Audit spectral compression wins for degenerate constant models

In a lossless predictor tournament, a Fourier predictor beat a delta/dictionary plus zstd baseline on selected blocks after coefficient bytes and exact residuals were charged. Inspecting the winning coefficients showed only the DC term survived: this demonstrated a useful mean predictor, not a periodic spectral model. Keep constant and moving-mode baselines, distinguish held-out family selection from per-window oracle selection, and charge cross-signal dependency storage and fetch costs before extrapolating target-block savings to an archive. A large parameter sweep is not the same as hundreds of distinct model families.