Syplheed-Reborn/docs/re/structures/xbg7-mesh.md

# XBG7 — mesh geometry (inside XPR2 model containers)

- **Confidence:** 🟡 `PROBABLE` for the single-stream layout (below); ❔ `HYPOTHESIS` / undecoded
  for the complex multi-stream body layout.
- **Parser in:** `sylpheed-formats/src/mesh.rs` (`Xbg7Model::from_xpr2`), tests
  `tests/mesh_disc.rs`. Container parsing reused from `src/texture.rs` (`Xpr2Header` /
  `Xpr2ResourceEntry`).
- **Applies to:** ship / weapon / prop models in `hidden/resource3d/*.xpr` (166 files).
- **Method:** clean-room — static hex inspection of the retail disc + geometric validation of the
  recovered triangles (non-degenerate area, indices in range, bbox matches the descriptor's stored
  size). No game code decompiled or copied.

## Where XBG7 lives

Models are ordinary **`XPR2`** containers (see the XPR2 texture doc / `texture.rs`). The 16-byte
resource-directory entries (from file offset `0x10`) carry `TX2D` texture resources **and** one or
more `XBG7` geometry resources:

```
entry = [ tag:4 ][ data_offset:u32 ][ descriptor_size:u32 ][ name_offset:u32 ]   (big-endian)
```

Offsets are relative to the directory base `0x10`. The `XBG7` *descriptor* (at `data_offset+0x10`,
`descriptor_size` bytes) is a **scene / material / node graph** — it holds node names
(`rou_f001_mnt1_root`, `Light`, …), a bounding value (`0x41F00000` = 30.0 ≈ the ship's ~30-unit
length), material names matching the `TX2D` channels (`_col` albedo, `_spc` specular, `_gls` gloss,
`_lum` luminance), and per-sub-mesh records. The **vertex / index buffers** live in the container's
shared **data section** (from `header_size`).

## Sub-mesh records (in the descriptor)

Read in file order by a sliding 4-byte scan; each is a big-endian tuple:

```
[ vtx_count:u32 ][ 0:u32 ][ idx_count:u32 ][ tail:u32 ]
   3..=65535        ==0     mult. of 3        1..=64
```

(For `rou_f001_wep_00`: `vtx_count=215`, `idx_count=1092` — matches the recovered geometry exactly.)

## The single-stream data layout — 🟡 PROBABLE (decoded, GPU-cross-checked)

For 36 of the 166 models (weapons, simple props) the data section is a straight sequence of
sub-meshes, carved from `header_size` in record order:

```
per sub-mesh block:
  [ 12-byte header  (contents undecoded) ]
  [ index buffer  : idx_count × u16 BE ]        triangle list (prim=4, GPU-confirmed)
  [ vertex buffer : vtx_count × stride bytes ]  ← declaration-driven
  (pad to 16 bytes → next sub-mesh)
```

The 12-byte header precedes the **index** buffer (the same block shape as stage
resources — see below); the vertex buffer follows the indices with no further
gap. (Earlier this was mis-modelled as `[index][12-byte gap][vertex]`, which put
the vertex buffer at the identical offset but read the index buffer 12 bytes too
early — turning the 12 header bytes into 6 junk indices = **2 leading degenerate
triangles** (a stray-triangle artifact) and dropping the last 6 real indices.
Skipping the header fixes the triangle list with no change to vertex coverage.)

**Vertex declaration.** The layout is **not fixed-stride**. The descriptor holds a declaration table
(right after the `(index_bytes, index_count)` marker) of `{offset:u32, format-code:u32, usage<<16:u32}`
big-endian triples, terminated by `offset == 0x00FF0000` / `code == 0xFFFFFFFF`:

| usage | element  | format code | format | size |
|-------|----------|-------------|--------|------|
| 0x00  | POSITION | `0x2A23B9`   | f32×3  | 12 B |
| 0x03  | NORMAL   | `0x1A2360`   | f16×4 (use xyz) | 8 B |
| 0x05  | TEXCOORD | `0x2C235F`   | f16×2 (u,v) | 4 B |

Stride = max element extent. Models **omit elements** → variable stride (20 = pos+normal, no UV;
24 = pos+normal+uv). Each element is read in **naive big-endian component order** (the raw file
bytes; see the endianness note below). Assuming a fixed stride-24 was why the old decoder mis-aligned
and declined the pos+normal-only models.

**Alignment pinned by the normals.** The vertex buffer starts at `index_end + 12` (a fixed 12-byte
header — NOT `align16`, which lands 4 bytes early on most models). The correct offset is the unique
one where recovered normals are exactly unit-length.

**Safety gate:** every index is validated `< vtx_count`, and when the declaration has a normal
element the mean recovered `|normal|` must be ≈1 (`[0.5, 2.0]`). A model failing either is rejected
(`MeshError::UnsupportedLayout`) rather than emitting garbage.

### Endianness — file bytes are naive-BE, `k8in32` is a red herring ✅

The Canary GPU capture reports every vertex stream with fetch `endian = 2` (`k8in32`, each 32-bit
word byte-reversed). This describes the **guest-memory** copy the GPU fetches — **not** the `.xpr`
file bytes. Reading the file with a `k8in32` transform breaks the normals (mean `|normal|` → 1.33);
the plain per-element big-endian read yields exactly unit normals. So the game **rearranges** the
vertex data between the on-disc `.xpr` and the uploaded buffer; the decoder reads the file directly
and must use naive BE.

## Stage containers — multi-resource, grouped pools ✅ (content-anchored)

`hidden/resource3d/Stage_S*.xpr` are not single models but **collections of
enemy / prop sub-models** — up to ~400 `XBG7` resources each (e.g. `Stage_S07` =
378). Their data layout differs from the weapon files:

- Each resource is a block `[12-byte header][index buffer][vertex buffer]`.
  Unlike weapons there is **no 12-byte gap between index and vertex** — the
  vertex buffer directly follows the indices.
- The **index count** is the descriptor's `(index_bytes, index_count)` marker
  (the *total* for the resource — a resource may have several sub-meshes summing
  to it), **not** the first sub-mesh record. The **vertex count** is a `u32`
  stored **32 bytes before** the marker.
- The blocks are **scattered through the data section, interleaved with the
  container's texture data**, in an allocation order that is **not** directory
  order and is **not stored** in any descriptor field we could find (the
  descriptor holds sizes — `rel 160 ≈ index_bytes+3`, `rel 164 =
  0x1000_0000 | (vertex_bytes+2)` — but no data offset). Reconstructing that
  allocation order is unsolved.

Because the offset is not stored, each resource's block is located by
**content**: parser `Xbg7Model::stage_models` does one `O(file)` pass per
distinct stride to find every **vertex-buffer start** (an offset whose NORMAL —
`f16×4` at vertex `+12` — is unit length while the *previous* stride slot's is
not, i.e. a run boundary; ~one candidate per block, not millions), then pins each
resource to the unique candidate where the `index_count` indices ending just
before it are all `< vertex_count`, reference nearly all vertices, and yield
non-degenerate triangles with real extent. This is fast (≤ ~2 s on a 70 MB
stage) and unambiguous (no two resources collide). Blocks that fail validation —
the few quantized hero bodies — are **skipped**, never emitted as garbage.

