Skip to content

Release v0.67.4

Choose a tag to compare

@mobiusklein mobiusklein released this 10 Oct 03:52
· 11 commits to main since this release

[0.67.4] - 2026-10-10

Fixed

  • Fix isolation window boundaries when no offsets or limits provided

[0.67.3] - 2026-10-05

Added

  • Count remaining iterator items and prevent repeated peaks (#64)
    Subtract consumed items from iterator lengths, cap oversized indexed skips, and stop front/back peak reads when their positions meet. Keep chromatogram iteration exhausted after returning None.

Preserve full source counts for seeking and the existing peak and reader size hints. Add regression tests in the existing source files.

Fixed

  • Check binary array casts and unaligned reads (#60)
  • Preserve bytes when opening streams (#61)
  • Use requested detail levels in time lookups and async reads (#62)
  • Correct mzML checksums and return file read errors (#63)
  • Replace old spectrum data in read_into (#65)
  • Return errors for empty iterator time seeks (#67)
  • Restore MGF reader state after seeking (#66)
  • Size loader chunks for the current Rayon pool (#68)
  • Return imzML opening errors (#69)

Removed

  • Forward serde and compression features (#59)
  • fix: forward serde features in model crates

  • fix: preserve requested compression backends

  • chore: remove unused direct memchr dependency

  • chore: stop requesting quick-xml serialization

[0.67.2] - 2026-09-29

Added

  • Add phf to reduce string matching
  • Add AsyncChromatogramSource

Changed

  • Use highest oberserved ion mobility and equivalent instead of custom params in TDF

Fixed

  • Experimenting with async grouping stream
  • Provide vectorizable calibration helpers for TDF
  • Reduce overhead when calling BinaryArrayMap::sort_from_indices

[0.67.1] - 2026-09-18

Added

  • Add BinaryArrayMap3D::params_of and siblings to bridge the gap between the parameter lists of DataArray that are split across the ion mobility dimension.

Fixed

  • Allow TDF readers to opt in to export their calibration models as CV parameters. They will use provisional IDs until ready.