0.2.0

Added

• Added CRAM output to nanalogue_sim_bam. The simulator now selects BAM or
  CRAM from the output extension and creates the corresponding BAI or CRAI;
  CRAM 3.1 output uses a generated external FASTA reference and FAI. The public
  write_cram_denovo API and detailed simulator CLI configuration help are
  also new.
• Added mm_suffix and drop simulation options. Simulated MM groups can use
  ?, ., or no suffix, and selected canonical bases can be omitted to
  generate non-zero MM distances and omit their corresponding ML entries.
• Added the public mm_groups/ParsedMmGroup MM-tag parser, resource-limit
  constants, and reusable input-validation helpers.
• Added reusable MM/ML parser and end-to-end CLI benchmarks with deterministic
  ModBAM fixtures, cache controls, baseline comparisons, and JSON results.

Changed

• Sped up modification parsing by scanning BAM's packed sequence directly,
  parsing MM distances without intermediate strings, flattening stored calls,
  caching candidate positions, and using a sparse CIGAR-segment coordinate map.
  End-to-end benchmarks improved representative sparse long-read commands by
  more than 2×, with gains varying by read and modification structure.
• Sped up dense, overlapping window-dens and window-grad workloads by
  reusing rolling reference-coordinate and base-quality calculations. Window
  basecall_qual values can differ by one Phred unit at rounding boundaries
  because of the changed floating-point summation order.
• Made Polars an optional, default-off dependency. Library consumers using
  curr_reads_to_dataframe, reads_table::run_df, window_reads::run_df, or
  Polars-specific error handling must explicitly enable the polars Cargo
  feature. CLI functionality is unchanged. Several Polars output columns,
  including sequence lengths, alignment bounds, and query positions, now use
  UInt32 instead of UInt64.
• Reduced the dependency surface by replacing thiserror, regex, csv, and
  itertools usage, and by incorporating only the required bedrs and
  openssl-probe functionality. Enabled libdeflate in rust-htslib.
• Hardened analysis of untrusted or pathological inputs with explicit limits
  for record capacity/count, CIGAR operations, sequence and contig lengths,
  modification types and annotations, MM/ML data, identifiers, paths, regions,
  sequencing-summary files, and peek records. This includes a maximum of 20
  modification groups per record and stricter ASCII/character rules for read
  IDs, contigs, and paths; displayed user-controlled error context is truncated
  after 300 Unicode characters. find-modified-reads, table/window commands,
  and peek now error on zero input records, including when applicable filters
  remove every record. read-info continues to return an empty JSON array, and
  read-stats returns zero-valued summaries. TSV/DataFrame window output also
  errors when records are present but no windows are emitted.
• Applied the HTTP, HTTPS, and FTP URL-scheme allow-list consistently to public
  URL-reader helpers and deserialization. Hardened the install instructions and
  script to require HTTPS and TLS 1.2 when using curl (and HTTPS when using
  wget).
• Sequencing-summary parsing now uses bounded plain TSV input rather than CSV
  quoting rules. Comments are accepted only before the header, and supplied
  summaries must contain a header and at least one read row.
• Simulator output paths must end case-insensitively in .bam or .cram, and
  simulated unmapped reads now use MAPQ 0 rather than 255.
• Breaking for library users: many coordinates and lengths now use bounded
  u32 types in InputBam, InputWindowing, GenomicRegion, simulation
  configuration and helpers, BamPreFilt, OrdPair, alignments, modification
  annotations, and relevant error payloads. Conversion from &GenomicRegion
  to bam::FetchDefinition is now infallible (From rather than TryFrom).
  The crate now exposes its own Bed3/StrandedBed3 types and GenomicBed3
  aliases; interval constructors are fallible and reject reversed coordinates.
  GenomicRegion and AlignmentInfo no longer implement Default. Only
  32-bit and 64-bit targets are supported.
• Breaking for library users: FiberAnnotation is now an 8-byte packed value
  representing single query/reference positions rather than intervals; its old
  end, length, and extra-column fields were removed. Construct it with
  FiberAnnotation::try_new(pos, qual, ref_pos) and use pos(), qual(), and
  ref_pos() accessors. FiberAnnotations fields are also private and its
  coordinate/quality accessors return iterators (starts() and
  reference_starts() become pos() and ref_pos()); from_annotations now
  returns Result. The maximum supported read length is now 2^24 - 1 bases
  (about 16.7 megabases), and the maximum contig length is u32::MAX - 1.
• Breaking for library users: callbacks passed to window_reads::run,
  run_json, and run_df now return Result<Option<_>, Error>: Some emits
  a window, None omits it, and errors are propagated instead of logged and
  skipped. Accordingly,
  threshold_and_mean_and_thres_win now returns Ok(None) rather than an
  error for a window below the requested density, and the obsolete public
  Error::WindowDensBelowThres variant was removed.
• Breaking for library users: TempBamSimulation::new now takes an
  AlignmentFormat, and TempBamSimulation no longer implements
  Deserialize. SeqCoordCalls::mod_calls now returns &[u8] rather than
  Option<&[u8]> and panics for an out-of-range position.
• Breaking for library users: external errors wrapped by Error now use boxed
  payloads, affecting direct construction and nested pattern matching;
  Error::CsvError was removed.
• (Project tooling, not code) Expanded locked and unlocked dependency CI to
  macOS and both default/all-feature builds, strengthened repository guardrails,
  and updated the pinned Zig release toolchain with download verification and
  a libdeflate compatibility workaround.

