Remote-sensing products provide near-real-time estimates of coral thermal stress, but they do not by themselves determine whether managers should monitor, act, defer, or redirect limited resources when reef condition is uncertain. We present ReefAction-POMDP as a POMDP-informed decision architecture that separates hazard prediction, ecological-state uncertainty, information acquisition, and constrained management allocation. The conceptual framework is a five-state dynamic POMDP; the empirical component reported here is deliberately narrower: a binary, one-step historical proof of concept using 3429 reef-date observations from 1724 reefs (1998-2017). A class-weighted random-forest observation model produced reef-blocked AUC = 0.785, average precision = 0.608, and Brier score = 0.166, whereas strict geographic blocking reduced AUC to 0.499, showing that pooled regional predictions did not transfer reliably to geographically distinct reef systems. Under normalized scenario assumptions, fixed-DHW, predicted-risk, risk-value, and EVSI-guided policies yielded 4.10%, 6.39%, 7.07%, and 7.18% simulated avoided loss at a 10% normalized budget. The 0.11-percentage-point difference between risk-value and EVSI policies is small and is not interpreted as evidence of practical superiority without uncertainty propagation. The numerical policy results are conditional mathematical comparisons, not estimates of avoided bleaching, coral survival, or intervention effectiveness. The principal contribution is therefore transparent decision logic and a reproducible research agenda for linking calibrated remote-sensing evidence to monitoring and allocation decisions, rather than a validated regional operational system.