apex_sim: production-realistic bench conditions + score-discrimination (AUC)

[jspaezp]

Stacked on #68 (apex-weighting-clean). Bench-only (apex_sim); no timsseek
production changes.

Makes apex_sim benchmark the apex-finder and the score as SEPARATE functions
under production-realistic conditions, and adds a signal-vs-noise score metric.
All additive: existing canonical suite / SimParams behavior preserved by
defaults.

Why

The old bench conflated two jobs and tested neither the way production runs them:

• Pass 1 (apex finder, weighted apex_profile) runs over a BROAD ~1695-cyc window
  (Phase 1) and inside a NARROW ~150-cyc calibrated window (Phase 3).
• Pass 2 (score / main_score) is meant to separate real signal from noise, but
  the bench only measured apex-cycle recovery, never score magnitude.
  It also pinned the apex to an integer index at the SAME position every run, and
  used a fixed interferent COUNT (so a wider window = lower density = easier).

Empirical anchors (derived from a real run)

250225_Desnaux_200ng_Hela ICC_off, default tol. Instrumented the two callsites

• read a persisted results.parquet (11743 targets at q<0.01):

• broad window (Phase 1): 1695 cycles (RT-unrestricted, whole gradient).
• narrow window (Phase 3): 82-206 cycles, median 152.
• peak width: rising/falling ~3 cyc => sigma ~1 (sim's 1.0 already realistic).

Changes

1. Density-scaled interferents: RandomPeaks.density_per_cycle (+ hardness)
   => count = round(density * n_cycles * hardness). Fixed density as the window
   grows; hardness>1 over-loads on purpose. count path preserved.
2. Sub-cycle apex jitter: SimParams.apex_jitter=Some(range) draws the realized
   apex = apex_cycle + U(-range,range) + U(-0.5,0.5) per seed (off-grid, moving
   target); SimData.realized_apex_cycle reports it and run_sensitivity scores
   against it. None keeps the historical on-grid fixed apex byte-identical.
3. ROC-AUC discrimination: run_discrimination scores matched present/absent
   pairs (absent = all fragments AND precursor zeroed, same seed => identical
   noise+interferents) and reports ROC-AUC of the Pass-2 main_score.
4. Realistic suites: broad_suite (1695 cyc) + narrow_suite (150 cyc), density
   0.41/cyc (= canonical 100/245, so window size is the only changed variable),
   jittered apex, per-scenario stressors incl a 3x-density hard variant.
5. Example prints canonical / broad / narrow recovery tables + a narrow AUC
   table. Canonical header kept as === summary so bench_out/score.sh still
   parses it. Each scenario is its own line (comment out to disable).

Results (n=2000, broad n=500, tol +/-2)

Same density, wider window is measurably harder:

high_noise+interf canonical 245cyc 86% broad 1695cyc 54% narrow 150cyc 90%
mismatched canonical 88% broad 59% narrow 93%

hard_3x_density (1.23/cyc) collapses recovery to ~0-8% (explicit stressor).

Score discrimination (narrow AUC, present vs pure noise):

clean 1.000 moderate 0.870 high_noise 0.840 mismatched 0.876 hard 0.680

med(present) > med(absent) in every scenario. The score separates well until
interference piles up, then degrades to ~0.68 -- the axis the old bench could
not see.

Follow-up (out of scope, bench-first)

Pass-2 apex relocation: score_at computes effective_apex from find_joint_apex on
the UN-weighted cosine_profile (global argmax), which drifts onto interferents
in narrow windows. The narrow-recovery + AUC metrics here exist to prove that
fix when it lands.

Validation

• cargo test -p apex_sim: 11 passed (density scaling, jitter determinism/sub-
  cycle, roc_auc, discrimination).
• cargo build -p apex_sim --release: clean.
• Full bench: bench_out/logs/realistic-bench-v2.log.
• Note: apex_sim has a pre-existing clippy field_reassign_with_default idiom in
  the suite base() closures; the new suites follow it for consistency.