fix(xbg7): degenerate index runs 582 -> 1 (grouped path + prefer a clean candidate)

Two follow-ups to the pad-scoring fix, both driven by the same invariant (a
correctly located index run has no degenerate triangles):

- anchor_grouped_meshes picked its pad by first-match too; scoring the pivot run
  the same way cleared every remaining ptc_pack composite (f102/f104/e107).
- anchor_pool_mesh now prefers a degenerate-free candidate over an earlier dirty
  one. examples/better_home.rs showed the last two resources each had exactly one
  degenerate-free, pool-covering block, sitting later in file order than the
  lookalike we took. First-match order is kept for every clean hit, and a dirty
  block is still used if nothing clean exists, so coverage cannot regress.

  degenerate index runs, disc-wide:      582 -> 11 -> 1
  captured index runs identical:         93/93 (unchanged)
  resources decoded / misses:            6 209 / 85 (unchanged)
  index runs changed / anchors moved:    590 / 10 (_rou_f402_dead x8, e201_bdy_03_m x2)

Cross-container minority decodes 89 -> 96, and that is progress: all seven new
rows are _rou_f402_dead, which now has a majority (32x25x8) for the first time, so
its seven wrong copies are named instead of hidden behind "no majority".

The last dirty run (_rou_f402_dead in Stage_S09) is blocked by distinct assignment
— its clean block is claimed by e_rou_f003_Near, both 24-vertex bounding boxes. A
winding-floor escalation for that case was written, measured to fire for nothing,
and reverted; the reasoning is kept as a comment.

Regression threshold tightened to 1. Suite green with --include-ignored apart from
the pre-existing known-failing cross-container consistency test.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NptfmpjdpNCKEez6d2xvA9

docs/re/structures/xbg7-mesh.md

@@ -1458,3 +1458,53 @@ panel lines, bridge tower and funnels. Same 31 991 triangles, same 4 placements,
 same bounds — only the wiring differs. Third time this project has learned it:
 **render the output**; a metric that cannot see a 1 600-unit slab could not see
 this either.
+
+### ✅ Following it through: degenerate index runs 582 → 1 (2026-08-13)
+
+The pad-scoring fix left 16 dirty sub-meshes. Two follow-ups cleared all but one.
+
+**1. The grouped path had the same first-match bug.** `anchor_grouped_meshes`
+picks `ib0 = vb0 − span − pad` the same way, so it got the same scoring (pivot
+run's degenerate count, then winding). That cleared every remaining `ptc_pack`
+composite — `f102_break.dat` (161 degenerate triangles), `f104_break.dat` (3
+sub-meshes) and `e107_break.dat` — 16 → 11.
+
+**2. Prefer a degenerate-free candidate over an earlier dirty one.** The last two
+resources were anchored on a *lookalike earlier in file order*:
+`examples/better_home.rs` scores every candidate `(start, pad)` for a resource and
+showed **exactly one** degenerate-free, pool-covering block for `e201_bdy_03_m`
+(`0x316383C`, ours was `0x248D054` with 4 degenerate triangles) and one for
+`_rou_f402_dead`. So `anchor_pool_mesh` now keeps first-match order for every
+clean hit and only searches on when the accepted block is provably mis-fitted,
+falling back to the dirty block if nothing clean exists (coverage can never
+regress). 11 → 1.
+
+| | first-match | + pad scoring | + these two |
+|---|---|---|---|
+| decoded sub-meshes with a degenerate index run (disc-wide) | 582 | 11 | **1** |
+| captured index runs identical (`Stage_S02`) | 76/93 | 93/93 | **93/93** |
+| resources decoded / misses | 6 209 / 85 | 6 209 / 85 | **6 209 / 85** |
+| sub-meshes whose index run changed | — | 580 | 590 |
+| of those, vertex anchor moved | — | 0 | **10** (`_rou_f402_dead` ×8, `e201_bdy_03_m` ×2) |
+
+**The consistency screen went 89 → 96 minority decodes, and that is progress.**
+All seven new rows are `_rou_f402_dead`: eight containers now agree on a
+**32 × 25 × 8** box, which gives the resource a *majority for the first time*, so
+the seven copies that still land elsewhere (`160×160×78`, `519×100000×519`,
+`856×463×803`, `779×5769×5769`) are finally **named** instead of hiding behind "no
+majority". Textbook case of consistency being the weaker witness: the number rose
+because the decoder got better.
+
+**The one remaining dirty run** is `_rou_f402_dead` in `Stage_S09`. Its
+degenerate-free block (`0x17AE620`, validates at pad 0) is **claimed by
+`e_rou_f003_Near`** — so distinct assignment blocks it, not a gate. Both are
+24-vertex bounding boxes: the identity class that needs descriptor-level data, not
+another heuristic. A winding-floor escalation for this case was written, measured
+to fire for **nothing** on the disc, and reverted (the comment survives in
+`anchor_pool_mesh`).
+
+**Next target, with the caveat that killed the last attempt:** six containers
+decode `_rou_f402_dead` to a *clean but wrong* block, so several degenerate-free
+candidates exist and file order picks badly. The majority span would separate
+them — but "flag, don't silently rewrite geometry on a vote" still stands, so this
+needs the box-identity data, or a capture of a stage that draws it.