**Coverage:** 5662 sub-models decode across the 22 stage files (e.g. `Stage_S07`
366/378, `Stage_S10` 7/9 — including the main enemy bodies `e003`/`e005`, their
LODs, weapons, and props). The viewer (`spawn_stage_models`) lays the decoded
sub-models out as a side-by-side "cast sheet", skipping the few huge skybox-plane
resources (300 k-unit quads). Cross-checked: `e003` = 2383 v / 1436 t bbox
23.5×9.5×32.5; `e005` = 2566 v / 1507 t; both 0 degenerate.

## Not yet decoded — ❔ the complex body layout

The hero-ship *body* meshes (`DeltaSaber_A/_T/_W.xpr` resource `f004`, and ~100 other models) still
decline. Two open sub-problems: (1) **multi-sub-mesh models** whose *first* sub-mesh decodes but a
later one's inter-mesh offset isn't yet handled (the `align16` advance is a guess) — these are
declined whole; (2) the big body meshes, where the data section does not start with an index buffer
and the geometry sits at descriptor-addressed offsets. NB the GPU capture showed **every** rendered
mesh is *single-stream* (just wider strides, e.g. 44 bytes = pos + f32×3 + colour + 2×f16×4), so the
body is likely single-stream-with-a-richer-declaration rather than the "separate streams" first
guessed — it was simply not rendered in the captured session (menu only). A capture taken *in a
mission* (where `DeltaSaber` renders) would hand over its exact declaration directly. `DeltaSaber_A`'s
5 XBG7 blocks are `f004` (body) + `_rou_f004_mnv01_L/_R`, `_mnv02`, `_turn180` (maneuver / pose).

## Evidence log

- 2026-07-12 — `rou_f001_wep_00.xpr`: XPR2 dir = 1×XBG7 (`f001_wep_00`) + 3×TX2D
  (`_col/_gls/_spc`). Data section starts with a u16-BE index run (max 214), then stride-24
  vertices. Descriptor record `[215,0,1092,4]` at desc `+0x2B0`; index-buffer byte size `0x888`
  (=2184=1092×2) at desc `+0x1A0`. Recovered mesh = 215 v / 364 t, 362 non-degenerate, median tri
  area 0.015 — coherent. → single-stream layout **PROBABLE**.
- 2026-07-12 — descriptor-driven sequential carving over all 166 models: **42 carve** under the
  index-only check. Real 3D extent confirmed on `rou_f001_wep_04` (bbox 3.7×1.2×3.5).
- 2026-07-12 (refine) — attribute ranges differed per model (`wep_00` UV≈attr0/2, `wep_04`
  attr4/5, `wep_03` attr1≈±60000 = garbage) → **refuted the fixed "6 half attrs, UV=attr0/2"
  reading.** Found the descriptor **vertex declaration** (offsets 0x0C/0x14, usages 0x03 NORMAL /
  0x05 TEXCOORD). Solved the vertex-base offset with a **unit-normal validator**: `index_end + 12`
  gives median `|normal| = 1.000` on every weapon model (`wep_00/02/03/04`), vs `align16` landing 4
  bytes early. UVs then land in `[0,1]`. Adding the normal gate: **25 models decode clean +
  normal-valid** (the rest — incl. `Stage_S*` degenerate blobs — correctly declined).
- 2026-07-12 (DYNAMIC) — added a cvar-gated draw logger to Canary
  (`command_processor.cc::LogDrawForRE`, cvar `log_draws`), captured the Ready Room / Briefings.
  **GPU ground truth confirmed the static layout exactly**: primitive `prim=4` = **triangle LIST**
  (settles the list-vs-strip question), and a stream with `f32x3 @offset0` + `f16x4 @3dw` +
  `f16x2 @5dw`, stride 6 dwords = 24 bytes — matching `POSITION@0, NORMAL@0x0C, TEXCOORD@0x14`.
  Revealed **stride varies** (24, 20, 28, 44 …) and that all streams are **single-stream** →
  parsed the declaration for variable stride: coverage **25 → 36** (e.g. `wep_05` is pos+normal,
  stride 20, no UV — previously mis-aligned). Also confirmed the endianness note above: fetch
  `endian=2` (k8in32) is the *guest* copy; file stays naive-BE. `Stage_S*` now decode (stride 20).
- 2026-07-12 (STAGE) — `Stage_S*.xpr` decoded as multi-resource containers. Found each geometry
  block is `[12B hdr][index buffer][vertex buffer]`, index count = descriptor marker (total, e.g.
  `e003` = 4308 spanning 2 sub-meshes), vertex count = `u32` at marker−32 (`e003` = 2383 → verified
  by max-index 2382 and 0 degenerate tris, bbox 23.5×9.5×32.5). Blocks are scattered among texture
  data with **no stored offset** (descriptor rel 160 = idx_bytes+3, rel 164 = `0x1000_0000 |
  (vtx_bytes+2)` are sizes, not offsets; block starts e.g. `e003`@0x10230, `e003_l`@0x5d000,
  `e005_l`@0x9b000 are not directory-ordered). Solved by **content anchoring**: a single per-stride
  pass finds vertex-run starts (unit NORMAL at +12 whose previous slot isn't), then match each
  resource by strict index+triangle validation. **5662 sub-models decode across 22 stages** (S07
  366/378), incl. the previously-declined main bodies `e005` (2566 v) and weapons — ≤2 s on 70 MB.
  Parser `Xbg7Model::stage_models`, tests `stage_models_{decode,sweep,quality_audit}`.
- 2026-07-17 — **triangle-LIST re-confirmed; a strip interlude refuted; winding-consistency gate
  added.** A 2026-07 change had briefly re-read the index buffers as triangle *strips* (to "fill
  holes"). Refuted objectively with a new `XVERIFY` diagnostic that compares both readings by
  **stored-normal agreement** (each triangle's cross-product face normal vs the sum of its vertices'
  stored normals): the LIST reading gives agreement **1.000** on every clean weapon (`wep_00/03/04/19`
  = only possible with correct topology + winding), the STRIP reading **~0.49** (random). The strip
  reading also over-generated ~2.5× the triangles (wep_00: 938 vs 364) — a hole-filling garbage soup.
  Reverted to LIST in both paths (`from_xpr2`, `read_pool_mesh`), matching the `prim=4` GPU capture.
  Added an objective **winding-consistency gate** `max(na, 1-na)`: a correct carve is internally
  consistent (agreement ≈1.0, or ≈0.0 for inverted-but-consistent winding — a real single-sided
  mesh), a mis-carve scatters to the ≈0.5 middle. `from_xpr2` declines sub-meshes below 0.90 (e.g.
  `wep_23` na=0.398 → declined instead of a spike-mess); the single-model **content-anchor fallback**
  gates at 0.85; the large multi-resource **stage** path stays ungated (its enemy meshes span a
  continuous 0.5–1.0 consistency range — a hard gate there dropped ~48/314 legit S07 blocks).
  **Routing fixed:** `decode_models` (CLI) and the viewer now route by `count_xbg7` (1 → validated
  records-based list decode, fallback to strict-gated anchor; >1 → stage anchor) instead of the old
  "whichever decoder yields more verts" rule — that rule let stage content-anchoring win on
  single-model weapon files and fabricate **phantom** blocks (a `wep_00` clone appearing inside
  `wep_19`), duplicates, and spike-mess anchors. Weapons now: 33 clean-decode / 26 declined (declined
  = genuinely multi-stream or un-carvable, shown as nothing rather than garbage); stage coverage
  unchanged (S07 314). `expand_triangle_strip` retained as an `XVERIFY`-only diagnostic.
- 2026-07-12 — `DeltaSaber_A.xpr` body: data does **not** begin with indices; plain-`f32×3` runs
  with ship-scale extent (span ≈27–34, matching bbox 30.0) found only at high offsets
  (`data+0x28634C`, …) → multi-stream, **undecoded**.
- 2026-07-18 — **GROUPED-POOL layout cracked → the hero ship (Delta Saber) fully decodes.** The
  detailed models (`DeltaSaber_*.xpr` + ~100 others) were declined for **location**, not format —
  their vertex format is the standard stride-24 triangle list. A resource's *several* sub-meshes
  don't interleave `[idx][vtx]` per block; they share **two grouped pools**: an **index pool**
  (buffers concatenated in descriptor-marker order, each **4-byte aligned**) followed by a **vertex
  pool** (each sub-pool `vtx_count × stride`, same order), with the index pool ending **exactly**
  where the vertex pool begins. So the whole resource pivots on one unknown, the first vertex-pool
  start `vb0` (= index-pool end, found by the unit-normal vertex-run scan); everything else is
  derived: `ib0 = vb0 − span`, `ib[i] = align4(ib[i-1] + idx_count[i-1]·2)`,
  `vb[i] = vb[i-1] + vtx_count[i-1]·stride`. Reversed statically from `DeltaSaber_T.xpr` and
  **cross-checked against a Canary GPU draw-log capture** (mission ship = `DeltaSaber_T.xpr`, found
  via the `--log_file_io` kernel hook): `f001` = body (idx@`data+0xC` = 0x5500C, vtx@0x61ACC, 10891 v
  / 8187 t) + **7 detail parts** (fins/cockpit/wingtips, markers at descriptor 0x3BEC…0x58FC) =
  **8650 tris**, and **every sub-mesh decodes at 0 degenerate / full coverage / winding-agreement
  1.000**. This is the layout the per-block adjacency anchor (`ib = vb − idx_bytes`) rendered as a
  **spiky phantom** (it read 24561 indices starting 2782 B too late, agree 0.64, 1277 degenerate).
  Insight: a single index marker reduces the grouped model to `index_end = vb0`, i.e. the existing
  adjacency `ib = vb − idx_bytes` — so grouped **generalises** the single-block anchor (n=1 is
  identical). Implemented as `anchor_grouped_meshes` (mesh.rs): `anchor_models` routes resources with
  >1 index marker to it (validated per sub-mesh; on failure falls back to the old first-marker
  adjacency anchor so stage coverage never regresses); single-marker stages/props keep the exact
  prior path. The shared acceptance test is factored into `validate_block` (the connectivity
  heuristic is relaxed for *derived* grouped parts, which are pinned by in-range + consistency, so
  small flat fins aren't mis-rejected). Render self-check: `sylpheed-cli mesh render DeltaSaber_T.xpr
  --only f001` (exact-name match excludes the `_rou_f001_mnv*` animation poses) → clean complete
  fighter. Test `hero_ship_grouped_pool_decodes`. Colours/UVs still pending the running-game oracle.
- 2026-07-18 (refinement) — **4-byte vertex-pool alignment + weapon recovery.** The grouped-pool
  rule "index pool ends exactly where the vertex pool begins" is really "the vertex pool is **4-byte
  aligned** after the index pool": `vb0 = align4(ib0 + span)`, so 0..=3 bytes of padding can sit
  between them. DeltaSaber's index pool ended already-aligned (pad 0), which hid this; **19
  weapon/`*_hangar` models** (single- and multi-marker: `wep_08/11/34/58/62/69/81/83…`) have pad 2
  and so decoded to *nothing* — the viewer then showed them as a flat 2D texture instead of a model.
  Fix: both anchors try `pad ∈ 0..=3` (`ib = vb − idx_bytes − pad` for the single-block adjacency
  anchor; `ib0 = vb0 − span − pad` for the grouped pivot), validated — a wrong pad reads shifted
  indices → agreement collapses < 0.85, so only the true pad passes. pad>0 in the ungated stage path
  is gated at a strict 0.85 to avoid a false anchor; pad 0 keeps its exact prior behaviour (stages
  unchanged). Result: all 19 now decode as clean models (e.g. `wep_34` 1243 v / 1233 t, a
  long-barrelled gun-pod; `wep_08` 3 sub-meshes / 478 t). Viewer routing already falls through
  `from_xpr2` → `anchor_models(0.85)` for single-XBG7 files, so the recovered grouped/padded weapons
  now preview as meshes.
- 2026-07-18 (refinement 2) — **pivot on the largest sub-mesh; all 19 recovered.** Three weapons
  (`wep_81`, `wep_81_hangar`, `wep_30_hangar`) still declined because the grouped pivot validated
  `markers[0]`, which for these is a tiny *elongated* lead bracket that fails the connectivity gate
  even when perfectly placed. Fixed by pivoting the alignment check on the **largest** marker (max
  index count) — the sub-mesh whose triangle-quality/connectivity signature most reliably confirms
  `(ib0, vb0)`. Once the pivot validates, markers up to it are read unconditionally (a legitimately
  tiny/flat lead part may fail the quality gates yet still be real), and markers after it stay
  validated so a stray trailing marker ends the chain. Result: **all 19 previously-declined weapons
  decode** (`wep_81` 460 t missile w/ tail fins; `wep_30_hangar` 334 t). DeltaSaber unchanged (its
  body IS the largest marker → same pivot). 7/7 disc tests green, stage quality audit unchanged.

