Rare co-occurrence patterns can be difficult to discover across institutions whose transactions remain distributed. We present FedRPM, a bounded-candidate protocol combining noisy summaries, pooled support estimation, and a verification stage. The reported proof-of-concept campaign uses twenty eight-client federations of UCI Adult under two Dirichlet skew levels, ten seeds per level, six planted carrier-concentrated itemsets per run, and a centralized Apriori support-band reference. FedRPM achieves mean recall 0.9564 and precision 0.9999 for supports between 0.5 and 2.0% at 67.22 KB of serialized communication, compared with recall 0.9234, precision 0.9896, and 72.49 KB for naive broadcast (reported paired Wilcoxon W = 6, p = 2.67e-05). The full-lattice federated reference reaches recall 0.9926 at 82.59 KB, or 22.9% more communication than FedRPM. Reducing the nominal per-coordinate epsilon from 8 to 0.5 lowers recall to 0.9221. These epsilon settings describe count-noise scales, not a demonstrated total record-level privacy budget. Secure aggregation is simulated, transmitted-content bytes are an exposure surrogate rather than a privacy guarantee, and adaptive thresholds and the feasibility controller are untested. The results therefore characterize candidate coverage and communication in a small simulation; an audited transcript accountant and the experiment artifacts are required to establish end-to-end privacy and reproducibility.