Refusal Changes the Future: Rollout-Equalized Credit Assignment for Sequential Memory Maintenance

A refused merge is a state transition: preserving two records changes retrieval and every later maintenance opportunity. Rollout-Equalized Refusal (RER) measures this dependence by freezing the pre-mutation refusal transition in paired reruns; a second intervention applies one state-conditional candidate-construction rule to both controllers. Its operator-conditional estimand admits exact pathwise accounting, and a two-step construction shows how equal refusal coverage or logged filtering can fail to identify the rollout quantity. At its complete-policy 200-slot operating point, MERQ reaches 78.4% task success and 3.2% method-blind, policy-level per-decision contradiction burden across 8,400 episodes, gains 1.6 task points over Compressive Memory, and uses 22.1% fewer slots than saturated LRU/LFU. Under the frozen MERQ-associated operator, its advantage over same-action-space RL is $+0.7$ task and $+1.8$ favorable contradiction points on StreamMem-Bench and has the same direction on LongMemEval, LoCoMo, and AgentOps-90d. Under the common construction rule, $+0.6[0.31,0.88]$ task and $1.3[1.02,1.58]$ favorable contradiction points remain; against independently trained Selective PPO at nearly identical refusal coverage, the margins are $+0.4[0.12,0.69]$ and $0.9[0.62,1.18]$. A 3,600-decision blinded audit measures 87.5% safe-merge recall, 92.4% unsafe-merge recall, and Fleiss' $\kappa =0.86$. These results establish refusal as a rollout-level evaluation object and localize MERQ's retained advantage to post-generation control under the declared operator and candidate-construction rule.