Qwen3-1.7B fine-tuned for HZPlan voice floor-plan editing (MCP tool calls), F0.9 rebalance recipe (same f09-dataset as the 0.6B). Fused from W&B f09-17-adapter, mlx q8.

Edit specialist with a QA trade-off (F0.8-era held-out packs, same scorer as all baselines):

Use for edit-heavy sessions (relative placements, compositions) at ~2× 0.6B latency; prefer the 0.6B (hzplan-edit-06-f09-q8) or 4B for QA-heavy use. The QA collapse at 1.7B (not seen at 0.6B on identical data) is an open investigation — see experiments/2026-07-10-h6-f09-rebalance.