the effective mass is the loss-weighted form, not either proposal

Agent-decided, under AGENTS.md §6, and settled by reading the papers rather than by choosing. Two candidate definitions were put to the round in advance — occasions on which the repair is awake, and probability mass the repair moves — and the reduction produces neither. Khot-Ponnuswami's equation (5) forces the inner player's own loss to zero, so AdaNormalHedge's adaptive quantity collapses to C_T = sum_t I(t) |p^T(M_f - 1) l|, weighing each occasion by how much the repair would actually have changed the incurred loss. A repair moving real probability across actions of equal loss scores full under both proposals and exactly zero under this one.