Easy to Detect, Hard to Judge: Signed MoE Route Flips under KV-Cache Quantization and the Limits of Token-Wise Selective Repair

Top-k MoE routing is discontinuous, so a numerical disturbance upstream of a protected gate changes which experts fire. Under simulated 4-bit KV-cache quantization read by a BF16 gate, the resulting route flips are signed: some help and some hurt, in near-equal measure, so a blanket correction nets almost nothing while a directional one could recover value. We ask whether the direction is available per token. On OLMoE-1B-7B the router margin predicts flip occurrence (AUC 0.772) but no tested local or cross-layer router statistic predicts flip sign above chance, and the deployable margin gate performs below a rate-matched random gate of equal budget on both architectures measured for it. Oracle evidence is architecture-dependent and we report it as such: on OLMoE, oracle-gated boundary smoothing is excluded beyond approximately 5.1% of the harmful component and oracle clean-route pinning shows no detected advantage over a rate-matched random gate (+0.4%), while on DeepSeek-MoE-16B the same oracle contrast is positive (+2.3%), so the strong oracle-null does not generalize and keeps an OLMoE subscript; both contrasts are N=384 and their interaction leaves it unresolved whether the architectures differ. A prefix reference-bearing intervention recovers roughly a quarter of the routed ceiling at a per-route byte cost far below a prefix-KV upgrade. We supply no method: this is an empirical limit on the tested token-wise predictors and interventions under one simulated disturbance, measured in teacher-forced NLL, with downstream task consequences unmeasured.