Distributional Plasticity Loss in an Exactly Solved Markov Decision Process

Deep continual-learning systems can lose the ability to benefit from new information. Explanations commonly concern changing network features, dead units, rank loss, bootstrapping or early-data bias. We isolate a different mechanism in a tabular Markov decision process with no function approximator, training dynamics or estimated transition law. The model has 1,152 live states, two controls, hidden polarity and volatility regimes, and substochastic transitions representing death. Exact dynamic programming shows that an adaptive survival oracle and the best member of a fixed thirteen-rule family are nearly tied in aggregate outcome: the 1,200-step normalised survival-area difference is 3.1879 x 10^-4 and the expected-lifetime difference is 0.917 steps. Their survivor distributions are nevertheless far apart. After 600 preconditioning steps, conflicted evidence states contain 35.73% of oracle-conditioned mass but 1.05% under the fixed policy, a 33.99-fold ratio; total variation is 0.3595. The fixed-preconditioned distribution has 14.50% less continuation option value under the original regime and 5.71% less after forced polarity reversal. A 64-kernel survival-gap and admission sweep shows the conflict-mass ratio remains between 11.43 and 281.07 throughout the identified interior, while entropy-effective support ratios lie between 1.111 and 1.499. Admission probability zero is an exact structural null. No sharp interior phase transition appears. A 30-case flip-rate and horizon sweep also rejects the predicted monotone timescale crossover. The reversal damage itself is negligible, approximately 2.74 x 10^-5 AUC. The result is a decoupling example: an operational signature of plasticity loss can arise entirely from policy-shaped survivor occupancy, but large support displacement need not imply large robustness loss. This deposit contains the manuscript in PDF and DOCX form together with the complete evidence archive: the full sparse action kernels generated from the printed rules, exact policy iteration, the 64-case family map, the 30-case timescale sweep, row-level metrics, source code, an independent verifier and a SHA-256 manifest. All results are exact; no trajectory sampling is used.