apex: apex_sim bench + unify finder + cull weighting, re-tune base profile

[jspaezp]

Net change vs main. History is noisy (it builds an apex-profile weighting
experiment and then removes it) -- squash on merge.

What this does

1. apex_sim test-harness (new). Synthetic DIA extractions fed to the real
   TraceScorer. Broad (1695-cyc, production Phase-1) and narrow (150-cyc,
   Phase-3) recovery suites, interferent density that scales with the
   window (0.41 peaks/cyc anchor from a real run), sub-cycle per-seed apex
   jitter, and a ROC-AUC bench separating a real peptide's score from a
   pure-noise twin. Window/width/density anchors derived from a real
   250225 HeLa run.

2. Dead code removed. score_for_viewer path, ViewerResult, full_results.rs,
   build_broad_extraction, find_joint_apex + its tests. Stale docs fixed
   (apex_profile C(t) was documented cos^3*I).

3. Apex finder unified. Pass-3 score_at now scores at Pass-1's apex cycle
   (effective_apex = cycle) instead of relocating via find_joint_apex on a
   separate profile. One apex source.

4. Apex-profile weighting: added, then culled. An intermediate experiment
   added ratio-free coelution + cross-row vote weighting to apex_profile.
   It tripled apex scoring cost (canonical 18->51 us/run, broad
   112->320 us/run) for recovery gains the sim barely discriminates, and
   the sim cannot reproduce the imperfect co-elution the passes were partly
   meant to handle. Net vs main: the weighting is GONE. apex_profile is the
   base composite profile only; runtime is back at the pre-experiment
   baseline.

5. Base profile re-tuned standalone. With the weighting removed, the base
   knobs were re-tuned (Optuna over canonical+broad+narrow recovery,
   rounded to robust non-degenerate values): cos_pow 0.5->0.25,
   i_exp 0.75->1.0, s_ratio 1.75->0.25, blur_passes 1->2. +1.5% summed
   recovery across all three window sizes at no runtime cost.

Runtime

              before feature   with weighting   this PR

canonical us 18 51 18
broad us 112 320 110
narrow us 12 33 12

Validation

• cargo build -p timsseek: clean, no new warnings.
• cargo test -p timsseek --lib: 88 passed.
• cargo test -p apex_sim: 11 passed.
• apex_sim bench (n=4000/800): summed pass2 recovery 907 -> 921 with the
  tuned base knobs; per-suite canonical/broad/narrow all up, none regressed.

Review notes (fable), open for decision before merge

• MEDIUM. Locator/scorer profile split. apex_profile (locator, cos^0.25)
  picks the apex cycle; score_at then reads the 11 features off
  cosine_profile (scorer, cos^3) AT that cycle. On non-co-eluting real data
  the two peaks can sit 1-3 cycles apart, so features get sampled off the
  true peak and deflate a genuine peptide's score. base_score is safe (it
  re-localizes independently); only the feature multipliers ride the
  locator apex. The apex_sim harness co-elutes everything at one cycle, so
  it is structurally blind to this -- "unify is bench-neutral" is asserted,
  not tested. Fix options: re-localize features in a small +/-window on
  cosine_profile around the locator apex, or add a non-co-eluting sim
  scenario (fragment apex offset from precursor/interference apex).

• LOW. Post-unify, pass1% and pass2% recovery columns are mathematically
  identical (Pass 2 no longer relocates). Collapse or relabel.

• LOW. run_discrimination's noise twin zeroes the expected (library)
  precursor intensity too, not just the observed data, making the AUC
  slightly easier than the real signal-vs-noise decision. Fragments are
  handled correctly (ratios preserved).

• The tuned constants are sim-derived and provisional; cos_pow=0.25 sits
  near the search-range floor. A real-data check (bench/wandb_bench.py) is
  the right gate before trusting the +1.5%.

Supersedes

Folds in the content of #68 (apex weighting -- now culled) and #69
(apex_sim realism). Both can be closed.