review D: LOESS boundaries, experiment provenance, audit q unit, transfer basis in TSVs (docs/29 findings 10, 15, 16, 19)

[RobbinBouwmeester]

Last of the four stacked code-review PRs (docs/29_code_review_2026-09-07.md, work package D). Base is review/c-desktop (#64); merge A, B, C, then this.

What changes

#10, LOESS boundaries. Loess::predict continues the boundary local fit (its value at the boundary grid point and the slope of the local line there) outside the anchor range instead of switching to the global least-squares line. Continuous at both ends; the global line is now only the degenerate fallback (fewer than four anchors, or a local window without spread). Test: the review's reproduction (y = 200 + 10x^2, span 0.3), endpoint continuity within 1e-3, extrapolation follows the local slope, linear far from the data.

HYE B01 before/after (merge check for #10). doxy, LFQ_Orbitrap_AIF_Condition_B_Sample_Alpha_01, identical DIA-NN library, native_tda, 32 threads per arm, C-branch binary (#64 head) against this branch; two RT sources. Unit: peptides.tsv rows selected on the stripped-peptide peptide_q_value at 1% (one row per stripped sequence here); PSM counts on the pooled q_value.

	imported DIA-NN iRT, before	after	DeepLC 4.1.1 re-predicted iRT, before	after
w_rt (identical cal.json per pair)	691 s	691 s	414 s	414 s
stripped peptides at peptide_q_value 1%	45,946	45,957	48,533	48,533
PSM-q 1% targets	50,332	50,335	53,124	53,127
PSM-q 1% decoy fraction	0.010	0.010	0.010	0.010
protein groups at pg_q_value 1%	6,410	6,410	6,519	6,519
extract accepted rows	2,183,445	2,183,899	1,961,773	1,961,800
candidates with a different window		208,130 (1.9%)		22,779 (0.2%)
wall (32 threads)	34:14	37:30	26:06	25:46

Neutral on both RT sources, as a boundary correction should be. The mechanism is visible in the first pair: 194,698 of the moved candidates have iRT above the anchor range (170); the old global line placed them at 10,625 to 11,946 s in a 9,000 s run, so they could never be extracted, and the continued local fit places them at 8,578 to 9,100 s, from which 454 produced an accepted extraction row.

#15, experiment manifest. experiment_manifest.json adds config_json (the resolved configuration, next to its hash), model_identities (RT source, fragment predictor, classifier that actually ran from the rescore report, feature schema id, MBR strategy), experiment.quant_q_filter.{configured,effective}, experiment.rescorer, and hashes the inputs at the start of the run (inputs_hashed_at: start). Relocatable references and the shared/per-run artifact inventory stay deferred, as agreed.

#16, audit. passed_precursor_fdr / FAILED_PRECURSOR_FDR read precursor_q; PSM q_value only on a table without it, with q_unit in the metrics JSON. A scored table with more than one source is refused. NO_PEAK_GROUP renamed to DID_NOT_SURVIVE_EXTRACTION (RejectionReason::DidNotSurviveExtraction), docs updated where the code appears. Tests: PSM and precursor q on opposite sides of the threshold; pooled table refused.

#19, report TSVs. is_transferred and transfer_q columns on both TSVs from both writers; the MBR worker writes transfer_q into the augmented scored table (NaN on non-transfers). The acceptance rule is unchanged and documented (docs/12, docs/15): a tighter threshold does not revoke a transfer; a transfer is not protein-group confidence. Tests: transfer basis exported, tighter threshold keeps the transfer, protein rows carry the columns; Python test checks transfer_q is finite on exactly the flagged rows and equals the accepted q. ci/check_smoke.py expects the two new trailing columns.

Guide. CLAUDE.md: retain_top_peaks alternatives do reach features/compete/rescore (peak_rank, selected_peak_rank); the rescore matrix is flat f32 with the ceiling checked before allocation; the top-K contract exists and awaits entrapment validation. Desktop comments no longer claim run-experiment writes no report. docs/29 gains an implementation-status section for A to D with both before/after pairs.

Local checks

Engine cargo fmt --check, clippy -D warnings, cargo test --workspace (258 tests); desktop clippy + cargo test --lib (78); pytest tests/python/test_mbr_worker.py (18); python -m compileall scripts; ci/gen_config_reference.py --check; ci/check_doc_refs.py; ci/check_desktop_ui.py.

🤖 Generated with Claude Code