release: 0.7.0

[RobbinBouwmeester]

Release 0.7.0. A minor release: it reads a new kind of data and changes what a search finds.

What 0.7.0 brings (CHANGELOG ## [0.7.0]):

• diaPASEF: Bruker .d is read natively with ion mobility through the library, extraction, features, a raw-event retrace and cross-run quantification (refactor(sidecar): one version-floor check for discovery, doctor and launch #185-docs, bench: diaPASEF recipe, completed guide and measurements (diaPASEF 7/7) #194, docs/35). On diaPASEF input a second extraction gate (feat(extract): the diaPASEF gate on by default on spectra that carry 1/K0 #211), an MS1 noise filter and search anchors are on by default, and the speed and memory stack is in (Land the diaPASEF speed/memory stack (#196, #198-#206) on main #209).
• Match-between-runs: run-experiment matches between runs with a second pass over an empirical library of the pooled first pass (feat(mbr): second-pass match-between-runs, the default #184).
• RT calibration: rt_im_train.multihead_anchors = "auto" refits the multi-head calibration on search anchors when the seed found few (feat(rt): search-anchored multi-head calibration (opt-in) #182, feat(rt): multihead_anchors = auto, the default #183).
• Rescore prefilter: rescore.prefilter_keep_decoys = 0.35 (feat(rescore): an out-of-fold prefilter before the classifier (rescore.prefilter_keep_decoys, default 0.35) #210).

Release commit

• The version bump in the same 8 files as 0.6.0: the workspace Cargo.toml and Cargo.lock, the desktop Cargo.toml, tauri.conf.json and Cargo.lock, docs/14, bench/pb_eval.py and sbom.cdx.json.
• CHANGELOG: Unreleased becomes ## [0.7.0] - 2026-10-18, with a preamble on why it is minor. The two diaPASEF gate entries are merged, and entries that still said opt-in are reconciled with the defaults.

Still to come on this PR

• Validation against the v0.6.0 tag, each version on its own defaults (hippogriff):
  • the six-file Astral HYE experiment (timing);
  • ProteoBench Astral entrapment, Orbitrap AIF E. coli entrapment and single-cell HeLa species entrapment, three NN seeds per arm;
  • 22 HeLa runs and one single cell.
• Then a second commit with the measured numbers in the CHANGELOG preamble.
• Dry runs: release.yml and docker.yml dispatched on this branch; neither publishes anything.

Tag only after main's CI on the merge commit has finished: the release gate requires a successful ci.yml run for the tagged SHA.

🤖 Generated with Claude Code

[RobbinBouwmeester]

Release check: v0.6.0 against the 0.7.0 candidate

Setup:

• Engines: the v0.6.0 tag against main at efdbfbc (the engine of this PR before fb9c060).
• Configs: each version on its own defaults, with one config holding paths, nn_torch and q_filter.
• Environment: one DeepLC 4.5 environment; hippogriff, timing arms alone on the host.

benchmark	0.6.0	0.7.0 candidate	FDR
six-file Astral HYE (run-experiment, NN seed 0)	133,135 peptides experiment-wide; 696,882 per-run precursors; 10:31, 23.3 GB	129,796 experiment-wide (-2.5%); 709,354 per-run precursors (+1.8%), per-run peptides +2.8%; 9:42, 22.9 GB
ProteoBench Astral entrapment (3 runs, 3 NN seeds)	91,778 precursors experiment-wide; 240,600 per run	90,472 (-1.4%); 249,097 per run (+3.5%)	paired FDP 0.86/0.90/0.89% vs 0.85/0.85/0.91%
Orbitrap AIF E. coli entrapment (3 NN seeds)	11,850 real precursors at FDP 1%	11,747 (-0.9%)	FDP at q 1% 1.06-1.09% vs 1.03-1.11%
single-cell HeLa B8 species entrapment (3 NN seeds)	15,787 real at FDP 1%	16,333 (+3.5%)	FDP at q 1% 0.78-0.96% vs 0.89-1.05%
22 single-cell HeLa runs (run-experiment)	355,171 human precursors per run; 24,128 peptides; 15:27, 44.2 GB	407,341 (+14.7%); 24,753 peptides; 12:55, 14.7 GB	species FDP 1.52% -> 1.00% (precursor level 1.69% -> 1.18%)
one single cell (Cell_A C10), cold / warm cache	5,835 peptides; 4:11 / 1:04	6,376 (+9.3%); 5:09 / 2:15

Reading:

• Experiment-wide counts: they fall on the Astral pools because second-pass MBR is now the default. The second pass searches only what the pooled first pass accepted, so experiment-wide counts fall by 1.4-2.5% while per-run counts rise. mbr.strategy = "none" restores 0.6.0. The candidate's own first pass on the six-file experiment was 132,515 peptides (-0.5%, one seed).
• Single cells: the cell and HeLa gains come from search anchors (multihead_anchors = auto, below 2,500 seed anchors). The extra minute per cell is that refit; the projection cache keeps it to about 70 s.
• AIF entrapment (-0.9%): this was the prefilter, so fb9c060 now limits it to pools of at least 1 M PSMs.
  • Evidence: the 0.6.0 and candidate competed tables rescore bit-identically, and on them the prefilter gave 11,644 against 11,798 real precursors (-1.3%, FDP unchanged), saving 2 s of a 10 s rescore.
  • Effect: with the guard, single runs of that size are scored as in 0.6.0.
• diaPASEF: 0.6.0 cannot read .d natively, so there is no baseline; see docs/35 section 16.

Dry runs on this branch: release.yml (4 engine archives, MSI, Linux desktop) and docker.yml are green.