D106 — Benchmark gates compare violation multisets

Decision

Native benchmark gates compare physical violation multisets; routing also evaluates error-severity starved_thermal and isolated_copper without requiring a reported track. Placement retains its rigid-single-footprint exclusion, except zone-dependent thermal/island findings. KiCad 10.0.x evaluated error caps are unknown, never clean or a passing comparison; warning-only caps remain diagnostic. Existing completion remains the capped DRC-list ratio, with saturation disclosed separately (doc 10 §4)

Alternatives

Count deltas by type; ignore all pad/zone-only findings; import an uncapped connectivity worker and change completion here

Reason

A removed clearance elsewhere cannot excuse a new one; refills can reveal copper damage while naming only pads or zones. Saturated lists cannot prove a zero delta. This isolates error-gate correctness from connectivity metric changes and preserves CLI/report compatibility