Add head subdivision variants and pick s48 after 1.5m/1.85m comparison

TODO.md

@@ -90,9 +90,13 @@ the rest of §C remains the broad review checklist.
       centre-profile evidence: nose tip −0.192, brow −0.150, chin −0.149, jaw step −0.11, forehead
       −0.126→−0.078. Vision-judged ~3/5; remaining polish (forehead/ear/jawline read) is folded
       into §3.2 comparison.
-- [ ] Compare two or three subdivision/loop variants at 1.85 m front/three-quarter/profile and a
+- [x] Compare two or three subdivision/loop variants at 1.85 m front/three-quarter/profile and a
       1.5 m portrait. Pick the variant that reads best before changing all 12 characters; keep the
       rejected variants local.
+      Compared `head_proto` at seg 24 (670 verts), 48 (1,558 verts) and 64 (2,470 verts) at 1.85 m
+      and 1.5 m, front/three-quarter/profile. Vision-judged: **s48 wins** — s24 is too faceted
+      (nose/lips melt, jagged silhouette), s64 is visually identical to s48 at both distances
+      (extra ~900 verts wasted). Rejected variants stay in `tmp/npc-parts/`.
 - [ ] Verify head weights stay on the Head bone, headgear clearance still passes `fit`, and the
       face does not bake recessed creases into dark AO patches.
 - [ ] Get owner sign-off on the one-variant prototype, then roll the approved head/face into the