Person re-identification (ReID) under severe occlusion remains a formidable challenge, as incomplete observations inevitably lead to fragmented features and corrupted distance metrics. While recent vision-language methods leverage text prompts to passively refine observed representations, they are prone to spatial drift under heavy occlusion and struggle to explicitly recover missing topological structures. Furthermore, existing feature recovery pipelines often rely on deterministic imputation, blindly fusing hallucinated features into the metric space without considering the inherent uncertainty of guessing. To break this bottleneck, we propose a unified framework that seamlessly integrates semantic probing with uncertainty-guided feature compensation. Specifically, we first introduce Text-Driven Proxy Pooling (TDPP) to actively retrieve reliable body-part evidence from corrupted observations using text semantics as spatial anchors. Subsequently, we propose Prompt-Guided Masked Feature Compensation (PMFC), which formulates occlusion recovery as a conditional variational masked modeling problem. By jointly optimizing reconstruction and KL divergence, PMFC explicitly encodes the recovery ambiguity into a posterior variance. During inference, this generative variance acts as a natural safety shield through our Uncertainty-guided Dynamic Retrieval (UDR) mechanism, strictly penalizing unreliable occluded features and preventing hallucinated noise from contaminating the metric space. Extensive experiments demonstrate that our framework achieves state-of-the-art (SOTA) or comparable SOTA performance across both occluded and holistic benchmarks (e.g., yielding 93.4% Rank-1 and 91.9% mAP on Occluded-ReID).