## The declined set, measured (2026-08-11)

The module note said "a few multi-stream / quantized bodies remain" declined.
Measured across all 166 `hidden/resource3d/*.xpr`:

- **6 294 XBG7 resources, 5 480 decoded (87.1 %), 814 declined**, in **31 of 166**
  containers. Worst: `Stage_S09` 64/380, `Stage_S06` 58/324, `ptc_pack` 57/136.
- The declined set is **not** "a few hero bodies". By name prefix it is **492
  `e*`** (enemy craft), **142 `f*`**, **73 `n*`**, **23 `eff*`**, plus destroyed
  variants (`_rou_f402_dead`, `_rou_f302_base_dead`) and one weapon
  (`_rou_e011_wep04`).

### A shortcut that does not work

The resource descriptor's third word looked like a format/stream flag — decoded
`g001…g003` carry `0x00010001` while declined `t170`/`t180` carry `0x00020004`,
which reads temptingly as `(streams << 16) | format`. **It is not that.**
Histogramming it over the whole disc puts decoded *and* declined resources at
every value:

```
word[2]      decoded  declined
0x00010001      4479       328
0x00010002       126        27
0x00010003       120       112
0x00010004       142        84
…               …          …
```

Its low half runs 1…0x52 and tracks sub-mesh count, not vertex format. **So
decodability is not declared in the descriptor** — it is a property of whether
the unit-normal anchor scan can locate `vb0`, which is exactly what the current
code already tests. Anyone attacking this should not spend time on the
descriptor: 229 of the declined resources even carry the *most* common
`0x00010001` with under 1 KB of data, i.e. they are small meshes the scan has too
little signal to anchor, not exotic formats.

## Silent mis-decodes: a detector, and how many there are (2026-08-11)

The [declined set](#the-declined-set-measured-2026-08-11) is the *honest* failure
mode — 814 resources the decoder refuses. This is the other kind: geometry that
decodes without complaint and is wrong.

