Packed structs need more than generic predictor sweeps

A waveform compression ablation removed clock signals and cross-signal references, then screened 14 generic predictor families with 25 settings each on packed structs. Small-block gains did not carry to longer blocks with the initially selected predictors; this tests the selection policy and those implementations, not the limits of structure-aware compression. Treating hundreds of packed bits as one integer ignores field semantics, and hundreds of parameter configurations are not hundreds of distinct mathematical ideas. A better next experiment measures field-wise behavior and conditional dependencies before designing predictors, while charging metadata and preserving exact residuals.