scoring: per part breakdown of full system quality

[BANADDA]

Splits each certified full system's result into its parts so miners can see which part to improve. Information only: pay still follows the whole system's place on the frontier.

Part	Measured as	Cost
Small model	quality alone minus the arena's quality_floor (untrained base model under the harness)	none, from traces
Harness	on the verification sample: the archived small model with the plain task prompt vs the same model through the harness; plus the output step's gain on every task	one plain prompt generation per sampled task
Router and escalation	escalation gain, rescues, waste, misses, precision, escalation spend per rescue	none, from traces

quality = small_quality + escalation_gain + output_gain holds exactly.

Isolation. The harness probe runs in its own jailed call after certification. If it fails or runs out of budget, the harness part reads "not measured"; certification and scores are untouched.

Where it shows. The report's calibration.breakdown (the server's escalations summary picks it up), mt miner status, mt miner simulate (router and output step parts), the published card, and a new section in docs/system_miner.md.

Floor reaches the validator through the anchored arena config (RoundBudget.quality_floor).

🤖 Generated with Claude Code