### The case that exposed it

`e303_wep_01` decodes from fourteen containers. In eleven it is a **49 × 23 × 42**
turret with organic vertices (`24.55, 0.00, 4.46` …). In `Stage_S02`, `S08` and
`S26` the *same* resource — **identical 172 vertices and 330 indices** — decodes
to **1600 × 2100 × 4800** of axis-aligned box corners:

```
Stage_S01   [   24.55    0.00    4.46]  [  24.55   9.84   2.91]   normals varied
Stage_S02   [   42.00 -900.00 2400.00]  [-600.00 -500.00 -500.00]  normals (0,0,1)
            [ -600.00 -500.00 -950.00]  [-600.00 -950.00 -950.00]  ← box face corners
```

The anchor scan located a **different buffer that happens to share the vertex and
index counts**, so every size-based check it makes passes. This is exactly the
"declined only for *location*, not format" risk the module notes describe — except
here it does not decline, it succeeds wrongly.

### The detector: cross-container bounds consistency

A resource shared by several containers must decode to the same bounds. That
needs no ground truth, and it measures the problem:

- **681 resources appear in ≥2 containers.**
- **125 of them decode to different bounds while reporting identical vertex and
  triangle counts** — a lower bound on silent mis-decodes (a resource wrong in
  *every* container is invisible to this test).

Examples: `_rou_f401` decodes as `62×25×10` in 16 containers and `4738×3147×4738`
in 2; `_rou_e011_wep05` produces **four** different spans across 8 containers.

### Repair candidate, and its limits

Taking the **majority span** across containers resolves **104 of the 125**; 14 are
exact 50/50 splits that a vote cannot decide. It agrees with the ground truth in
the one case that has independent evidence — `e303_wep_01`, where the 11-container
majority is the turret the render and the runtime capture both support.

🟡 **It is a heuristic and is otherwise unvalidated.** For
`_rou_e302_base_break` the majority is the *larger* span (`685×1206×1444`, 8 of
15) and nothing yet says which is right. Use the detector to flag; do not silently
rewrite geometry on a vote.

### ROOT CAUSE: the candidate list is container-global

Traced 2026-08-12. `anchor_pool_mesh` (the **per-block** path, which is the one
that handles single-sub-mesh resources like `e303_wep_01`) walks a candidate list
built by `vertex_run_starts(bytes, data_base, stride)` — **one scan of the whole
container per stride**, shared by every resource of that stride. It accepts the
**first** candidate that validates.

So a resource is anchored to *whatever block matches its signature first in file
order*, and nothing ties that block to the resource it belongs to. Two resources
sharing `(stride, vertex count, index count)` are interchangeable to this search.

**The wrong block is not distinguishable by quality.** Tracing the accept for
`e303_wep_01`:

```
Stage_S01   ACCEPT vb=4600480   pad=0  span=  49 ×  23 ×  42   passes 0.85 = true
Stage_S02   ACCEPT vb=18403456  pad=0  span=1600 × 2100 × 4800 passes 0.85 = true
```

Both clear the strict winding-consistency gate, because the wrongly-taken block
**is** real, coherent geometry — just another resource's. That rules out a whole
family of fixes: no threshold, no scoring, no "pick the best candidate" changes
this, and the earlier attempt to add best-of-N selection in the grouped-pool
anchor duly changed nothing.

**The search space has to be constrained instead — and the fix is now pinned
down.**

**The correct block is already in the candidate list.** Enumerating *every*
validating candidate for `e303_wep_01` in `Stage_S02` gives exactly two:

```
vb = 18 403 456   span 1600 × 2100 × 4800   ← what the decoder takes, only because it is first
vb = 52 257 440   span   49 ×   23 ×   42   ← correct: the same size all 11 good containers give
```

So nothing needs to be found that the scan is missing; the wrong one merely
appears earlier in file order.

**Locality picks the right one.** Recording each resource's accepted anchor in
descriptor order shows that global **monotonicity is refuted** — only 25 of 47
steps increase in `Stage_S01` and 130 of 248 in `Stage_S02`, i.e. no better than
chance. But *neighbourhood* holds strongly: in `Stage_S02` this resource's
descriptor neighbours anchor at **51 974 668** and **52 218 424**, its correct
candidate is **52 257 440**, and the block it wrongly takes is at **18 403 456** —
two thirds of the file away from its own family.

**Proposed rule:** among candidates that validate, prefer the one **nearest the
anchors of the neighbouring resources** (equivalently: decode in descriptor order
and prefer candidates close to the previous resource's anchor), falling back to
first-match when there is no neighbour yet. That needs no new format knowledge,
and it selects `52 257 440` here.

### ⚠️ WITHDRAWN (2026-08-12) — it halved inconsistency but regressed the twin mirror

`anchor_pool_mesh_near` tries candidates in order of distance from a reference,
and `anchor_models_filtered` runs **two passes**: pass 1 anchors first-match to
learn where resources land, then pass 2 re-anchors each resource preferring its
**neighbourhood** — the median anchor of its ±2 descriptor neighbours. A resource
with too few anchored neighbours keeps pass 1's result, so nothing regresses to
guesswork.

| | decoded | shared | inconsistent | e106 mirror |
|---|---|---|---|---|
| shipped (today) | 5 480 / 6 294 | 681 | **125** | ✅ matches capture |
| neighbourhood anchor | 5 480 / 6 294 | 681 | 63 | ❌ flipped |
| + refining the map | 5 480 / 6 294 | 681 | 51 | ❌ flipped |

**Refining matters** because pass 1's anchor map contains the very mistakes the
neighbourhood is meant to correct, so a resource beside a mis-anchored neighbour
inherits a bad reference. Re-anchoring against the improving map and repeating
converges quickly — two rounds, with a third changing nothing.

On its own metric this looked complete: coverage unchanged, inconsistency
halved, `e303_wep_01` decoding to 49 × 23 × 42 in *all* containers, and `e106`
rendering as a destroyer instead of a slab
([before](../captures/e106-static-assembly-volume-bug.png) ·
[after](../captures/e106-static-assembly-fixed.png)).

**It was reverted anyway.** `ship::tests::static_assembly_matches_runtime_capture`
is gated on `SYLPHEED_ISO` and therefore skips in an ordinary `cargo test`; run
with the ISO it fails:

```
e106_bdy_01: static M row0 [-1.0, 0.0, 0.0] != captured [1.0, 0.0, 0.0]
```

**Why:** `e106_bdy_01` and `e106_bdy_02` are a mirrored pair whose two vertex
buffers hold the same geometry reflected in X, and **both resources currently
decode to the *same* buffer** — identical vertex count, identical span, identical
`mean_x`. `apply_twin_mirrors` decides which instance to reflect from the sign of
that `mean_x`, so which of the two buffers gets picked flips the decision:

```
before the change   both twins decode with mean_x = −66.83  → mirror bdy_02  (matches the capture)
after  the change   both twins decode with mean_x = +66.83  → mirror bdy_01  (contradicts it)
```

The runtime capture is ground truth, so a change that contradicts it does not
ship.

**Correction (measured after the fact):** the first write-up of this said "two
distinct resources sharing one decode is itself the bug". **That is wrong.**
Sharing is normal here — 1 043 of 5 480 decoded resources (19 %) share geometry
with another resource, and of 1 242 related pairs, **1 241 are identical in every
container they co-occur in**, which is what legitimate asset reuse looks like.
A mirrored pair like `bdy_01`/`bdy_02` is *supposed* to share one geometry, with
the reflection applied at placement — exactly what `apply_twin_mirrors` does.

