v0.3.0

Added

• Sequence-conditioned, context-mixed quality for long reads — the quality
  coder now conditions each score on the read's bases and recent qualities instead
  of read position (which carries no signal on long reads). It mixes several
  context models of increasing richness (coarse/mid/rich) with adaptive,
  confidence-gated weights — a logistic mixer — because a per-block adaptive model
  can't exploit a richer single context but can blend a well-trained coarse one
  with a sparse rich one. On PacBio HiFi this takes the quality stream (the dominant
  share of a HiFi archive) below CoLoRd, lossless. The mode is chosen automatically
  by mean read length and recorded in a self-describing header byte, so short-read
  archives are byte-identical to before; long-read archives decode the sequence
  first and feed it to the quality decoder. The mixer is fixed-point/integer
  throughout, so archives are bit-identical across platforms. New fqzcomp API:
  encode_seq/decode_seq/needs_sequence; new random-access projection helpers
  decode_quality_with_seq/quality_needs_sequence.

• Long-read (ONT / PacBio) support — platform-aware compression for
  Nanopore and PacBio reads: --platform illumina|nanopore|pacbio (with header
  auto-detection), long-read quality binning (--quality-bin ont|hifi, matching
  CoLoRd's cutpoints), and a dedicated long-read overlap sequence codec that
  drives its minimizer sketch from the detected platform.

• Long-read overlap sequence codec (fqxv-lroverlap) — a new cross-read
  overlap codec (minimizers → overlaps → layout → consensus → per-read banded
  edit script → rANS) wired into the container as the sequence path for long-read
  blocks, auto-selected and kept only when it beats order-k. It reaches CoLoRd
  parity on ONT/HiFi (e.g. ~0.653 → ~0.067 bits/base at depth) and codes its edit
  streams (substitutions, ops, insertions) with per-stream context models rather
  than a flat order-0 code — worth a further −4.3% on the ONT sequence stream,
  losslessly. A WFA aligner (HiFi) and an AVX2 anti-diagonal aligner accelerate
  encoding, both byte-identical to the scalar reference.

• Closed-syncmer seeding for ONT — the overlap codec seeded from window
  minimizers, which a base error in any k-mer of the window can deselect; at
  Nanopore's ~10% error that is the dominant loss of shared anchors. ONT now seeds
  with closed syncmers (a k-mer is an anchor when its minimal s-mer sits at its
  first or last position), so selection depends only on the k-mer's own bases and
  an intact shared k-mer is co-selected regardless of neighbouring errors. Density
  is 2/(w+1) as before, so k, anchor count and specificity are unchanged — only
  conservation improves: ONT overlap-coded sequence goes 1.631 → 1.559 bits/base
  against an order-k baseline of 1.808, widening with per-block coverage. Seeding
  is encode-only, so archives stay decodable by any reader. PacBio keeps window
  minimizers, which are already near-optimal below ~1% error.

• Content-based platform detection — platform was detected from read-name
  grammar alone, so SRA-reformatted runs (bare SRR… headers) recorded unknown
  and were handed the Nanopore sketch. When names carry no platform signal, fqxv
  now classifies long-read runs by mean per-base quality, which separates the
  platforms cleanly (measured across 24 corpus accessions with ENA ground truth:
  Nanopore 6.9–23.5, PacBio 36.9–84.5). Ambiguous data stays unknown, so a wrong
  platform is still never recorded. On a real HiFi run this restores the
  low-divergence WFA path and cuts encode time ~31%.

• Shared whole-file reference for long reads — the overlap codec reaches
  CoLoRd parity within a block, but the container re-assembled and re-stored the
  same consensus reference in every 256 MiB block. It now assembles one consensus
  over the whole file and stores it once in a framed region between the header
  and the first block, coding every block's reads against that frozen frame — so
  the genome is stored once, not once per block. Auto-selected for long-read input,
  behind a whole-file never-worse gate against order-k, and gated on the
  GLOBAL_REFERENCE feature bit (which is set in the plain layout now as well as
  the reorder layout, so a reader without it refuses rather than misreads). On
  PacBio HiFi this drops the sequence stream ~0.102 → ~0.084 bits/base at two
  blocks, widening with block count. New fqxv-lroverlap API: Reference,
  build_reference, encode_against/decode_against.

• Python bindings (fqxv via PyO3 / maturin) — read-only access from
  Python: a streaming record iterator and column projection (fetch just names,
  sequence, or quality) over an existing archive.

• compress --verify — an opt-in read-after-write check that round-trips the
  freshly written archive back to the original records before exiting, backed by a
  public verify_roundtrip in the library.

• Block sync markers + footer-independent recovery — each block carries a
  sync marker so a reader can resynchronize and recover blocks even when the
  footer index is missing or truncated.

• compress -f/--force — compression now refuses to overwrite an existing
  output unless --force is given.

• Remote / parallel column projection — the footer row-group index now
  records, per group, a (offset, len, crc32c) triple for each of the three
  coded streams (names, sequence, quality). A client can fetch the archive tail,
  parse the index, and issue a single range request for just one stream — read
  names are <1% of the archive, so an ID-only client fetches ~100× less than
  before, and a sequence-only client (k-mer screening, classification) skips the
  quality stream. The joint block content digest can't verify a single fetched
  stream, so each stream carries its own CRC-32C.

