release: 0.3.0

[RobbinBouwmeester]

Version bump for v0.3.0. main already carries the content (#74, #75, #76); this makes the tag gate accept the tag.

Why minor rather than patch

The DeepLC floor moves from 4.1.1 to 4.4.0, so every environment must be rebuilt before upgrading. mumdia doctor, interpreter discovery, sidecar::require_deeplc_version and both worker scripts enforce it from one constant, and every shipped environment pins deeplc==4.4.0.

What is in it

• Multi-head retention-time calibration, rt_im_train.multihead_calibration, off by default. deeplc.predict returns one of the base model's 6,543 LC-setup heads on that setup's own gradient, and the per-run LOESS cannot reorder peptides, so that setup's elution order survives into the result. Set the field to N and the worker fits MultiHeadRidgeCalibration over the N best-correlating heads against the run's own confident seed PSMs.
• sha2 0.11 in the desktop application, with the hex formatting it breaks.
• Nine desktop test fixtures made unique per process.

Measured, both HYE acquisitions

Four pooled six-file experiments from one binary, differing only in that field:

	AIF base	AIF multi-head	Astral base	Astral multi-head
stripped peptides at 1%	80,842	84,725 (+4.8%)	102,942	117,652 (+14.3%)
protein groups at 1%	10,707	10,963 (+2.4%)	11,458	12,276 (+7.1%)
empirical decoy fraction	0.0100	0.0100	0.0100	0.0100
PSMs into the pooled rescore	11,897,730	11,271,816	7,167,118	3,510,539
RT residual, in-sample median	67-78 s	19-26 s	15.8-16.0 s	2.0-2.2 s

Identical decoy fraction throughout, so no part of the gain is a loosened threshold. Costs 1.4x to 1.7x wall clock. Details in docs/08_rt_im_train.md section 4d.

It stays off by default in this release. CLAUDE.md requires an empirical null as well as two acquisitions before a sensitivity default changes, and an unchanged decoy fraction is not that test. Entrapment has not been run; flipping the default belongs in a later release once it has.

Changes here

0.3.0 in the workspace manifest, the desktop crate and tauri.conf.json; both lockfiles refreshed; CHANGELOG.md closed off with a fresh [Unreleased]; docs/14 and bench/pb_eval.py quote the new version; sbom.cdx.json regenerated (175 components, no dangling references).

Validation

Engine fmt, clippy -D warnings, cargo test --workspace (281); desktop fmt, clippy, cargo test --lib (83), and two concurrent runs to prove the fixture fix; pytest tests/python (68 passed, 12 skipped); compileall over scripts, bench, ci; gen_sbom.py --check; gen_third_party_licenses.py --check; gen_config_reference.py --check; check_doc_refs.py; check_desktop_ui.py; check_workflows.py.

After merging

With a green main run for the merge commit: Releases, Draft a new release, tag v0.3.0 on main, publish. Worth a manual pass through the rehearsal installer first, since the desktop start, stop and converter paths still have unit tests only.

🤖 Generated with Claude Code