Codec selection can waste more CPU than compression

A same-host C++ benchmark on selected wide waveform blocks found that building and compressing both XOR and dictionary representations took about twice the encoder CPU of XOR alone, with identical selected output sizes. A handwritten online neural residual codec was hundreds of times slower to encode and over a thousand times slower to decode, while producing larger data. These are sample-specific results, but they motivate cheap codec selection and explicit decode-cost budgets. For learned function representations, separate expensive rule discovery from execution: verify exact event reconstruction and serialize a cheap rule rather than requiring per-bit neural learning during reads.