Matryoshka prefilter: adopt SearchDimension for large-corpus query latency

[autholykos]

Problem

Pituitary's vector search runs at full stored dimension today. For small corpora (< ~5k chunks) that's fine. For larger corpora — especially the cross-repo governance use case (#173) and multi-repo workspaces (#228) — full-dim cosine across thousands of chunks becomes the dominant query cost.

Proposal

Adopt stroma v2's matryoshka prefilter via SearchParams.SearchDimension:

1. Run the first-pass vector scan at a smaller truncated dimension (e.g. 256 for an embedder trained with matryoshka loss).
2. Rescore the shortlist with full-dimension cosine.

Result: latency drops roughly proportionally to the prefilter dimension, recall stays within ~1-2% of full-dim for matryoshka-trained embedders.

Why this matters for Pituitary

• Unlocks cross-repo governance (RFC: Cross-repo spec governance #173) and multi-repo workspaces (bench: add multirepo benchmark workspace coverage #228) at realistic corpus sizes without a CPU budget explosion.
• Enables richer retrieval features (wider shortlist for rerankers, contextual retrieval, spec-family-scoped search) by making each individual vector scan cheaper.
• Improves watch-mode / pre-commit hook responsiveness.

Implementation notes

• Configuration surface: runtime.search.prefilter_dimension = 256 (or similar). Default 0 = full-dim, no prefilter.
• Only meaningful for matryoshka-trained embedders (most OpenAI models, some OSS models). Gate behind embedder capability detection or explicit opt-in to avoid recall regressions on non-matryoshka models.
• Three call sites today embed VectorSearchQuery: internal/index/search.go, internal/analysis/semantic_terminology.go, internal/analysis/repository_similarity.go. All three need to thread the prefilter config.
• Benchmark against representative corpus to confirm recall stays within tolerance.

Acceptance criteria

• SearchParams.SearchDimension threaded through all three vector-search call sites
• Config surface with safe default (off)
• Embedder-capability gate (don't enable on non-matryoshka embedders without explicit opt-in)
• Benchmark: latency reduction AND recall@k on a representative corpus at ≥5k chunks
• Documentation explaining when to enable and at what dimension

Context

Unlocked by stroma v2.0.0 (merged in #337). Part of the Phase 4 stroma adoption plan.