Self-healing infrastructure is useful only when it acts before a service-impacting event and when the action itself does not create a larger disturbance. Existing telemetry anomaly detectors mostly answer whether a system is abnormal now ; conversely, many self-healing architectures leave the predictive trigger and remediation policy underspecified. This paper introduces ProHeal, a concrete predictive self-healing method that couples event-onset forecasting with a risk-controlled alert gate and a disruption-constrained remediation policy. First, ProHeal formulates failure forecasting as predicting the onset of a labeled operational anomaly within a future horizon while explicitly excluding already-abnormal states. A robust multi-resolution precursor encoder combines normalized state, absolute deviation, 5/15/30-minute mean shifts, 15-minute volatility, and sparse global deviation summaries. A LightGBM hazard learner then estimates multi-horizon onset risk. Second, a block-max conformal gate converts scores into alert episodes under an explicit exceedance budget, reducing sensitivity to isolated correlated score spikes. Third, a matchedstate drain/migration policy selects a healthy historical continuation subject to a boundary-disruption constraint; a PCA residual model trained only on normal training telemetry provides an independent health-risk evaluator for counterfactual replay. Experiments use two real, publicly downloadable industrial servermonitoring sources: eBay Pooled Server Metrics (PSM) and three representative Server Machine Dataset (SMD) entities. All splits are chronological. On PSM, ProHeal attains AUPRC 0.0127, 0.0684, and 0.0883 for 5-, 15-, and 30-minute onset forecasting, respectively, corresponding to 2.26×, 1.94×, and 2.09× the AUPRC of a strong LightGBM lag-feature baseline. At the primary 15-minute operating point (α = 0.03), the gate generates 0.784 false alert episodes per 1000 healthy minutes and detects 11.1% of test events with a median four-minute lead. Across three SMD entities, the method is strongest in AUPRC on one entity, second-tier on another, and weak on the sparsest one, documenting cross-entity heterogeneity rather than masking it. Ablation identifies windowed mean shifts, volatility, and global deviation summaries as the dominant predictive components. Under 2% Gaussian telemetry noise and 2% random missingness, 15-minute PSM AUPRC remains 0.0472 and 0.0509. Finally, in a policy-only replay over 18 PSM test events, constrained matched-state drain reduces independent mean health risk by 65.7% relative to no action while using 45.0% less boundary disruption than random drain; 88.9% of events have a candidate satisfying the primary disruption budget. Code, exact result tables, data-download instructions, and cryptographic hashes accompany the paper.