fix(migrate): parse PGLite string embeddings before re-serializing for Postgres

[ShadowRaptor]

Summary

gbrain migrate --to <postgres-url> (e.g. PGLite → Supabase) fails on the very first page with:

invalid input syntax for type vector: "[[,-,0,.,0,1,2,5,4,2,7,2,5,,,-,..."

The corrupted vector value is the embedding being iterated character-by-character with commas.

Root cause

pglite-engine.ts:getChunksWithEmbeddings returns the embedding column as a JSON-stringified array (e.g. "[0.1,0.2,...]"), not a Float32Array. PGLite's pgvector returns vector columns this way.

The migrate then passes those chunks to postgres-engine.ts:upsertChunks which serializes embeddings:

const embeddingStr = chunk.embedding
  ? '[' + Array.from(chunk.embedding).join(',') + ']'
  : null;

When chunk.embedding is a string, Array.from(string) iterates characters (["[", "0", ".", "1", ",", ...]). Joining with , produces the malformed "[[,0,.,1,,,0,.,2,..." and pgvector rejects it.

The sister method pglite-engine.ts:getEmbeddingsByChunkIds already does the right thing — parses the string back to Float32Array. getChunksWithEmbeddings was missing the same defensive parse.

Fix (defense in depth)

1. pglite-engine.ts getChunksWithEmbeddings — parse string embeddings to Float32Array at source, mirroring the pattern in getEmbeddingsByChunkIds.
2. postgres-engine.ts upsertChunks — also accept already-stringified pgvector-format embeddings (pass-through), in case any other call site hands us a string.

13 lines added, 2 lines changed.

Reproduction

Any PGLite source brain with chunks.embedding populated:

gbrain init                          # creates PGLite brain
# add some content with embeddings
gbrain migrate --to <postgres-url>   # fails on first page

Verification

After patch: migrate completes successfully, embeddings transfer intact.

Verified on a real 153-page PGLite → Supabase migration (472 chunks, 1536-dim text-embedding-3-large vectors). Pre-patch failed immediately; post-patch completed in ~3 minutes with all embeddings intact.

Notes

• No new tests added — happy to add them if you want a specific test format. The fix follows the same defensive pattern already in getEmbeddingsByChunkIds, so it's symmetry rather than a new contract.
• Found while migrating a 0.10.2 brain. Verified bug still present on master (81b3f7a).
• Separate (lower-priority) finding from the same investigation: gbrain put returns the misleading error Page not found: <slug> when the slug contains uppercase letters. I'll file that as a separate issue.

🤖 Generated with Claude Code