• Random-access API — a public, IO-free Index (parse from a seekable
  reader or a fetched suffix buffer via Index::from_suffix), Index::byte_ranges
  to turn (groups, stream) into byte ranges to GET, Index::verify_stream,
  and per-stream (decode_names/decode_sequence/decode_quality) and
  whole-block (decode_block_contents) decoders. The caller drives fetching
  (local File, object_store, async HTTP, …).

• compress --block-reads N — set the reads-per-row-group directly,
  decoupling random-access granularity from the --level effort knob. Smaller
  groups give finer remote access and more parallelism at some ratio cost.

• Per-stream content digests — the plain block payload now carries three
  xxh3-64 digests (names, sequence, quality) in place of the single joint digest.
  A post-decode mismatch (a codec round-tripping CRC-valid bytes into
  wrong-but-in-bounds output) now names the offending stream instead of only the
  block. Each digest still folds in n_reads and its stream's per-read lengths,
  so boundary pinning is unchanged; cost is 16 extra bytes per block.

• Crash-safe compress output — compression writes to a sibling temp file and
  atomically renames it into place only once the whole archive (header, blocks,
  and footer trailer) is on disk. Interrupting a stream mid-run (Ctrl-C on
  sracha get -Z | fqxv compress -, say) or hitting any error no longer leaves a
  corrupt, footer-less .fqxv at the destination: the partial temp is removed on
  a ? bail and by a SIGINT/SIGTERM/SIGHUP handler, and the destination path only
  ever holds a complete archive.

• Live compress progress — the compress indicator now reports how much data
  has been processed and the rate. With a known input size (a file) it renders a
  percentage bar; for a stdin stream of unknown length it shows a bytes + rate
  readout, so a pause waiting on an upstream producer reads as 0 B rather than a
  hang.

Changed

• The on-disk format is stable at 1.0. Versioning moved from a single
  monotonic FORMAT_VERSION integer to FORMAT_MAJOR.FORMAT_MINOR: a reader
  refuses a differing major and tolerates a newer minor, and additive features are
  gated behind required-feature bits so a reader that predates a feature refuses
  the archive outright rather than misreading it. That contract is now a stability
  guarantee — archives written by a 1.x release stay readable by later ones, and a
  major bump would be announced as a breaking change. The 1.0 format carries the
  long-read overlap codec, the extended per-stream footer index, and the
  per-stream block digests below.
• platform has its own header byte — the platform tag no longer shares bits
  with the flags byte (a prior collision made Illumina --order any archives
  undecodable).
• inspect now sums per-stream sizes straight from the footer index instead
  of seeking to each block header — one footer read is the whole metadata cost.
• fqxv-dna primitives crate — the 2-bit ACGT lookup and reverse-complement
  helpers are extracted into a shared leaf crate, and the fqxv-reorder monolith
  is split into focused modules.
• Default log verbosity is warn — routine per-run info diagnostics now
  require -v, so they no longer interleave with and smear the live compress
  indicator on a shared stderr (-vv = debug, -vvv = trace with targets).
• --verify verifies before publishing — the read-after-write check now runs
  against the temp file, so a failed verification leaves the unverified output off
  the destination path (kept aside for inspection) and exits non-zero, rather than
  leaving a suspect archive in place.
• --help layout — the flags most runs need are listed first and the
  rarely-touched knobs are grouped under an Advanced heading, so the default
  help is short enough to read. -h prints a one-line summary per flag and
  --help expands the detail (when a knob changes the data or the guarantees,
  the long form says so).
• Rust edition 2024 — the workspace moved from edition 2021. The MSRV is
  unchanged at 1.95, so this is invisible to downstream builds on a supported
  toolchain.

Performance

• Single-end compress now streams through a block pipeline instead of
  buffering the whole input.
• Reorder read storage flattens cl_reads into an arena, cutting ~168 MB of
  peak memory with byte-identical output.
• Quality models are sized to the alphabet actually present in the file
  (~17% faster quality coding, ~45 MB less memory).

Fixed

• Silent quality corruption on certain inputs and a --block-reads
  compression-bomb path, both surfaced by CLI stress testing.
• fqzcomp no longer rejects legitimately compressible quality (which could
  drop data).
• Read-name headers are preserved byte-exactly.
• Long-read memory — the overlap codec's overlap/layout/placement and the
  banded aligner's DP matrix are now bounded, fixing OOMs on high-error ONT and
  large amplicon inputs (with O(n) placement).
• Assorted CLI fixes: --order any no longer aborts on a truncated FASTQ that
  preserve rejects, --estimate handles empty input, --interleaved 0 is
  rejected, and --verify honors --threads.
• Python binding tests now build their own tiny .fqxv fixture with the CLI
  instead of latching onto whatever archive sits at the repo root (the suite ran
  for minutes against a large sample and could fail on a half-written one); they
  now finish in well under a second. The Info repr also shows the container
  version as format=major.minor instead of the raw packed integer.

Security

• Decode-path hardening — a fuzzing-driven pass (cargo-fuzz targets over the
  public decode entries, plus a weekly fuzz schedule) closed multiple
  decompression bombs across the rANS, quality, sequence, name, reorder, and
  container decoders: untrusted length/size headers can no longer trigger
  unbounded allocation.

Documentation

• Long-read benchmark tables, a format-comparison page, a Python API reference,
  and --quality-bin ont|hifi documentation.