v1.15.0

Fixed

• TUI zebra stripe readable on light terminals. The default theme's
  alternating-row background is a near-black grey that only reads as a subtle
  stripe on a dark terminal; on a light-background terminal (e.g. JupyterLab's
  web terminal) it became a hard black band that swallowed the default-foreground
  text. vv now detects the terminal background (OSC 11 query, then COLORFGBG,
  overridable with VV_BACKGROUND=light|dark) and, on a light terminal, uses a
  subtle light-grey stripe instead. Dark terminals are unchanged; detection
  failure falls back to the previous behaviour.

Added

• --arrow / --feather output — write an Arrow IPC file (Feather v2), the
  zero-copy interchange format for pandas / polars / R arrow. Same column
  projection / --filter / - (stdout) handling as --parquet; streams
  chunk-by-chunk. --compression accepts zstd (default), lz4, or none
  (Arrow IPC body compression). Round-trips through vv and is read by pyarrow.
• --count — print the row count and exit. Instant on formats that carry a
  count in metadata (Parquet, LociSSD); reflects a -r region and, with
  --filter, counts only matching rows. A wc -l for any supported format.
• --describe --json / --describe --ndjson — machine-readable per-column
  statistics (the same count/nulls/min/max/mean/distinct as the text table) as a
  JSON array or one object per line, for reproducible QC in pipelines. Integers
  are exact (not %.6g-rounded), floats round-trippable, strings JSON-escaped;
  numeric columns omit distinct, all-null columns report null min/max/mean.
• Reference-aware pileup — vv x.bam --pileup -f ref.fa. With a reference
  FASTA (needs a .fai), the --pileup output fills the ref column from the
  reference and renders read bases matching it as . (forward) / , (reverse),
  byte-for-byte like samtools mpileup -f (vv applies no BAQ, i.e. -B); deleted
  bases are filled from the reference too. Without -f the previous behaviour is
  unchanged (ref N, literal bases). The current contig is fetched once and
  cached. Verified byte-identical against samtools mpileup -B -f (matches,
  mismatches, deletions, ref-column case). -f without --pileup is a clean
  usage error.

Changed

• LociSSD v4 reads faster. The colblock read_chunk decoded the Start
  column twice whenever End was also present (once for the LENGTH/mask input,
  once as an output column) — the universal case for these interval files. Start
  is now decoded once and the array reused. ~40% less time on both a full
  --tsv scan and a -r region query over a 1 M-row file; output byte-identical.

Fixed

• LociSSD v4 reader hardening (memory safety). The colblock decoders trusted
  on-disk offsets from the (untrusted) file: the DICT and ARENA string codecs did
  assign(blob + o0, o1 - o0) with no o0 ≤ o1 ≤ blob_len check (a non-monotone
  offset underflowed the length to ~4 GiB → out-of-bounds read), n_dict + 1 was
  computed in 32-bit (wrap at UINT32_MAX), a -r region query on a file lacking
  Start/End indexed columns out of range, and zstd_inflate had no output ceiling
  (a tiny chunk could inflate to gigabytes). All now validate and reject with a
  clean error (non-zero exit), never an OOB read or OOM. Crafted-input fixtures
  (tiny.v4.baddict/badarena/nocoord/zbomb.lociss) guard each path.