docs(map): correct decoy_thresh comment (strict domination, ties keep transcript)

The doc comment described `best_decoy >= decoy_thresh * best_valid` (drop on
tie), but the code uses strict `<` (`best_valid < decoy_thresh * best_decoy`): a
fragment is assigned to a decoy only when the decoy is *strictly* better, so at
the default decoy_thresh=1.0 a decoy that merely ties the best transcript does
NOT dominate and the transcript is kept. This matches C++ salmon's
`bestScore < decoyThresh * bestDecoyScore`. Comment-only.

Investigated + reverted an experimental C++ --denyDecoyOrphans flag: testing
showed C++ already applies this exact strict-domination rule (so a faithful
mirror is a no-op), and a blunt "drop any orphan co-occurring with a decoy"
variant over-dropped (90.9% vs Rust default 93.7%). The real C++/Rust residual is
seeding (Rust surfaces dominating genome pairs C++ does not), not decoy policy.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B7JMur5DmDpECddErpi2JS

Author: rob-p

crates/salmon-map/src/score.rs

@@ -78,7 +78,11 @@ pub struct ScoreConfig {
     pub score_exp: f64,
     /// drop mappings whose weight falls below this
     pub min_aln_prob: f64,
-    /// drop the read if `best_decoy >= decoy_thresh * best_valid`
+    /// drop the read only when a decoy is *strictly* better than the best
+    /// transcript: `best_valid < decoy_thresh * best_decoy` (so at the default
+    /// `decoy_thresh = 1.0` a decoy that merely ties the transcript does **not**
+    /// dominate — the transcript is kept). Matches C++ salmon's
+    /// `bestScore < decoyThresh * bestDecoyScore`.
     pub decoy_thresh: f64,
     /// keep only the best-scoring mappings (each weight `1.0`)
     pub hard_filter: bool,