What actually matters is **which of two mirrored buffers is canonical**. The disc
holds both an X+ and an X− version; the engine treats one as the base, and
`apply_twin_mirrors` was tuned against that. The neighbourhood anchor moved these
resources to the *nearer* buffer, which is the other one — hence the flip. So the
real fix is not "give each twin its own buffer" but **pin which buffer is
canonical**, with the capture as the oracle.

**Exactly one pair is provably mis-anchored by this test**: `e105_bdy_02_l` /
`e105_brg_m` share a decode in 7 of the 15 containers holding both and differ in
the rest — two names cannot be the same geometry only sometimes.

**The filtered path needed care.** `models_named` (what the viewer's ship
rendering uses) drops non-wanted resources, which would leave a filtered decode
with no neighbourhood at all — and silently keep the old behaviour. Resources are
now collected regardless of the filter, but only the asked-for ones and their ±2
neighbours are decoded in pass 1, so a filtered decode stays proportional to what
was asked for.

**51 remain**, and they cluster in `_l` (LOD) and `_dead` variants — `e001_l`,
`e010_bdy_01_l`, `e011_bdy_01_l`, `e016_l`, `e104_bdy_05_l`, `e106_eng_02_l`,
`e501_01_l`, `_rou_f301_base_dead`, `_rou_f302_base_dead`, `e303_base_dead`.

**A tempting explanation, tested and false.** The obvious reading is that a
variant shares its base's vertex and index counts, so the two are mutually
confusable *and* adjacent, defeating locality. Checked across every container:
**2 714 variant/base pairs, and exactly zero share identical counts.**

**What is actually happening: one region is a universal false positive.**
`e010_bdy_01_l` is 171 verts / 90 tris and **no other resource in its container
shares those counts** — yet in `Stage_S02` and `S26` it decodes to
**1600 × 2100 × 4800**, the *same* bounds `e303_wep_01` (172 verts / 110 tris)
produced before the fix. Differently-shaped resources are landing on the same
place. So the attractor is not "another mesh with my shape" but a region of
**round, axis-aligned box data** that validates for many different `(vtx, idx)`
shapes at once — every index lands in range and the triangles are coherent boxes.

That also explains why the neighbourhood fix helped so broadly: it steers
resources away from a single strong attractor rather than resolving many
pairwise confusions.

**Selection-based fixes are exhausted — tested.** The proposed tiebreak was
implemented as a *last resort* (accept the attractor only if nothing else
validates), first keyed on "all coordinates multiples of 50" and then on the
sharper **"all coordinates integral"** — the attractor reads `(42, −900, 2400)`,
`(−600, −500, −950)` while real geometry carries fractions like
`(24.55, 9.84, 4.46)`. **Neither changed anything: still 51.**

That null result is itself the answer. A mechanism that defers the attractor
whenever another candidate exists, changing nothing, means **no alternative
candidate validates for any of the 51** — the correct block is *not in the
candidate list at all*. Both attempts were reverted rather than kept.

So the residual is **not** a selection problem, and no reordering, scoring or
tiebreak will move it. The frontier is `vertex_run_starts` — the unit-normal
run scan that builds the candidate list — which does not emit a start for these
resources' real vertex buffers. That is where the remaining 51 live.

The ignored test
[`mesh_consistency_disc.rs`](../../crates/sylpheed-formats/tests/mesh_consistency_disc.rs)
still asserts the target state and now records 63 rather than 125; the remaining
cases are where the neighbourhood is itself wrong or absent.

### Where the mis-decode is *not*: the grouped-pool anchor

An attempt to fix it by making `anchor_grouped_meshes` choose the **best-scoring**
`vb0` (rather than the first candidate clearing the 0.85 gate) changed **nothing**
— still 5 480 of 6 294 decoded and still 125 inconsistent — and instrumenting the
pivot loop shows why: for `e303_wep_01` it never runs. **The resource has a single
sub-mesh, so it is decoded by the per-block adjacency path
(`anchor_pool_mesh`), not the grouped-pool anchor.**

So the silent mis-decode lives in the **per-block** anchor. That is worth knowing
before anyone else spends time on the grouped-pool pivot, which is the more
prominent and better-documented of the two and the natural first suspect.

The change was reverted: it was untargeted, unproven, and added a scoring path
with no demonstrated benefit. (Its one reusable idea — that several `vb0`
candidates can clear the gate and first-in-scan-order is an arbitrary tiebreak —
still applies to whichever anchor turns out to be at fault.)

## Two follow-ups on the anchor (2026-08-12)

**The descriptor does not address the geometry.** If it did, the whole
candidate-scan could be replaced by direct addressing. It cannot: across `e106`'s
resources in `Stage_S01`, `anchored_vb − entry.data_offset` ranges from
**1 199 052 to 4 173 988** with no constant or stride. `data_offset` locates the
*descriptor*, and nothing in the first six descriptor words tracks the vertex
pool. The scan is necessary.

**Cross-container agreement does not prove legitimate reuse.** The earlier
measurement — 1 241 of 1 242 related pairs identical in every container — was
read as "sharing is normal". It is weaker than that: a **systematic** error is
invisible to a consistency test, because it is consistent. The same dump shows
`e106_bdy_02` and `e106_bdy_03_m` anchoring to the identical offset
(`1 505 556`), and `e106_bdy_01_l` with `e106_brg_01_m` (`4 251 208`) — a hull
half and a *different* body's medium LOD, or a hull half and a bridge LOD. Those
are different parts; one of each pair must be wrong.

So the honest position is: **sharing is common (19 %), some of it is certainly
legitimate (a mirrored twin pair genuinely shares one geometry), and some is
certainly not** — and cross-container consistency cannot tell them apart. A test
that can: compare a shared pair against a runtime capture, which is ground truth
for what the engine actually draws.

### The capture answers it at population level — and the logs are still on disc

`/sylph-home/re/shipcap/xenia_ship_capture_*.log` (kept from the 2026-07 capture
sessions) carry the raw per-draw lines the baked table was distilled from:

```
DRAW vbase=0x150CCAC0 stride=28 vcount=1 indices=1 prim=1 vs=0x…
DRAW vbase=0x150CCAC0 stride=24 vcount=10891 indices=18 prim=4 vs=0x…
```

`vbase` is the GPU vertex base — **ground truth for which buffer the engine draws
a part from**, which is exactly the oracle the sharing question needs. Over the
three logs, restricted to the ship-geometry stride 24:

- **6 093 draws from 2 291 distinct vbases**
- **only 77 vbases (3.4 %) are drawn more than once**

⚠️ **That 3.4 % measures less than it first appears — corrected 2026-08-12.**
The capture code (`command_processor.cc`, `CaptureShipDrawForRE`) de-duplicates
by **(vbase, WVP-transform hash)**, so a buffer drawn many times *at one
transform* — which is what a mesh split into per-material sub-draws looks like —
appears **once**. The figure therefore counts buffers drawn at *several
placements* (multi-instance parts), not buffers serving several parts. It is not
the population-level evidence about sharing it was first written up as; the
`bdy_01_l`/`bdy_02_l` result below is direct evidence and stands on its own.

