Q4 — Embedding model bottleneck: GTE-Large FP16 → Qwen3-Embedding-0.6B

[longobucco]

Problem

GTE-Large FP16 (~670 MB, 512-token context, English-centric) is outperformed by Qwen3-Embedding-0.6B as of April 2026:

Model	MTEB avg	Context	VRAM
GTE-Large FP16	63.1	512 tok	~670 MB
Qwen3-Embedding-0.6B	68.4	8192 tok	~360 MB

Qwen3-Embed-0.6B is smaller, better, and multilingual — relevant for Italian lecture notes + English whitepapers. The 8192-token context also eliminates the child-chunk truncation issue (R4) entirely.

Fix

When `@qvac/sdk ≥ 1.0` ships `QWEN3_EMBEDDING_0_6B`, update `QVAC_EMB_SRC` in `workers/qvac-service/src/models.js`. Re-ingest all documents after the switch.

Effort: Small (config change + re-ingest)
Impact: High — direct quality improvement, smaller memory footprint