Did the World Change or Did My Sensor Break? Cheap Physics Residuals Attribute Deployment Faults for Frozen Manipulation Policies

A deployed robot policy that starts failing faces a differential diagnosis: did the world change (an object got heavier, a surface more slippery) or did a sensor break (bias, freeze, noise)? The two families demand opposite responses—re-adaptation vs. re-calibration—yet deployment monitors rarely distinguish them. We inject both families into frozen behavior-cloned policies on three simulated manipulation tasks (object mass $\times 2/\times 4$, friction $\times 0.5/\times 0.1$; end-effector bias, frozen object observations, proprioceptive noise; n=150 per condition per task) and evaluate four training-free residual detectors computed only from the policy-visible stream: a cross-modal parity check, a post-contact-freeze check, a temporal-jitter check, and a command-tracking check. Three findings. (1) Sensor faults are cheaply attributable: the freeze and jitter detectors are perfectly specific (AUC 1.000 on their targets, 0.500 on everything else), and the parity check is perfect (1.000 on all sensor faults) exactly where the observation vector carries a mathematically exact redundancy—on tasks whose redundancy is only approximate ($\sim 2$ cm frame offset), a 2 cm bias becomes invisible, making redundancy precision a sensor-design requirement, not an implementation detail. (2) Moderate world-changes are stealthy: mass $\times 2$ and friction $\times 0.5$ shift no detector above AUC 0.56—and are also behaviorally benign, absorbed by the impedance controller. (3) A pre-declared severity sweep shows that harm onset precedes visibility onset: friction $\times 0.1$ collapses the precision task from 0.35 to 0.05 success while the best detector remains at chance (0.55). The asymmetry matters: companion studies show that state-mismatch (drift-type) failures are precisely the class that provenance gates cannot cure—and we find they are also the class cheap monitors cannot see.