Before more codec tuning on a compressed format, measure the headroom with strong compressors and rank work by read cost

For a random-access, block-compressed data format (waveforms, time series, columnar), it's tempting to keep tuning transforms and codecs. A cheaper order of work:

1. Fix the read path first. A point query that decodes a whole block to return one value wastes most of its work. Stopping the decode at the query position gave a ~3.4x faster multi-signal point read in a prototype, more than any codec change was likely to give. 2. Measure headroom before codec work. Compress representative transformed block payloads with zstd --ultra -22 --long, xz -9e, brotli -q 11 and a context-mixing compressor (zpaq/paq). Treat the best as an empirical bound (not a true entropy bound). If your current size is within ~10-15% of it, stop tuning the codec. Order-0 entropy is far too weak a bound to decide this. 3. Look for structure, not entropy: exact duplicate streams, constant-offset copies, periodic signals (store a descriptor plus exceptions, answer queries in O(1)) and shared trained dictionaries are often worth more than a better entropy coder, and they also speed up reads. 4. Choose block size per stream with a cost model: size + weight x expected query decode cost, from the stream's change rate and width. Fit the per-operation costs from a benchmark rather than guessing.

Stratify every measurement by value width. Wide values usually dominate bytes even when they are a small fraction of changes.