[RobbinBouwmeester]

Summary

Two related fixes to the second-pass trace pad. This changes default output and therefore needs an explicit decision against the 3D byte-identity rule, which is why it is a draft.

1. 3D second passes are padded for quantification (43a544b). run_experiment requested TracePad::for_quant only when the spectra carry 1/K0. On 3D data the narrow second-pass RT window then truncated the fixed integration window: on ProteoBench Astral 2-35% of condition A and 73-94% of condition B quantities were integrated over fewer than 7 samples. The pad now applies to every second pass with extract.emit_rt_window, as it already did for 4D.
2. Pad points stay outside the f32 window bound (16b7cb6). features::clip_rows cuts a padded trace back with chromatograms::window_span, which compares in f32, while extract decides in f64. A pad point just outside a bound whose RT rounds onto the f32 bound was read back as inside, so it leaked into a few feature rows (Astral 38 of 814,524). This leak already exists on main for 4D second passes (28-55 rows per run on diaPASEF). A pad point that would round inside is now stored one f32 step outside.

Measured

• 3D (Astral seed 0, AIF seeds 0 and 2): features, psms_competed, scored_combined and every per-run scored.parquet are byte-identical to main. Only second-pass chromatograms and quantities change. Clipped integrations go to 0. ProteoBench Astral species-weighted |eps| improves by 0.011-0.019 with better CV and human SD on every seed (removal of a uniform condition offset, not an accuracy gain in the narrow sense); AIF, Low Input and diaPASEF quantification are unchanged within noise.
• 4D (diaPASEF, one seed): the boundary fix removes main's leak, so second-pass features and identifications change by seed-level amounts (peptides -4, precursors -17, protein groups -17). Not yet measured with seeds and entrapment.

Before merging

• Decide on the 3D byte-identity break (3D second-pass quantities and chromatograms change on purpose; CI smoke passes, its expectations do not cover these columns).
• diaPASEF seeds and an entrapment run for the 4D identification change, or split 16b7cb6 out if 4D should wait.

🤖 Generated with Claude Code

[RobbinBouwmeester]

diaPASEF validation of the boundary fix (16b7cb6): base a836630 against 6bf15dc, NN seeds 0-2. The decision rule was written before the seed 1, seed 2 and entrapment numbers were looked at.

The fix removes the 4D leak on every seed: 23-55 second-pass feature rows per run differed on ProteoBench diaPASEF HYE (about 144,000 rows per run), 0-7 on the E. coli entrapment experiments. The first pass, the second-pass library, psms_extracted and run_windows stay byte-identical.

• HYE 3 + 3, seed means, branch minus base: peptides +2 (seed spread 337), precursors +10 (410), protein groups -9 (84), per-run precursors on run_psm_q +100 (331). PSM decoy fraction 0.0099 in every arm. Units: experiment-wide peptide_q_value, precursor_q, pg_q_value; per run on run_psm_q.
• Quantification (ProteoBench 0.18.6 offline, min_obs 3): ions +95 (spread 337), global median |eps| +0.0001, CV +0.0001. Species-weighted |eps| +0.0014 / +0.0017 / +0.0023 (base 0.2017, branch 0.2035, base spread 0.0015), from E. coli (+0.0050). This follows from the reshuffled second-pass identifications, not from changed quantities: of about 99,900 ions per seed quantified in the same runs in both arms, at most 4 change, by under 4e-5 relative. The branch accepts slightly more weak second-pass E. coli precursors in condition A. Three seeds do not decide whether this is systematic.
• Entrapment: the existing E. coli diaPASEF entrapment is a single run without a second pass, so it cannot exercise this change. Three E. coli diaPASEF replicates with a 1:1 human entrapment proteome were therefore run as one experiment with the default second pass. Measured precursor FDP at precursor_q 1%: 0.614% (base) against 0.613% (branch), seed means, base spread 0.013 pp; peptide level 0.758% against 0.759%; real precursors 18,056 against 18,046 (spread 106). A control with three copies of one run gives 0.330% against 0.338%. Limit: the leak is rare on these runs (0-7 rows per run), so entrapment tests the change at low exposure; a matched-FDP count is not informative because both arms stay below 1% over the whole list.

Conclusion: no change in 4D sensitivity or FDR beyond seed noise. The one consistent movement is a +0.0018 species-weighted |eps| on diaPASEF from the reshuffled second-pass identifications.

The branch is now merged with main (after #229, #230, #232, #238). The docs that #229 added described the trace pad as 4D-only; those passages and the quant.fixed_scan_halfwidth reference are updated. cargo fmt, clippy -D warnings and cargo test --workspace pass (0 failures). Marking ready for review; what remains is the decision to accept the deliberate change of 3D second-pass quantities.

🤖 Generated with Claude Code