Bug Description

gbrain embed (both --stale and --all) hangs indefinitely on PGLite backend without making any HTTP requests to the embedding provider. The process runs for 90+ seconds with zero output and zero network activity, then requires SIGTERM to kill. gbrain dream --dry-run correctly detects unstale chunks, but the actual embed execution deadlocks internally.

Environment

• gbrain version: 0.35.1.0
• OS: macOS 12.7.6 (Intel Core i7-6700HQ)
• Backend: PGLite (~/.gbrain/brain.pglite)
• Embedding providers tested: Ollama (nomic-embed-text), MiniMax (embo-01), custom OpenAI-compatible proxy → all same behavior
• Installation method: bun add -g gbrain

Steps to Reproduce

1. Have a brain with unstale chunks (any page with embeddings not yet generated):

   gbrain dream --dry-run  # confirms N chunks pending embed

2. Run embed:

   gbrain embed --stale
   # or
   gbrain embed --all

3. Observe: process hangs with no output. No HTTP requests reach the embedding provider (verified via proxy logs and network monitoring).

4. SIGTERM required to stop the process (kill <pid>).

Expected Behavior

gbrain embed --stale should iterate through unstale chunks, send embedding requests to the configured provider, and update the vector index in PGLite.

Actual Behavior

Process hangs silently. Key observations:

• No HTTP requests are ever sent — verified with a local proxy (http://localhost:4000) that logs every incoming request. Zero requests logged.
• No stdout/stderr output — not even progress indicators or error messages.
• gbrain dream --dry-run works correctly — it detects the unstale chunks, so the database query path is functional.
• Same behavior across all providers — Ollama (local), MiniMax (remote), and a custom OpenAI-compatible proxy all exhibit identical behavior, ruling out provider-specific issues.
• PGLite-specific — this appears to be a deadlock or blocking issue in the PGLite embed execution path.

Additional Context

• gbrain doctor and gbrain stats work normally on the same PGLite database.
• gbrain search (keyword/tsvector) works fine.
• The issue is not a provider timeout — no requests leave the process at all.
• Previous attempts included splitting large pages into smaller ones (2-3 chunks each) to rule out payload size issues — same result.
• Multiple Ollama restarts and full system reboots were attempted — no change.

Workaround

None found. Keyword search (gbrain search) works, but semantic/vector search is unavailable until embed completes.