Fix -wslr/-wslm .sitelh invariant under per-category (safe_numeric) scaling

[famulare]

Fixes the .sitelh invariant documented in the file's own header:

#   LnLW_k: Logarithm of (category-k site likelihood times category-k weight)
#           Thus, sum of exp(LnLW_k) is equal to exp(LnL)

Fixes #533. See that issue for the full root-cause write-up and reproduction.

The bug

PhyloTree::computePatternLikelihood() writes the per-rate/mixture-category
site log-likelihoods (-wslr/-wslm/-wslmr). It chose between per-category
and per-pattern scaling with:

if (params->lk_safe_scaling || leafNum >= params->numseq_safe_scaling)

But the likelihood kernel decides its scaling mode via PhyloTree::safe_numeric
(set in setLikelihoodKernel()), which is broader:

safe_numeric = (params && (params->lk_safe_scaling || leafNum >= params->numseq_safe_scaling)) ||
    (aln && aln->num_states != 4 && aln->num_states != 20);

The two conditions differ only on the num_states clause, so they disagree
exactly when the writer condition is false but safe_numeric is true: a
non-4-state-DNA / non-20-state-protein alphabet (codon, morphology, custom),
run without -safe, on a tree below numseq_safe_scaling (default
2000) tips. In that case the kernel scales per category and stores
scale_num per category (scale_num[ptn*ncat_mix + c]), while the writer took
its per-pattern branch and read scale_num as a flat per-pattern array
(scale_num[ptn]). This mis-indexes the scale counts and applies a single,
wrong scale step uniformly to all categories. On numerically-extreme sites
(categories underflowing at different scale counts) every LnLW_k comes out
exactly LOG_SCALING_THRESHOLD (≈ 177.4457) too low, breaking the invariant.
(-safe or a ≥2000-tip tree make the writer condition true as well, so they
mask the bug rather than cause it.) .siteprob is unaffected because its
posteriors are normalized.

The fix

Use the tree's own safe_numeric flag to select the writer branch, so the
writer always matches the kernel's actual scaling layout.

         double *lh_cat = _pattern_lh_cat;
         double *out_lh_cat = ptn_lh_cat;
         UBYTE *nei2_scale = nei2->scale_num;
-        if (params->lk_safe_scaling || leafNum >= params->numseq_safe_scaling) {
+        // NOTE: must match the kernel's per-category vs per-pattern scaling choice
+        // (PhyloTree::safe_numeric, set in setLikelihoodKernel), otherwise scale_num
+        // is mis-indexed and the per-category log-likelihoods are off by a scale step.
+        if (safe_numeric) {
             // per-category scaling

(same change in the external-branch and internal-branch blocks; two if
conditions).

Verification

Reproduction (codon, deep branches, 5 well-separated FreeRate categories,
seed 7 — fully self-contained via --alisim, see the companion issue for the
commands and the verification script):

build	rows	violations (|dev|>1e-3)	max |dev|
2.4.0 (before)	667	18	177.445710
this PR (after)	667	0	0.000096

Example site 85, before → after:

before:  LnL=-3.8118   LnLW = -181.9381 -182.2600 -183.3261 -188.9624 -215.2784
after:   LnL=-3.8118   LnLW =   -4.4925   -4.8143   -5.8805  -11.5167  -37.8327

Regression / no-change checks:

• DNA (GTR+F+R4, 4-state, per-pattern path): .sitelh is
  byte-identical before and after the patch.
• Protein path is likewise on the unchanged (per-pattern) branch.
• Standard end-to-end run on the bundled example/example.phy
  (GTR+R4 -wslr): invariant holds (max|dev| ≈ 9e-5).
• Morphology (MK+R) and profile-mixture (-wslmr) codon runs: invariant
  holds (0 violations).
• Total tree log-likelihood is unchanged (-10226.6026 before and after),
  and sum_i LnL_i still equals the tree log-likelihood — confirming the
  change touches only the per-category reporting, not the likelihood engine.

Scope

Two-line logic change in tree/phylotree.cpp (plus comments). No behavior
change for DNA/protein; no change to the likelihood engine, tree search, or
any other output.

---

AI-assistance disclosure: this fix was diagnosed and drafted with the help
of Claude Opus 4.8. All changes were reviewed and verified by Mike Famulare
before submission.