Remove the b<->y intensity swap from generated decoys

[brendanx67]

Generated (reverse) decoys were badly miscalibrated: measured against an FDRBench
entrapment oracle they carried roughly 12x the reported error rate, ~11.8% true FDP
at a claimed 1% q on Stellar and 7.6% on Astral. This traces to a single defect.

RecalculateFragments mapped target b_k to decoy y_{n-k} and carried the copied
relative intensity along with the relabel. The reasoning behind it -- "which decoy
ion covers the same residues?" -- was sound, but fragment intensity is dominated by
ion TYPE, not by residue coverage: y ions are systematically more intense than b
ions, so relocating intensity across the swap inverts the strongest structural
feature of the spectrum. A source review of six open-source implementations found
five of six map intensity to the same ion label; this swap was unique to Osprey.

Dropping it recovers a near-honest null with no predictive model:

dataset	true FDP @ 1% q, before -> after
Stellar	11.81% -> 1.47%
Astral	7.64% -> 2.03%

For scale, Carafe-predicted library decoys measure 1.92% on the same Astral cell, so
the fixed generated decoys reach parity with model-predicted ones. On an 82-file
SEA-AD set the fixed construction recovers 19% MORE real identifications than library
decoys at a matched TRUE 1% FDP, so this is not a sensitivity trade.

The paired decoy-win fraction tells the same story from a decoy-independent angle:
within a target/decoy pair an honest decoy should beat a known-false entrapment
target about half the time. Before the fix it won 0.205 of the time; after, 0.473,
against 0.501 for library decoys.

The second commit adds the similarity gate every surveyed implementation already
carries (EncyclopeDIA 0.4 fragment overlap, SpectraST 0.7 spectral similarity,
OpenSWATH 0.5 sequence identity). Its effect at library scale is nil -- ~1e-4 of
peptides excluded, entrapment FDP unchanged within noise -- and it is kept for
robustness at small library scale, where palindromes and low-complexity runs are a
far larger fraction.

Two things measured and REJECTED, recorded so they are not re-litigated: Skyline's
+10-unit precursor m/z shift (net negative here, 1.47 -> 1.96% Stellar and
2.03 -> 3.49% Astral -- it buys MS1 separation, but a decoy-only handicap is
spurious separation), and both similarity gates as an FDP lever.

Cross-implementation parity, both sides built from matching source:

dataset	Rust	C#	result
Stellar 3-file	35,222 precursors	35,222	PASS at 1e-9
Astral 3-file	126,474 precursors	126,474	PASS at 1e-9

Stage 7 protein FDR and full blib content both compare clean per-column at 1e-9.

The matching ProteoWizard/pwiz change is on branch
Skyline/work/20260726_osprey_regression_redesign, which also re-records the C#
regression goldens against this construction and adds a libdecoy+entrapment
regression dataset with FDR calibration spot checks. The two must land together:
the C# golden re-recording is only defensible alongside this change.

cargo fmt, clippy -D warnings and cargo test all pass.

---

Added 2026-07-27: the ladder had a second defect, and the invariant it relies on was enforced on one path

theoretical_ladder dropped every y ion when any residue was unknown. It derived y as
total - prefix, so one residue missing from the mass table made total NaN and killed
the whole y series - not just the ions spanning it, which is what the comment above it
promised. A LEADING unknown emptied the ladder outright, and an empty candidate ladder is
read by is_candidate_acceptable as "accept": the gate silently switched itself off for
exactly the peptides it could not evaluate.

Not hypothetical. The Astral reference library (1,148,928 distinct peptides) carries 60
selenocysteine (U) peptides; the Stellar library carries none. With half their ladder
missing, the inflated overlap ratio pushed the plain reversal past 0.4 and forced the
cycling fallback for three of them:

GPSPPPMAGGUGR        was PSPPPMAGGUGGR       now GUGGAMPPPSPGR
GFVCIVTNVASQUGK      was FVCIVTNVASQUGGK     now GUQSAVNTVICVFGK
SYSSKPPLWAAELCVTUSS  was YSSKPPLWAAELCVTUSSS now SUTVCLEAAWLPPKSSYSS

y ions now come from suffix sums, mirroring the prefix sums used for b ions.

MIN_PEPTIDE_LENGTH = 6 is now enforced by all three format loaders (DIA-NN TSV,
blib, elib) rather than assumed. The overlap gate compares the FULL ladder, so its two
permutation-invariant rungs put a structural 1/(n-1) floor under the ratio; at length 3
that floor is 0.5 against a 0.4 budget and every candidate would be rejected, dropping the
peptide from the search with no diagnostic. At the enforced minimum it is 0.2. The bound
cannot fire on real input - the Astral library's shortest peptide is 7 residues - so it
exists to make a malformed library fail loudly instead of quietly changing what is
searched.

Tests. Neither the swap removal nor the overlap gate had any. Four were added, mirroring
OspreySharp's DecoyConstructionTest one for one: ion type + ordinal preserved with the m/z
read off the decoy's own ladder; the gate rejects an isobaric near-duplicate (AILLAK ->
ALLIAK, where I/L isobaricity makes every rung coincide) and accepts an ordinary
tryptic reversal, so a gate that rejected everything cannot pass; generation never emits a
rejected candidate; and the ladder drops only the ions spanning an unknown residue. Both
implementations assert the same rung counts, so a change on either side that is not mirrored
breaks both suites.

Also restores the doc comment for generate_all_with_collision_detection, which the
overlap-gate commit left attached to MAX_FRAGMENT_OVERLAP, and rewrites the docs'
ion-swap section to describe what the code now does.

The matching OspreySharp PR is ProteoWizard/pwiz#4480 and must land with this one. It
re-records the Astral golden for the corrected ladder and adds the regression coverage that
would have caught the swap.

Gates: cargo fmt --check, clippy -D warnings, cargo test all green.