[brendanx67]

Brings Rust in line with the C# implementation, which shipped this change first
(ProteoWizard/pwiz#4593). Without it the two diverge substantially on Astral, so
this restores the cross-impl gate.

The problem

The first-pass training subsample represented each precursor by its BEST
observation across all files. Every training row was therefore a maximum over
however many runs were batched together, while the model it produces scores
ordinary per-run rows. The mismatch grows with the batch: mean coelution_sum
is 1.19 at one file and 5.75 at 82, against 1.18 in the population actually
being scored.

The change

Represent each precursor by ONE uniformly sampled run, contributing that run's
best candidate peak, at the same row count and the same memory. A size-1
reservoir decides: the k-th run a precursor appears in takes the slot with
probability 1/k, drawn from a SplitMix64 mix of (base_id, k, seed) rather than
a shared RNG, so thread scheduling cannot move the choice and a re-run trains on
the same rows.

Both halves matter. Pass 1 is pre-compaction, so a precursor carries several
candidate-peak rows per run. Drawing per ROW rather than per RUN would weight a
run by how many candidates it produced and leave a random candidate as the
training row, which measured -17.4% identifications on a 3-file test in C#.

Targets and decoys of one base_id are separate precursors that share it, so
both the counter key and the draw take the decoy bit -- masking it off would make
a target and its paired decoy decide identically at every k and land on the same
run for essentially every precursor.

No opt-out switch. C# keeps one only as an A/B lever for the change itself; Rust
tracks the shipped default.

Results

82 SEA-AD Astral files, matched true entrapment FDP, experiment scope:

	pass 1	pass 2
baseline (cross-run maximum)	34,552	35,327
one sampled run	45,943	52,911
	+33.0%	+49.8%

A wash at 3 files on every dataset carrying a true-FDP oracle, so this is a
large-batch effect, as the mechanism predicts.

Cross-impl parity

Before this change, Rust and C# diverged on the 3-file Astral set by 2,096,935
of 6,226,744 scores, 454 precursors, and 85 protein-group winner flips.

Compare-EndToEnd-Crossimpl.ps1 now reports PASS at 1e-9 on both Stellar and
Astral -- precursors, blib content, Stage 7 protein FDR, and FDR sidecars.

Also: pearson_correlation_raw on degenerate input

A constant series has zero variance and no defined correlation. Clamping the
denominator to 1e-30 and dividing by its root manufactures an arbitrary tiny
value out of numerator noise. This returns 0.0 instead, matching both C# helpers
(PearsonCorrelation.Pearson and TukeyMedianPolish.PearsonCorrelationRaw).

It is in this PR because it has no independent motivation -- it earns its place
as the last thing the parity gate needed. It surfaced as
median_polish_residual_correlation = -8.47e-7 against C#'s exactly 0 on one
precursor whose every fragment pair was degenerate, moving one SVM score by
8.6e-8 (~48 million ULPs at that magnitude, so not rounding noise).

test_pearson_constant_input previously asserted only is_finite(), which the
clamped form satisfies. It now pins the value, plus the both-constant and
all-zeros cases.

Testing

• cargo fmt --check, cargo clippy -D warnings, cargo test -- all pass
• New reservoir_takes_slot_matches_cross_impl_vectors: 21 pinned
  (base_id, seen, expected) values shared with C#, including decoy-bit-set
  cases and true values at seen=5/7/82 so a broken implementation cannot pass
  by returning false
• New reservoir_spreads_winners_evenly_over_runs: mirrors the C# uniformity
  test over 82 runs and 200,000 precursors
• Compare-EndToEnd-Crossimpl.ps1 on Stellar and Astral, both PASS at 1e-9