planner_study.validate_mvprop¶
Mechanistic proof for MVProp: its forward IS value-iteration with ONE learned quantity, the passability w=sigmoid(gain_conv). Claim: distillation makes w ~ 1-wall (true passability), which is a LOCAL map-independent function -> generalizes by construction. Test: (1) does w recover the walls? (2) is w goal-invariant (pure passability)? (3) does the field match the oracle wavefront (and MSP's not)? (4) does w-recovery hold across OTHER map distributions?