Federated learning (FL) deployments in the wild often suffer from missing-not-at-random (MNAR) participation, where client availability depends on power, connectivity, and workload conditions. Such biased activation distorts global model calibration and fairness across heterogeneous devices. We introduce CoPA–Fed, a server-side reliability auditing framework that frames participation bias and coverage control as queryable reliability signals aligned with data-management principles. CoPA–Fed continuously estimates client participation propensities from lightweight telemetry (availability, latency, gradient, and loss signals) and instantiates classical inverse propensity weighting (IPW) within federated conformal calibration to approximately maintain target coverage under MNAR participation. The framework produces two actionable diagnostics: a \(\varGamma \) -sensitivity ladder quantifying robustness to adversarial up-weighting and participation-fairness deciles measuring coverage across rarely to frequently active clients. CoPA–Fed is optimizer agnostic, plug-and-play, and imposes negligible overhead; clients only share small validation statistics. Experiments on HAR, EMNIST (Balanced), and PathMNIST under non-IID and MNAR conditions show that CoPA–Fed achieves near-target coverage ( \(\tau {\approx }0.90\) ), improves worst-decile coverage relative to baselines, approaching target levels, and yields robustness certificates around ( \(\varGamma ^{\star }{\approx }3\) ).