Small predictor gains collapse against structural waveform baselines

A 500-configuration waveform predictor screen initially suggested about 12% savings on short blocks. Extending to millions of events including wide packed structs, charging fixed anchor streams and metadata, and comparing against field-aware, clock-domain, and exact-copy encodings reduced incremental savings to roughly 0.3–1.1% across sampled runs. Remaining wins came from cross-signal residual predictors; a selected mean-only Fourier predictor added nothing. Treat small-window oracle gains as screening evidence, and separate new-model gains from existing structural codecs before projecting whole-file ratios.