feat: IVF index structure (opt-in for large codebases)

[Lucas-Bur]

What to build

Add an optional IVF (Inverted File Index) structure for faster search on large codebases. Instead of comparing the query against ALL vectors (O(n)), partition vectors into clusters at index time. At query time, find the nearest N clusters to the query, then brute-force only those clusters.

Decision: Opt-in for large codebases

• Default: Brute force (current behavior). Exact results, fine for <50k vectors.
• Opt-in: Enable via config.index.ivf.enabled: true or pix index --ivf.
• Clustering: K-means with sqrt(n) clusters. Store centroids + cluster assignments per vector.
• Query: Find nearest N centroids, search only those clusters.

Acceptance criteria

• IVF index built at index time when enabled
• Centroids stored in .pix/ivf-centroids.bin
• Cluster assignments stored per vector in .pix/ivf-assignments.json
• Query uses IVF when enabled, brute force when disabled
• Configurable cluster count and search width
• pix status shows IVF enabled/disabled
• Adapter test verifies IVF search returns similar results to brute force
• Performance test shows speedup on large index (>50k vectors)

Blocked by

None - can start immediately