Add sort-free Rust FDR kernels (q-values, keep-best, fused finalize)

[GeorgWa]

Adds src/fdr.rs with array kernels for the FDR hot spots, mirroring the pandas implementations in alphadia/fdr/fdr.py:

• fdr_q_values: lexicographic argsort by (proba, decoy, tiebreak) plus the running target/decoy FDR and reverse running minimum. Parallel sort via rayon.
• fdr_keep_best: per-group arg-min over the score, ties broken by lowest original index.
• fdr_finalize: fused and sort-free — a per-group arg-min for keep-best, then a fixed-bin histogram of target/decoy counts over proba whose cumulative sums give the q-value per bin. Replaces a chain of up to four O(N log N) sorts.

Motivation is the pandas multi-key sort / cumsum / groupby overhead, which dominates FDR at ~10M+ candidates.

Python side: MannLabs/alphadia#841.

🤖 Generated with Claude Code

[GeorgWa]

Context: the tie-break argument, and why the signature is not frozen

Draft for now — this wants to be reviewed together with MannLabs/alphadia#841, which is the only consumer.

Why fdr_q_values takes a tie-break

q_values_in_order accumulates target/decoy counts along the sort order, so rows with an equal proba have no defined position relative to each other and the resulting q-values depend on which one comes first. argsort_fdr therefore sorts by (proba, decoy, tiebreak) to make the order total, mirroring df.sort_values([score, decoy, extra_sort_columns]) in the Python reference.

The tie-break arrives as &[i64] because Rust cannot compare arbitrary Python objects. On the Python side that is fine for precursor FDR (precursor_idx) but not for protein FDR, which breaks ties on a string accession; #841 maps non-integer columns to order-preserving integer codes rather than branching on dtype.

The constraint is that the pandas implementation stays live — always for fragment competition, and everywhere while these kernels are unreleased — so the two must agree numerically, not merely each be deterministic.

The signature may lose its third argument

There is an open proposal to assign one q-value per exact equal-proba block instead of ranking within ties. If that lands:

• fdr_q_values drops tiebreak entirely;
• argsort_fdr becomes a single-key sort, and par_sort_unstable_by becomes safe because intra-block order stops being load-bearing — currently the total order is what makes the unstable sort acceptable only in combination with the tie-break key;
• the two extra comparison levels and one i64 array come out of the hot path.

So this is a small perf win on top of a correctness one, and it is worth not treating the current signature as settled while reviewing.

Grouping would be on exact f64 equality, deliberately. The ties in question come from bit-identical classifier outputs; scores differing by one ULP are genuinely different and belong in separate blocks. That is a different mechanism from q_values_histogram here, which bins by value ranges — lumping distinct scores together and quantizing q-values to bin width. That approximation is defensible for fdr_finalize at 10M+ candidates but should not be confused with exact tie grouping, and it is why fdr_finalize needs no tie-break argument today.

Status

Draft until reviewed alongside #841. Note that until these kernels ship in a released alphadia-search-rs, the Python side has _RUST_FDR_AVAILABLE = False, so nothing exercises them end to end and the parity tests there skip at module level.