fix: reduce DiskANN index storage overhead

[amandeavor]

Summary

• store DiskANN neighbor vectors as float8 for both content and tag indexes
• cap max_neighbors at 20 so high-dimensional embeddings do not inherit the much larger libSQL default
• assert the persisted SQL definitions and exercise direct vector_top_k lookups through both tuned indexes

Scope

This deliberately changes newly created indexes only. Existing shards retain their current index definitions until a separate failure-safe rebuild/VACUUM migration is designed, so this patch does not risk leaving an existing shard without an index.

Addresses the new-shard portion of #277.

Validation

• bun test tests/turso-vector-search.test.ts -t "inserts and searches memories" — 1 pass
• bun run typecheck — pass
• bun run build — pass
• full local Windows suite — 446 pass; 2 unrelated platform/timing failures: an EBUSY file rename in the legacy migrator test and the existing 30-second 200-vector ANN test timing out at 35.5 seconds. The latter was slower and eventually crashed Bun when repeated with the unmodified upstream index definition.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lindixu6-hash]

Reviewed the new-shard index definition and direct-query coverage. The change is intentionally scoped to newly created content/tag indexes; existing-shard rebuild and VACUUM remain a separate migration concern. Local verification at 88221cd: targeted vector tests 3/3, full suite 447/447 with 1 skip, typecheck, and production build all pass. The approved GitHub package-smoke matrix is also green on Ubuntu, Windows, macOS Intel, and Apple Silicon.