What Normalization Erases: Absolute-Amplitude and Morphological Features as an Interpretable Impact-Detection Leg for End-of-Line Compressor Screening

End-of-line (EOL) vibration screening with deep anomaly detectors normalizes every recording before inference, which removes the absolute-amplitude scale that separates impact faults in reciprocating refrigerator compressors. The objective of this work was to design and evaluate a dual-path screening architecture in which an interpretable absolute-amplitude specialist ran beside a normalized deep generalist on a single rig, single sensor and single compressor population. Per-recording normalization lowered the maximum-peak AUC from 0.927 to 0.513; with features and classifier held fixed, a global gain changed nothing, whereas every per-recording scaling tested reduced single-peak separability to AUC 0.50–0.82. The specialist was a Random Forest on ten features—the three largest amplitude peaks and seven scale-free spectral-flux descriptors—with nested out-of-bag thresholds. Under leave-one-session-out validation it attained 78.2%/84.9% recall at a 0.5%/1.0% false-positive rate (FPR) (session-clustered bootstrap 95% CI 73.2–87.5%/81.0–93.3%), with a realized FPR of 1.2% at the 1% target (6/500; Wilson 95% CI 0.55–2.59%). On an unseen fault unit it detected 24/32 and 26/32 recordings at the production thresholds (Wilson 95% CI 57.9–86.7% and 64.7–91.1%), with no triggers on 20 abnormal-noise recordings. At equal FPR on an independent healthy session its recall did not differ significantly from that of the deployed deep CNN (26 versus 23 of 32 at 0.5%; exact McNemar p=0.45), but its thresholds transferred to a second healthy session with far less false-alarm inflation (1.97% versus 28.1% at a nominal 0.5%). These external results were preliminary (n=32 and n=20). Spectral kurtosis performed near chance (AUC=0.517). The specialist has been running in production since June 2026.