fix: parse pgvector embeddings before cosine rescoring

[leonardsellem]

Summary

Fix a Postgres/Supabase search bug where gbrain query could render [NaN]
scores after cosine rescoring.

The root cause was that getEmbeddingsByChunkIds() could return pgvector values
as strings (for example "[0.1, 0.2, ...]") instead of Float32Arrays. Those
string values were then passed into cosine similarity math, which produced
NaN scores.

This change adds a shared embedding parser and uses it in the Postgres path
before cosine rescoring.

Changes

• add parseEmbedding() in src/core/utils.ts
• use parseEmbedding() in rowToChunk()
• use parseEmbedding() in PostgresEngine.getEmbeddingsByChunkIds()
• add regression tests for pgvector-string parsing in test/utils.test.ts

Why this is needed

Before this fix, hybrid search itself was returning valid RRF scores, but the
Postgres/Supabase cosine rescoring step converted them into NaN because the
embedding type coming back from the DB was not normalized.

After this fix, query results render normal numeric scores again and semantic
ranking works as expected on the Postgres engine.

Test Plan

Run the targeted unit tests:

bun test test/utils.test.ts test/search.test.ts

Manual verification against a real Supabase-backed brain:

gbrain query "what happened in ayor pmo meetings"

Before this fix, the query output showed [NaN] scores.
After this fix, the query output shows normal numeric scores like [0.8315].

[garrytan]

Hey @leonardsellem — thank you. The Supabase NaN-score bug was a sharp catch. PR #196 (v0.12.1 hotfix) bundles your fix with the JSONB double-encode work since they're the same Postgres-string-typed-data class. Your parseEmbedding() helper is the canonical implementation; rowToChunk delegates to it everywhere now.

One small adjustment per Codex outside-voice review during planning: parseEmbedding now throws loud on malformed vectors (e.g. '[abc, def]') instead of silently returning null, to prevent a future silent-NaN class. Non-vector strings still return null (treated as 'no embedding here'). Your tests pass unchanged.

Re-implemented rather than cherry-picked because the file overlapped heavily with #187's rowToChunk update — single coherent commit was cleaner. Co-authorship preserved.

Mind if I close this PR in favor of #196? Thanks for the fix.

[leonardsellem]

Hi @garrytan, no worries, of course you can close this PR and merge #196 instead, as long as it makes gbrain even better! Thank you for bringing such an upgrade in the way I declutter my organic brain!