Removed

• Removed the public get_u8_tag and
  utils::filter_by_ref_coords::WindowState APIs. The simulation helpers
  generate_random_dna_modification and generate_reads_denovo are now
  private.

Fixed

• Corrected read-stats median and N50 boundary calculations, including exact
  halfway totals and zero-length values.
• Made MM/ML parsing reject malformed records instead of panicking or silently
  accepting them. This includes non-canonical numeric modification codes,
  oversized headers, gaps above 128,000,000 bases, zero-length sequences,
  mismatched MM/ML counts, MN tags that disagree with sequence length, and
  mapped reads without a CIGAR. Canonical U is handled as BAM-encoded T,
  and MM/ML parsing handles padding (P) CIGAR operations without panicking.
• Enforced sorted query coordinates and strictly monotonic mapped reference
  coordinates in modification annotations, including during deserialization,
  so reference filtering and window bounds cannot silently use invalid order.
• Applied complete validation consistently to parsed, converted, built, and
  deserialized genomic regions and intervals. Alignment intervals must be
  non-empty; BED intervals may be empty but cannot be reversed.
• Fixed probability-exclusion bands to reject exactly the representable
  byte-valued ML probabilities within the inclusive bounds; a band containing
  no representable probability now excludes nothing.
• Deferred each read-info JSON delimiter until its record has been parsed and
  rendered successfully, avoiding a dangling object for a failing first
  record. Moved missing-index warnings from stdout to stderr so they no longer
  corrupt structured output.
• Gradient commands now reject one-element windows, and invalid
  sequence-display flag combinations are rejected by the CLI.
• Fixed simulator validation and cleanup: empty window/modification schedules
  are rejected, failed temporary simulations remove their directories, and
  aliased or hard-linked output paths cannot overwrite one another.
• BAM/CRAM writers now reject unusable index paths and out-of-range CRAM thread
  counts before creating alignment output.
• Explicitly flushed subcommand output and propagated flush failures, preventing
  buffered-output errors from being silently lost.
• Read-ID filter files now permit comments only before the first ID and reject
  blank/whitespace-only lines and malformed IDs.
• Fixed HTTPS certificate discovery on macOS by checking the standard macOS
  certificate bundle path.
• (Project tooling, not code) Fixed install.sh to use the ARM release archive
  name produced by the release workflow.
• (GitHub workflow, not code) Docker publication dry runs now authenticate with
  Docker Hub so invalid release credentials fail before a real push.
• Updated the locked networking dependencies to h2 0.4.16, rustls 0.23.45,
  and rustls-webpki 0.103.15 to address security advisories.
• Updated locked event-listener from 5.4.1 to 5.4.2 to address
  RUSTSEC-2026-0221, and replaced the yanked chacha20 0.10.0 with 0.10.2.