Two more field semantics, read off the same patch rather than guessed:
`vcount = fetch.size × 4 / stride` is the **buffer's capacity**, not the draw's
vertex usage (which is why it matches a decoded resource's vertex count so
exactly), and `indices` is `VGT_DRAW_INITIATOR.num_indices`, **that draw's**
index count. So the 119-vertex twin logging `indices=21` against our 246-index
marker most likely means the engine issues the mesh as several sub-range draws
and the log keeps the first — likely, not proven.

**How to use it per-part:** `correlate_capture` already matches a draw to a
resource by `vcount` plus decoded positions. The same match yields, for each
part, the `vbase` the engine used — so two resources that our decoder gives the
same geometry can be checked directly: different `vbase` in the capture ⇒ our
shared decode is wrong. That is the per-part oracle any future anchor work should
be validated against, and it needs no new capture run.
### ✅ It was run — and the capture gives file-offset ground truth

`examples/shared_vbase_check.rs` does the per-part check above, and then goes one
step further than planned. Three results, in order of strength.

**1. A draw's `vbase` *is* the container file offset plus a constant.** Vertex
POSITION is `f32×3` big-endian at vertex offset 0, so a draw's dumped positions
are a value pattern that can be searched for in the `.xpr` itself. Doing that for
every draw in `xenia_ship_capture_01/02.log` and histogramming `vbase − offset`:

```
xenia_ship_capture_01.log: 11 distinct vbases located, 208 not in this container
   vbase - offset = 0x1A94FFF4  ×8      ← same constant in log 02
```

The 208 "not in this container" are draws whose geometry lives in `Common.xpr`,
a weapon pack or a backdrop — expected. The eight that do belong to `Stage_S01`
share **one** constant, and the *same* constant in a second run, so the container
is uploaded contiguously and **a capture names the exact file offset of every
buffer the engine drew**. Log 03 loaded the container at a different address, so
the constant is per-run, not baked.

**2. Read against our anchor scan, that is a defect list.** `GameMesh` now
carries `vbuf_offset` — the offset the anchor scan actually placed a sub-mesh at
— so the comparison is exact
([`captures/stage-s01-capture-truth-offsets.txt`](../captures/stage-s01-capture-truth-offsets.txt)):

| drawn offset | vcount | our resource anchored there | our resources with that vcount |
|---|---|---|---|
| `0x3b3ee8` | 119 | `e106_bdy_01_l`, `e106_bdy_02_l` | `bdy_01_l`, `bdy_02_l` |
| `0x3c55d8` | 119 | — **nobody** | `bdy_01_l`, `bdy_02_l` |
| `0x3dd2c4` | 146 | `e106_bdy_03_l` ✅ | `bdy_03_l` |
| `0x40763c` | 179 | `e106_bdy_04_l` ✅ | `bdy_04_l` |
| `0x40e418` | 51 | — **nobody** (ours sit `0x5d0` earlier) | `brg_01_b_02`, `brg_01_l` |
| `0x444ccc` | 58 | `e106_eng_01_l` ✅ | `eng_01_l` |
| `0x44a32c` | 44 | — **nobody** | `eng_02_l` |
| `0x45705c` | 82 | `e106_wep_02_01_l` ✅ | `wep_02_01_l` |
| `0x38788` `0x4b8b8` `0xb6574` `0xdbbac` `0x133da0` `0x162840` | 181, 93, 41, 77, 76, 60 | — nobody | mostly none (other objects in the stage) |

Four of the ship's drawn buffers are anchored exactly right. Two are the twin
collapse below. **Two are mis-anchors of a size we do have**: the engine's
51-vertex bridge buffer is at `0x40e418` while both our 51-vertex bridge
resources sit at `0x40de48`, and its 44-vertex `eng_02_l` is at `0x44a32c` while
ours is elsewhere entirely. The remaining six belong to other objects in the
stage (`n041`, `n042`, `e303`), only two of which we decode at the right size.

> **A trap worth recording.** The first version of this table located our
> resources by *searching the container for their leading vertices* instead of
> asking the decoder, and it read much worse — full and `_m` resources appearing
> to start inside their own `_l` buffer. That was an artifact: **the same leading
> vertex run occurs at several offsets in one container** (`e106_bdy_03`'s first
> eight positions occur at four, `bdy_01`'s at three). That multiplicity is
> itself the reason the anchor scan is ambiguous — but it makes a position search
> useless for asking where a resource *was* anchored. Hence `vbuf_offset`.

**3. The twin pair is an anchoring error, and the mirror is in the data.** For
`e106_bdy_01_l` ≡ `e106_bdy_02_l` the capture shows **two** 119-vertex buffers
per run, `0x3b3ee8` and `0x3c55d8`; the first is byte-for-byte what we decode,
and the second is its **exact X-reflection** (every dumped position matches ours
with `x` negated). So the container carries both halves as separate baked
geometry, the engine draws each from its own buffer, and our decoder returning
one buffer for both names is the defect — which `correlate`'s mirror flag and
`ship::apply_twin_mirrors` have been compensating for downstream all along.

That settles the question this section opened with, for this pair: **not
legitimate reuse.** It also pins what the withdrawn neighbourhood-anchor fix
could not: `0x3b3ee8` stays with whichever twin we already decode there, and the
other twin must move to `0x3c55d8`. The invariant is checkable without a capture
— *mirrored twins must decode to X-reflected buffers, never identical ones*.

**4. Root cause, for these three: selection, not the run scan.**
`mesh::debug_vertex_run_starts` exposes the candidate list the anchor scan works
from. `Stage_S01` yields **15 710** stride-24 candidate starts, and **all three
capture-proven offsets are in it** — `0x3c55d8` (the mirrored twin), `0x40e418`
(the drawn bridge buffer) and `0x44a32c` (`eng_02_l`). The scan sees the right
offsets; `anchor_pool_mesh` walks the list in ascending order and takes the first
that validates, so an earlier lookalike wins — our bridge resources sit `0x5d0`
before the buffer the engine drew.

This is scoped: it says the *current* decoder's e106 mis-anchors are selection
failures. It does not overturn the earlier finding that the residual 51 *under
the withdrawn neighbourhood fix* had no validating candidate at all — a different
population, and the two can both be true.

What the twins suggest as the fix: selection is **per-resource and greedy**, so
two resources can and do claim one buffer while a validating buffer sits unused.
An assignment that is distinct by construction — each candidate used at most once
— resolves the twin case by shape rather than by heuristic. Whether the proven
offsets actually validate for their resources is the next thing to test; if they
do, distinctness alone is the fix.

**5. Do the proven offsets validate? Two of three — and that splits the fix.**
`mesh::debug_try_anchor(bytes, name, vb, max_pad)` asks `validate_block` directly
(`examples/try_anchor.rs`):

| resource | proven offset | verdict |
|---|---|---|
| `e106_bdy_01_l` / `e106_bdy_02_l` | `0x3b3ee8` **and** `0x3c55d8` | **accepted for both, at both** (v=119, idx=246, pad=0) |
| `e106_brg_01_l` / `e106_brg_01_b_02` | `0x40e418` | **accepted for both** (v=51, idx=126, pad=0) |
| `e106_eng_02_l` | `0x44a32c` | **rejected** — and still rejected with the pad widened to 64 |

So the twin case is exactly what it looked like: the correct block is perfectly
acceptable and simply lost the first-match race, and a **distinct assignment**
(each candidate buffer claimed by at most one resource) fixes it — the two
resources have two accepted offsets between them. The bridge pair is weaker:
`0x40e418` is accepted by both, our current `0x40de48` is accepted too, so
distinctness would separate them but not choose correctly.

