Wire vendored nomic-embed-code (pretrained-embed) to brain search — replace hashing-trick 256d #455

Goal

Replace the current hashing-trick 256-dim static embedding in brain search with the vendored nomic-embed-code model (already in codebase via pretrained-embed feature flag, 31MB, 768-dim).

Context

• Current: embed_code() in src/embed/tokens.rs uses a deterministic hashing trick (256-dim vectors)
• Already exists: vendored nomic-embed-code model in vendored/nomic/ (768-dim, 31MB binary)
• Feature flag pretrained-embed already gates this code path
• Brain search (FTS5 + vector + graph RRF fusion) in src/index/brain.rs has no_provider mode fallback

Scope

1. In src/index/brain.rs: replace no_provider mode with pretrained_embed provider
2. Ensure 768-dim vectors are stored in usearch index instead of 256-dim
3. Handle migration: if existing index has 256-dim vectors, flag for re-index
4. Keep hashing-trick as ultimate fallback (no model available scenario)

Files

• src/index/brain.rs — core brain/index logic, provider selection
• src/embed/mod.rs — embed module entry
• src/embed/tokens.rs — current hashing-trick + pretrained model interface

Trade-offs

• +768-dim real embeddings (vs 256-dim hashing-trick) → better semantic search
• +31MB binary overhead (vendored model)
• +ONNXRuntime dependency for pretrained path
• Fallback chain: pretrained-embed → hashing-trick → FTS5-only