before ranking — the "filtering is not all embeddings" requirement.
• Per-artifact dedupe with over-fetch (collapse multiple surface-hits to the best surface per artifact).
• Deferred seams: hybrid BM25 + vd.reciprocal_rank_fusion, and ef-style reranking. ir_02 finds hybrid materially helps short, identifier-heavy capability text (skills/packages) — so this is the highest-value next retrieval increment.