Score left and right separately - averages hide chirality breaking that mirror augmentation does not prevent

Symptom

Policies developed quantified left/right asymmetry (e.g. C2-700 turned right at 82% but left at 67% - a 15 pp gap; push tolerance 40/40 symmetric on the root vs 17/40 on a deep-trained descendant), and averaged metrics would have reported healthy midpoints.

Context

The repo had policy-level symmetry-breaking evidence strong enough to make separate scoring a battery rule: "左右必须分开打分 … 平均 vy 跟踪会把它掩盖". Notably, chirality broke and never recovered even though mirror augmentation (command-level mirror_prob 0.5) was on the whole time - augmentation reduced but did not prevent asymmetry, and once broken it stayed broken through subsequent rungs. PASS conditions therefore carried explicit symmetry budgets (left/right tracking gap <=10 pp), and sim's predicted asymmetry (700: right faster than left) was flagged for direct real-robot timing confirmation.

Change

Battery rule: every directional skill reports left and right (CW/CCW) as separate rows with a max-gap budget; mirror augmentation treated as mitigation, not proof of symmetry.

Outcome

The 700-vs-A800 asymmetry gap (15 pp vs 7 pp) became a first-class selection criterion; C4 product shipped with a measured 5 pp gap.

Mechanism

Averaging over mirrored conditions cancels antisymmetric error exactly where it matters; and symmetry lost during training is a lineage injury (like plasticity loss) that later rungs do not spontaneously heal, so it must be gated, not assumed.

Applies when

• evaluating turn/sidewalk/push-recovery or any mirrored skill
• relying on mirror/symmetry augmentation
• selecting between checkpoints with similar average scores