Dead, weak, saturated, locally bursting, and intermittently persistent traces in land shot gathers may become narrow vertical artefacts after sorting and stacking. Their joint detection and reconstruction is difficult because the corrupted support is unknown, and valid high-energy or high-curvature events can resemble anomalies. We formulate restoration as a decision problem involving uncertainty in both anomaly support and predictor reliability: multi-scale energy, persistence, spatial compactness, neighbour mismatch, receiver recurrence, and acquisition context first define a classified trace-time support; independently validated predictors then estimate only that support, with abstention when evidence or reference coverage is inadequate. Bilateral frequency–space (f–x) prediction, one-sided spatial autoregression, and geometry-aligned adjacent shots are selected by withheld-healthy-trace validation rather than unconditional fixed weighting. A controlled benchmark injected 192 known anomalous segments into 48 spatially separated field records and reserved 16 records (64 anomalies) as a held-out test set that was not used for parameter selection. On that test set, full evidence fusion increased detection F1 from 0.737 for energy-only screening to 0.843, with 0.780 precision and 0.917 recall. With known supports, the calibrated hierarchy achieved a median normalised root-mean-square error (NRMS) of 0.936 and a median signal-to-noise ratio (SNR) improvement of 14.56 dB. It significantly improved NRMS, SNR, and relative amplitude-spectrum error relative to the previous workflow and singular value decomposition reconstruction, while being statistically comparable to linear interpolation and traditional f–x prediction. Field-record, common midpoint, and stacked-section examples provide complementary qualitative evidence that narrow trace-related artefacts are reduced without broad alteration of reflection geometry. The contribution is therefore a reproducible, evidence-and-validation framework for support-local anomalous-trace restoration, rather than a new universal interpolator or a gather-wide noise-attenuation operator.