`eng_02_l` is a different failure: the offset the engine drew from is **not
acceptable at all**, so no selection policy can reach it. That is the
"residual" class this file describes above, now with one member pinned to a
concrete offset for the first time.

❔ **An open discrepancy, recorded not explained.** The capture's `DRAW` lines
carry an `indices=` field that does not agree with the descriptor's index count:
the 119-vertex twin draws log `indices=21` where our marker says 246, and the
44-vertex draw logs `indices=12`. Whether that field is an index *count* of a
sub-range, a different unit, or a Xenia-side artifact is unknown — it may matter
for `eng_02_l`, whose block validation is exactly what an index-count mismatch
would break.

**6. Why `eng_02_l`'s real block is rejected: the connectivity heuristic.**
`mesh::debug_find_index_buffer` scans the *whole container* for an index buffer
that validates against a known vertex buffer, instead of assuming adjacency
(`examples/find_ib.rs`). For `e106_eng_02_l` at the capture-proven `0x44a32c`,
with the connectivity test on, **nothing in the container validates**. With it
off (`SOFT_IB=1`) the nearest hit is `ib 0x44a29c` — `vb − ib = 144 = 72 × 2`,
i.e. **exact pad-0 adjacency**. So the index buffer is exactly where the decoder
assumes it is; the block is thrown out by one heuristic.

That heuristic is `mean_edge / bbox_diag > 0.28 → reject`. Measured on the real
blocks:

| block | mean edge | bbox diagonal | ratio | verdict |
|---|---|---|---|---|
| `eng_02_l` (24 tris) `ib 0x44a29c → vb 0x44a32c` | 109.21 | 261.96 | **0.417** | rejected (cap 0.28) |
| `bdy_02_l` (82 tris) `ib 0x3c53ec → vb 0x3c55d8` | 131.70 | 786.66 | 0.167 | passes |
| `bdy_01_l` (82 tris) `ib 0x3b3cfc → vb 0x3b3ee8` | 131.70 | 786.66 | 0.167 | passes |

This is precisely the false positive the check's own comment predicts — "a small
flat sub-mesh legitimately has large edges relative to its own diagonal" — caught
in the wild for the first time, with the runtime naming the block it rejects. A
24-triangle engine LOD is coarse by construction, so its edges *are* a large
fraction of its size.

(The twins' two real blocks having **identical** mean edge and diagonal is a free
corroboration that they are mirror images: reflection preserves lengths.)

So the residual class is not one bug. `eng_02_l` has its vertex start in the
candidate list *and* its index buffer exactly adjacent, and still fails — a
**validator** problem, not a scan or selection one. Raising the cap is not the
fix to reach for blind: the threshold trades against false anchors, and now that
a capture can name true blocks, it can be **calibrated** against them rather than
guessed. Not changed here.

**7. Calibrating the cap: a real trade, not a free win.** `XBG7_EDGE_CAP`
(and `XBG7_SMALL_TRIS` / `XBG7_EDGE_CAP_SMALL` for a triangle-count-aware
variant) make the threshold sweepable without changing the default;
`examples/edge_cap_sweep.rs` reports coverage and cross-container consistency per
setting. Over the whole `resource3d` directory:

| cap | resources decoded | anchors moved vs 0.28 | shared | inconsistent | consistent → **inconsistent** | inconsistent → consistent |
|---|---|---|---|---|---|---|
| **0.28** (shipped) | 5 480 | — | 678 | 125 | — | — |
| 0.35 | 5 955 | — | 707 | 140 | — | — |
| 0.42 | 6 069 | 254 | 714 | 153 | **18** | 3 |
| 0.45 | 6 093 | 254 | 716 | 153 | 18 | 3 |
| 0.28, but 0.45 below 64 tris | 6 090 | 236 | 716 | 151 | **16** | 3 |

**Nothing is ever lost** — every resource that decoded at 0.28 still decodes —
and the capture-proven case is fixed: at any cap above 0.417, `e106_eng_02_l`
anchors at exactly `0x44a32c`, and the four e106 parts that were already correct
stay correct. So on the only ground truth available, relaxing is a strict
improvement (4 → 5 of the ship's drawn buffers correct).

But it is bought: ~590–610 resources that previously decoded not at all now do,
**~240–254 existing anchors silently move**, and 16–18 shared resources go from
cross-container consistent to inconsistent (against 3 repaired). The
triangle-aware variant barely narrows that — almost everything the looser cap
admits is a small block anyway.

**So the cap is not changed here.** The evidence says 0.28 is too tight and that
mean-edge-over-diagonal is a weak discriminator for coarse LODs; it does not say
0.45 is right, because the 254 movers have no oracle. Deciding needs either a
capture covering more ships and stages (the same `--truth` method extends to any
container the engine drew from) or a discriminator that does not degrade for
coarse geometry. Both are recorded as the next step rather than guessed at.

**8. Provenance correction — the capture is `Stage_S02`, and the oracle is
4× bigger than reported.** `examples/capture_truth_scan.rs` runs the
offset-locating pass over **every** container in `resource3d` and reports each
one's modal `vbase − offset`. `Stage_S02` places **64** buffer-matches at a
single constant (`0x17FE3FF4`); `Stage_S01`, which sections 1–7 above used,
places only 16. The logs also contain `f101` (the ACROPOLIS escort, 25 448
verts), `f105`/`f106` (TCAF cruiser and destroyer) and `e105` — a Stage-02 cast.
**The loaded container was `Stage_S02`.**

Why `Stage_S01` nevertheless produced a consistent constant: the two containers
carry the shared block **verbatim and contiguously**. The e106 twin buffers sit
at `0x3b3ee8`/`0x3c55d8` in `Stage_S01` and `0x2d1fee8`/`0x2d315d8` in
`Stage_S02` — the same `0x116F0` apart. So the earlier findings are still true
statements about `Stage_S01`'s content (both mirrored buffers are in it, and our
decoder collapses them), and they reproduce in `Stage_S02`; only the claim that
`0x1A94FFF4` was *the loaded container's* base was an artifact.

The `Stage_S02` table
([`captures/stage-s02-capture-truth-offsets.txt`](../captures/stage-s02-capture-truth-offsets.txt))
places **46 drawn buffers**: 34 have a claimant (29 of them a single resource of
exactly the drawn size), **12 are claimed by nobody**. It also resolves the six
mystery buffers of the `Stage_S01` table — 181, 93, 77, 76, 60, 41 vertices —
as **`e105` parts** (`bdy_04_l`, `bdy_05_l`, `brg_l`, `eng_01_l`, `wep_01_l`),
another ship in the same mission, several of which are the same "right size,
wrong place" failure as `eng_02_l`.

❔ **A new defect class to chase**: `_rou_f105_break` — a destruction composite —
claims a run of **twelve consecutive** drawn offsets whose sizes match *other*
ships' LODs (`f105_bdy_01_m`, `f105_bdy_03_m`, `e106_wep_02_01_l`, …). Either the
break model genuinely mirrors those buffers or our decoder is handing a whole
address range to one composite. Not resolved here.

