Separable Levels of Observability in a Truth-Known Memory Simulator

We built a simulator where the true memory dynamics are known, then asked how much of that truth survives under non-oracle partial observations (synthetic branch-summary measurements, not calibrated calcium/EEG models). Oracle finite-difference consistency, propagation calibration, and frozen-writing ablations are separated from non-oracle measurement claims. Observable charts are selective: some motifs support mechanism discrimination, the robust order result is chain inversion, and a lagged branch summary cannot recover a response-generating local operator. The contribution is a map of measurement strength, not a claim that dendritic accessibility is consciousness or that any chart is empirical MNPS.