Wide packed structs: changed-field XOR columns beat byte-sparse XOR coding by ~35%; zstd prefers fixed-width columns

For 385/512-bit packed-struct waveform values (4-state, value plane + X/Z plane), compressed with zstd-9 per block, values only, exact round trips:

- plain XOR vs previous value + drop the X/Z plane when the block has no X/Z: baseline; - sparse gaps (positions of nonzero XOR bytes + those bytes): -7 to -9% on regular structs, but +5 to +7% on busy, irregular ones; - sparse bitmap: -3 to -14%; - changed-field (CF) coding (per value a bitmask of RTL fields that changed, then one fixed-width XOR column per field): -36 to -43%; - CF with gap-coded field entries: 13-23% WORSE than plain CF, because zstd compresses fixed-width columns better than variable-length entries. - On some 385-bit structs, a value dictionary (a reference to an earlier identical value) halved the size and beat CF by 15-18%, so pick per struct (adaptive) instead of a single codec.

CPU on a 2.1 GHz Xeon: - sparse gaps encode 0.13 / decode 0.025 us/value; - CF encode 0.54 / decode 0.09 us/value, with a word-level incremental decoder (u64 planes, per-field shift/mask).

Encoder pitfall: per-byte vector push_back into two candidate buffers (XOR and dictionary), plus 2 zstd calls per block, cost 4-5x an optimized word-wise XOR loop (1.0-1.2 vs 0.2-0.25 us/value).

Also, a ZSTD_CCtx created inside an encoder object and never freed leaks ~4 MB per object. With one encoder per signal over thousands of signals, RSS climbs to GBs and std::vector eventually throws length_error under ulimit -v. It only showed up at -O3 with many files.