Fix U2 background subtype-clobber (silent species-correction disable)

[glarue]

Summary

Fixes a silent bug in the species U2 background correction that disables the
correction on annotations carrying non-canonical splice dinucleotides, inflating
scores on compositionally-divergent species and producing high-confidence U12
false positives.

Root cause

SpeciesBackground._build_final_pwm_sets maps all non-canonical splice dnts to
'gtag' (via the FIVE/THREE_DNT_TO_SUBTYPE defaults) and assigned the corrected
matrix into matrices[('u2', pwm_subtype)] in sorted-dnt order with
last-writer-wins (added in c406498 for streaming/in-memory determinism). Because
T-starting 5′ dnts and most non-AG 3′ dnts sort after the canonical GT/AG, a
handful of spurious non-canonical introns (e.g. 3× TT) win the gtag slot. Their
blend weight w = n/(n+n0) is ~0 at low n, so the corrected background collapses to
the human U2 prior — silently reverting the species correction.

Impact

Conditional: harmless on near-human-composition species (revert-to-human is a no-op
— which is why gold validation passed and this went unnoticed), but it inflates raw
scores on compositionally-divergent species (AT-rich protists, etc.). Observed: a
confirmed U12-loss ciliate went 1 → 95 HC U12 on an annotation that happened to
carry a few non-canonical introns.

Fix

Iterate subtypes by (-n, dnt) and first-writer-wins, so the highest-n (canonical)
dnt defines each pwm_subtype and rare low-n non-canonical dnts cannot clobber it.
The sort is order-independent, preserving the streaming/in-memory determinism
c406498 sought. Applied to the 5′/3′ and BPS loops.

Verification

• New regression test (test_low_n_noncanonical_dnt_does_not_clobber): fails
  pre-fix (background reverts to the human prior), passes post-fix.
• Full unit suite green (657 passed).
• End-to-end: the loss ciliate goes 95 → 3 HC; a clean-annotation run of the same
  species is unchanged (1 → 1); the dumped gtag background goes
  0.0006 → 0.136 (≈ the correctly-corrected value).

Validation

• Regression test (test_low_n_noncanonical_dnt_does_not_clobber): fails pre-fix (background reverts to the human prior), passes post-fix.
• Unit suite: green (657 passed on main-line; 12 on v2.4.x).
• Gold recall preserved (buggy→fixed HC): A. thaliana 291→288, D. melanogaster 20→19, Chlamydomonas (loss) 3→3 — negligible; the fix does not disturb validated species.
• Divergent bearers preserved: gut fungi (Neocallimastix/Piromyces/Anaeromyces) and Basidiobolus HC change ≤10 — real U12 calls retained; core_fraction stays high (e.g. 0.97→0.77) so they remain trusted.
• ⚠️ One integration test (test_modesep_pipeline.py::test_continuous_discount_preserves_strong_TPs) fails — but it is pre-existing on main (fails on unfixed main too) and is unrelated to this change. (Not present on the v2.4.x line.)