router v2: split entropy into rival mass + smear

[solarkyle]

rival mass (candidates 2-5) and tail smear predict errors very differently, see assets/figure5b_states.png. current router uses one entropy number. retrain with the split features from the uncertainty_shape traces, might beat v1. analyze_router.py has the training code

[solarkyle]

ran it on all 5 models. the reframe holds everywhere: the entropy metric correlates with tail smear at 0.93-0.999 on every model, so it was always a noise gauge. smear beats rival mass for predicting wrong answers on all 5. and 'conflicted is the safest state' holds on all 4 gemmas (18-35% wrong vs 53-75% for mushy). qwen flat as usual.

but the v2 router with split features is basically a wash: combined AUC moves at most +0.01. since entropy is nearly identical to smear, the old router was already using the good feature. so the split matters for understanding, not accuracy. maybe rival mass helps on other tasks or as a separate 'deliberating' flag, open to ideas