Hy3 Preview routes most refactor steps

This is a solid signal for hybrid agent routing: cheap open models can absorb the long tail of routine coding work, but you still want a stronger model for tangled dependency edits and path-sensitive refactors.

• –The cost split is the point: most steps were cheap enough to keep the overall run near single-digit dollars until Opus had to rescue the hard cases.
• –`enable_auto_tool_choice` plus parser compatibility mattered as much as the model choice; orchestration friction can dominate the actual routing logic.
• –`no_think` on routine steps is a practical cost lever, but it only works when the task is already easy enough to classify correctly.
• –The failure mode is predictable: circular imports and missing-module hallucinations are exactly where a 21B-active MoE still loses to a frontier model.
• –The DeepSeek V4 comparison suggests accuracy parity is not enough if search-loop latency makes the agent materially slower.