Reranker dies with SIGTERM during deep_search on large index (4.6GB, 40 chunks)

[tilleulenspiegel]

Environment

• qmd version: 1.0.6
• GPU: NVIDIA GeForce GTX 1080 Ti (10.9GB VRAM, 8.7GB free during crash)
• Index size: 4.6 GB / ~559k vectors
• Reranker: Qwen3-Reranker-0.6B-Q8_0-GGUF

Behavior

When running a deep_search query that expands to ~30 parallel sub-queries, the reranking step dies with SIGTERM after processing ~40 chunks. The parallel searches complete successfully, but the reranker is killed before returning results. VRAM is not the bottleneck (8.7GB free at time of crash).

Steps to reproduce

1. Index a large collection (~14k files, 4.6GB, ~559k vectors)
2. Run a deep_search query (expands to ~30 sub-queries)
3. Reranking phase starts with ~40 chunks
4. Process receives SIGTERM and dies — no results returned

Expected behavior

Reranking completes and results are returned.

Notes

• MCP timeout was increased from 15s to 60s — SIGTERM still occurs, so this is not a timeout issue
• qmd embed was not running simultaneously (VRAM contention ruled out)
• Happens consistently on this index size