Freeze the M0 evaluation bars

[jmjava]

Summary

• Freeze the general-coding floor: tests, repair, multi-file edit, and explanation. A release fails when any one of those drops more than 5 points below the pinned base model.
• Freeze the Embabel gain: compile rate, full-task success, hallucinated-API rate, planner success, and repair iterations. The lift counts only when every metric meets its minimum.
• Freeze the host budget: 7B–14B first. A 32B-class model (30B–34B) waits until docs/runner-stable.md records a runner that fails a known-bad fixture. Downstairs hosts score first. The IDE workstation stays the MCP host.

Test plan

• python3 scripts/check_m0_bars.py --self-test rejects a quiet edit of the 5-point drop
• python3 scripts/check_m0_bars.py --against origin/main accepts this initial freeze
• GitHub Actions m0-bars is green
• Merge only after that check is green

Made with Cursor