v0.5.1

Performance

• fqxv compress --estimate is now 15–60× faster — always subsecond on a
  block-sized sample. It predicts the archive from the sample's measured entropy
  (what the real entropy coders converge to) instead of coding it: static order-k
  sequence entropy, a k-mer duplication sketch that captures long-read cross-read
  redundancy, order-1 quality entropy, and the real tokenizer for names — each
  scaled by a small per-platform calibration factor. The measurement passes run in
  parallel across read chunks, and the sample is capped by read count for short
  reads and by one block of bases for long reads (so the sketch sees a real block's
  coverage). Predicted archive size stays within ~1% of the previous coding-based
  estimate across Illumina, PacBio HiFi, and Nanopore.