Embedding materialization hangs on 23K+ chunks during build

[marklubin]

Problem

fastembed hangs during _materialize_search_surface() when processing ~23,797 reference chunks with modes=["fulltext", "semantic"]. Process shows 0% CPU — deadlock or blocked I/O.

Investigation so far

• fastembed standalone: 23K docs in 70s (works fine)
• fastembed from asyncio executor thread: 12s (works fine)
• The hang is in the synix indexer path, not fastembed itself
• FTS5-only materialization for same chunk count works (takes ~16 min)

Environment

salinas: AMD CPU, 120GB RAM, Python 3.13.12, fastembed 0.8.0, onnxruntime 1.24.4

Current workaround

Reference surface uses modes=["fulltext"] only. Main surface (7 artifacts) has semantic enabled.

Next steps

• Add logging inside _generate_embeddings and _materialize_search_surface
• Test with actual build path at increasing chunk counts (100, 1K, 5K, 10K, 23K)
• Check if the hang is in SQLite write after embeddings, not the embedding generation
• Check ONNX thread settings (OMP_NUM_THREADS=1)

See docs/issues/fastembed-hang-23k-chunks.md