Qwen3-1.7B fine-tuned for HZPlan voice floor-plan editing (MCP tool calls), F0.8 recipe (compositions + declarative phrasing + robustness). Fused from box adapter lora-f08-17-final (trained 2026-07-08, 225 W run), mlx q8.

The intended sweet spot between the 0.6B (fast, capacity-bound on F0.8 compositions) and the 4B (best quality, slow on phone). No full eval pass yet — smoke: 6/8 on the live agent loop (T1 2/2, T2 4/5 incl. add-object-centered, T3 0/1), ~4.2 s/case median on Mac q8.

Siblings: hzplan-edit-06-f07a-q8, hzplan-edit-06-f08-q8, hzplan-edit-4b-f08-q8.