[WIP] rescore: LDA path + env-tunable GBM config

[jspaezp]

What

Exploration of faster Phase 5 rescoring. The GBM rescorer dominates the
pipeline (~180-220s of a ~300s run on the Skyline library). Two additive,
opt-in paths; the default GBM rescore is unchanged.

Not for merge yet - captures the exploration and leaves the tuning knobs in
place.

Results (Skyline library, 250225 Hela off, 2.3M candidates)

GBM baseline on this run = 46539 targets @ 1% FDR, ~182s training.

config                         targets@1%   %GBM    train    speedup
default (1000 it, depth-6)       46539      100%     182s      1x
GBM 200 it / 31 leaves / lr.10   45566      97.9%     66s      2.75x
GBM 100 it / 15 leaves / lr.15   44160      94.9%     33s      5.5x
LDA (linear ceiling)             37984      81.6%     ~10s     ~18x

Path 1: env-tunable GBM config (the likely mergeable win)

The forust default (1000 iterations, depth-6, unlimited leaves) is heavily
over-provisioned. GBMConfig::from_env adds runtime overrides so the
accuracy/speed knob can be swept without recompiling:

TIMSSEEK_GBM_ITERS  _MAX_DEPTH  _MAX_LEAVES  _LR  _EARLY_STOP

200 it / 31 leaves keeps 98% of IDs at 2.75x. Far better trade than LDA.

Path 2: Sage-style shrinkage LDA (TIMSSEEK_RESCORE_MODEL=lda)

• Two-pass Fisher LDA, flat row-major matrix, rayon-parallel and
  deterministic (fixed-chunk reductions). Ridge shrinkage keeps the
  collinear 108-dim feature set solvable without pruning.
• Rank-based inverse-normal transform (van der Waerden) for the per-feature
  normality assumption; missingness indicators + abs() of signed error
  features as linear-friendly feature engineering.
• ~40x faster (Phase 5 ~180s -> ~10s), but caps at ~82% of GBM IDs on the
  hard Skyline library. This is a genuine linear-model ceiling, confirmed
  offline against sklearn LDA / QDA / logistic (all cluster at 79-82%);
  HistGradientBoosting matches forust at 99.6%, so trees are required for
  the last ~18%. LDA is useful for QC/triage, not a drop-in replacement.

Methodology

TIMSSEEK_LDA_DUMP=/prefix writes the raw feature matrix + labels, which
drove an offline numpy/sklearn harness (feature selection, transforms,
model zoo, tree-config sweep). Kept local under bench_out/ (gitignored).
Feature selection did not help - LDA wants all 108 features; INT is worth
~+5600 IDs; the gap is model capacity, not FE.

Open question

Which operating point should become the default: keep 1000/depth-6, or
move to ~200/31 (2.75x, -2% IDs)? Leaving that decision for the merge.

[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     31.98
  moderate_noise                  91.6    91.6        0     31.54
  high_noise+interference         57.8    57.8        1     31.71
  heavy_interference               5.8     5.8       85     35.67
  mismatched_library              56.3    56.3        1     31.37
  absent_top_fragment             16.5    16.5       69     31.45
  absent_precursor                57.8    57.8        1     31.53

=== broad apex-finding (n=500, tol=±2 cycles) ===
  scenario                      pass2%  pass1%   medErr    us/run
  broad_clean                    100.0   100.0        0    189.56
  broad_moderate_noise            50.6    50.6        1    186.13
  broad_high_noise+interf         27.6    27.6      308    185.27
  broad_hard_3x_density            0.8     0.8      487    185.42
  broad_mismatched_library        41.0    41.0      161    184.56

=== narrow recovery (n=1000, tol=±2 cycles) ===
  scenario                      pass2%  pass1%   medErr    us/run
  narrow_clean                   100.0   100.0        0     21.52
  narrow_moderate_noise           81.3    81.3        0     21.27
  narrow_high_noise+interf        65.8    65.8        1     21.11
  narrow_hard_3x_density          10.9    10.9       34     21.10
  narrow_mismatched_library       62.8    62.8        1     21.09

=== narrow score discrimination (AUC, n_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

commit 809299d