Preserve library RT coordinates and calibrate search dimensions independently

[jspaezp]

Library RT was treated as acquisition seconds, and a failed RT fit prevented m/z and mobility calibration. This change preserves the source RT domain and calibrates searchable dimensions independently.

• One RT descriptor per library: absent, seconds, normalized index with optional scale, or unspecified. Measured minutes convert to seconds; normalized values retain negative values and zero. Mixed availability or axes are rejected.
• mzSpecLib reads effective attributes and grouped units through rustyms PR #6, pinned at c033092. JSON loading and prediction preserve explicit axes and declared scales.
• Source geometry exposes LibraryRT<f32>. A borrowed ExtractionQuery carries separately resolved ObservedRTSeconds<f32> bounds. Collectors own reusable buffers and retain no library lifetime. Index calls accept PeakTolerance; RT is resolved once, preserving existing tolerance arithmetic.
• Isotope extraction shifts the collector's fragment buffers directly. No owned target materialization or library-axis clone per scored candidate. Python query adapters also borrow their input geometry.
• Missing/failed RT fits search the full acquisition and still measure m/z and supported mobility at score-selected apexes, without ridge filtering. Unmeasured dimensions retain configured tolerances. Missing RT preserves decoy competition, rescoring and q-values; it does not select raw mode. Prescore heaps remain bounded and mergeable.
• The viewer uses calibrated acquisition coordinates or full-run extraction without an RT fit, including libraries measured in seconds. Direct-query CLI/Python seconds inputs explicitly denote acquisition coordinates. Existing CLI output fields remain compatible.
• Results format 5 includes library-axis metadata; calibration format v4 supports partial calibration. Older calibration files must be regenerated. FAIMS/absent mobility never supplies sentinel calibration measurements.

Validation:

• 142 query/reader, 201 scoring, 108 search CLI, 12 query CLI, 10 viewer and 17 calibrt unit tests passed. Existing ignored tests remain ignored. Carafe, library-building integration and doctests passed.
• Tests cover source/acquisition type separation, clipping, explicit full-run selection, invalid RT rejection, collector lifetime independence, source RT remaining unchanged, index use of resolved bounds, and buffer reuse.
• Workspace/all-target Clippy passed with -D warnings -A clippy::too_many_arguments; the argument-count exception covers existing CLI/viewer/Python signatures. Dashboard feature checks, task fmt and git diff --check passed.
• Release search on the supplied Astral HeLa mzML, with 948,957 library entries and 474,327 supplied decoys, completed for all four RT variants. All retain rescoring and finite output scores/q-values. Missing/failed RT preserves m/z calibration, leaves RT predictions/residuals unavailable and has zero RT-boundary skips. Absent RT omits RT features; constant RT produces non-finite features that rescoring culls. Python bindings were compile-checked, not exercised in a Python runtime.

Library RT	Targets at q ≤ 1%	RT behavior
Intact	28,159	Fitted
Removed	23,754	Full run; m/z calibrated
Shuffled, seed 20260923	11,013	Broad spurious curve accepted
Constant zero	21,300	ZeroRange; full run with m/z calibration

Shuffled RT exposes an existing estimator limitation: the fitter accepts a monotonic path through noise. Its ridge half-width averages 411 s versus 96 s for intact RT. Fit-quality rejection remains separate work; no scientific threshold is added here. These estimated q-value counts on one acquisition do not establish independent FDR accuracy or the 1,000-file storage requirement.

Closes #115.

Follow-up work:

• Audit DIA-NN TSV/Parquet RT provenance: these readers still convert RT values to seconds and use the default seconds axis. Verify normalized exports, tag their actual axis, and add reader fixtures.
• Add fit-quality rejection for shuffled/uninformative RT, using representative positive and negative datasets to choose acceptance criteria.
• Viewer's calibration Apply writes a uniform RT tolerance where a real search uses a position-dependent one #83: make viewer extraction use per-query ridge widths and the same fallback tolerance as search.
• Represent mobility observables explicitly and degrade searches safely #117: represent mobility observables and absence explicitly; the calibration changes here do not implement that schema.

[jspaezp]

for sure type this ...

[github-actions]

Apex-finder bench

cargo run -p apex_sim --release --example bench -- 1000 2

Sensitivity + timing across canonical scenarios

=== summary (n=1000, tol=±2 cycles) ===
  scenario                      pass2%  pass1%   medErr    us/run
  clean                          100.0   100.0        0     27.39
  moderate_noise                  92.7    92.7        0     26.72
  high_noise+interference         59.3    59.3        1     26.73
  heavy_interference               7.4     7.4       88     26.88
  mismatched_library              59.8    59.8        1     26.61
  absent_top_fragment             16.0    16.0       65     26.65
  absent_precursor                59.3    59.3        1     26.71

=== broad apex-finding (n=500, tol=±2 cycles) ===
  scenario                      pass2%  pass1%   medErr    us/run
  broad_clean                    100.0   100.0        0    162.39
  broad_moderate_noise            50.6    50.6        1    158.24
  broad_high_noise+interf         27.6    27.6      308    159.42
  broad_hard_3x_density            0.8     0.8      487    158.27
  broad_mismatched_library        41.0    41.0      161    158.98
  broad_measured_density          49.6    49.6       13    187.94

=== narrow recovery (n=1000, tol=±2 cycles) ===
  scenario                      pass2%  pass1%   medErr    us/run
  narrow_clean                   100.0   100.0        0     18.66
  narrow_moderate_noise           81.3    81.3        0     18.29
  narrow_high_noise+interf        65.8    65.8        1     18.22
  narrow_hard_3x_density          10.9    10.9       34     18.29
  narrow_mismatched_library       62.8    62.8        1     18.31
  narrow_measured_density         83.6    83.6        0     20.89

=== narrow score discrimination (AUC, n_seed_pairs=1000) ===
  scenario                         AUC   med+signal    med-noise
  narrow_clean                   1.000     4.329e14      7.199e7
  narrow_moderate_noise          0.870      4.054e9      3.054e8
  narrow_high_noise+interf       0.830      2.761e9      3.602e8
  narrow_hard_3x_density         0.684      3.644e9      1.244e9
  narrow_mismatched_library      0.842      2.172e9      3.575e8
  narrow_measured_density        0.858      4.341e9      2.562e8

commit 245d852