**9. The cap, scored against ground truth — and what it reveals about the
selection.** Re-running the sweep against the 46 capture-named `Stage_S02`
buffers (the oracle section 7 lacked):

| cap | exact anchors | claimed by several | unclaimed | vs 0.28 |
|---|---|---|---|---|
| **0.28** | 29 | 4 | 12 | — |
| 0.35 | 30 | 3 | 12 | +1 exact, 0 lost |
| 0.42 | **31** | 4 | **10** | +2 exact, **0 lost** |
| 0.45 | 31 | 4 | 10 | +2 exact, 0 lost |

So against the runtime, relaxing is **monotone**: two more buffers get exactly
the right resource and **no previously-correct anchor is lost**. That is the
opposite reading from the cross-container consistency metric, which showed 18
regressions — and the two measures can be reconciled by looking at what those 18
actually are.

Of the 18, only two touch a drawn buffer, and one is decisive: the 32-vertex
buffer at `0x24ce5f4` is **unclaimed at 0.28** and at 0.42 is claimed by **both**
`f303_body_l` and `e302_barrel_l`, two 32-vertex resources. That is not a
mis-placement — it is a **collision**. The looser cap admits the true block *and*
lets a second resource grab it, because selection is per-resource and greedy.

**Conclusion: the cap and the selection have to move together.** Relaxing alone
buys real anchors and pays in collisions; distinct assignment alone (section 5)
fixes the twins but cannot reach blocks the validator still rejects. The pair of
changes is the fix; either alone is a half-measure, which is why neither has been
made yet.

### ✅ Fixed: distinct anchor assignment (2026-08-12)

The selection is no longer per-resource-greedy. After the parallel decode, the
models are walked in container order: the first claimant keeps a buffer, and any
later resource that chose the same one is **re-anchored past every buffer already
claimed**. A resource that finds no free candidate keeps its collided decode, so
coverage can never regress. Grouped-pool models are untouched.

Measured against the 46 capture-named `Stage_S02` buffers, and disc-wide:

| | exact anchors | unclaimed | resources decoded | shared inconsistent |
|---|---|---|---|---|
| greedy, cap 0.28 (before) | 29 | 12 | 5 480 | 125 |
| **distinct, cap 0.28 (now)** | **40** | **4** | 5 480 | **46** |
| distinct, cap 0.42 | 45 | 0 | 6 069 | 62 |

Nothing decodes that did not decode before, cross-container inconsistency drops
by **63 %**, and 11 more of the ship's drawn buffers get exactly the right
resource. The cap stays at its shipped 0.28 — that is a separate change with its
own evidence (section 9), and this one is worth being able to revert alone.

**One convention changed, and it is the point of the fix.** With the twins
sharing a buffer, the reflection had to be synthesised downstream:
`ship::apply_twin_mirrors` flipped one hull, and `correlate` baked an X-flip into
`e106_bdy_02`'s captured matrix (`diag(-1,1,1)`). Now each twin decodes to its
own, already-mirrored buffer, so **the mirror lives in the data** and both
placements are proper rotations. Re-emitting the block from the capture confirms
it independently (`1 0 0 0 1 0 0 0 1 264.04343 …`). The embedded placement row
and the two assertions that encoded the old convention were updated, each with
the reason in place; the full suite including the disc- and ISO-gated ship tests
is green.

### ✅ Fixed: the connectivity cap is 0.42 (2026-08-12)

With distinct assignment in place the cap was swept again against the 46
capture-named buffers. It saturates:

| cap | exact anchors | unclaimed | resources decoded | shared inconsistent |
|---|---|---|---|---|
| 0.28 (old) | 40 | 4 | 5 480 | 46 |
| 0.35 | 42 | 3 | 5 955 | 56 |
| **0.42 (now)** | **45** | **0** | 6 069 | 62 |
| 0.45 | 45 | 0 | 6 093 | 62 |
| 0.60 | 45 | 0 | — | — |

Nothing above 0.42 anchors anything more, so **0.42 is the least permissive value
that captures the whole measured gain** — 45 of the ship's 46 drawn buffers get
exactly the right resource, none is left unclaimed, and 589 more resources decode
than at 0.28 with nothing lost.

The one metric that worsens is cross-container consistency (46 → 62). That is the
*proxy*, and this file already documents why it is the weaker witness: a
systematic mis-anchor is invisible to it because it is consistent. Where the two
disagree, the capture wins. Full suite green including the disc- and ISO-gated
ship tests.

✅ **Closed — the grouped composite is a false alarm.** `_rou_f105_break` (the
f105 destruction model) claims **twelve consecutive** drawn buffers while
`f105_bdy_01_m`, `f105_bdy_03_m` and `f105_eng_01_m` sit elsewhere, which looked
like the twins' bug in the grouped-pool path. It is not.
`examples/locate_draw.rs` counts how many copies of a captured buffer a container
holds, directly and X-mirrored:

| drawn buffer | vcount | direct copies | mirrored copies |
|---|---|---|---|
| `0x1bf596c` | 2446 | **3** | 0 |
| `0x1c0fe2c` | 2336 | **6** | **6** |
| `0x1c211cc` | 1402 | **4** | 0 |
| `0x1bf40f4` | 261 | 1 | 0 |

The container stores these parts several times over, and every live LOD checked
(`f105_bdy_01_m` at `0x1ecb4c0`, `f105_bdy_03_m` at `0x1f902a4`, `f105_eng_01_m`
at `0x1ffb884`) is anchored on a **direct** copy — byte-identical geometry to the
one the engine drew. So the composite taking "the drawn" copy costs nothing: the
decode is the same vertices either way. (The draws use the ordinary ship shader
`0xEEA84C59D7F95371`, the same one as the validated e106 hull draws, so these are
intact-ship draws, not debris.)

⚠️ **This also calibrates the oracle metric.** "Exact" in the tables above means
*anchored at the offset the engine drew from*, which is stricter than correct: a
resource anchored on an identical copy is equally right. The 45/46 figure stands
as a lower bound, and an "unclaimed" row is not automatically a defect.

❔ What the copy counts *do* leave open: with six direct and six mirrored copies
of one buffer, a resource landing on a **mirrored** copy would be a real defect
and would look identical to a correct decode in every count-based metric. Only a
capture (or the twin-pair invariant) can catch it.

### Distinct assignment extended to grouped pools (2026-08-12)

`anchor_grouped_meshes` now takes the same `taken` set: a pool whose start another
resource already claimed is skipped, and a grouped model that collides is
**re-placed whole** past everything claimed (or keeps what it had, so coverage
cannot regress).

| | oracle exact / 46 | unclaimed | resources decoded | shared inconsistent |
|---|---|---|---|---|
| single-mesh distinctness only | 45 | 0 | 6 069 | 62 |
| **+ grouped pools (now)** | 45 | 0 | 6 069 | **56** |

The runtime oracle is unchanged and cross-container inconsistency falls another
10 %. The suite, including the disc- and ISO-gated tests, stays green.

**It clears the `n206` collapse too** — a correction to what was written here
first. `n206_02` now anchors at `0x342d984` instead of sharing `0x33b6e54` with
its twin. The earlier "no alternative pool validates" reading was wrong twice
over: `mesh::debug_grouped_report` (`examples/why_rejected.rs`) shows that pool
**ACCEPTED at pad 0** under the production gates, and `0x342d984` is in the
candidate start list. What actually misled the check was the audit itself —
it classified twins by decoded **geometry**, and `0x342d984` is a *direct*
(unmirrored) copy of `0x33b6e54`, so the pair still looked "identical" after it
had been separated.

The audit now distinguishes the two: identical geometry is only a collapse when
it comes from **one buffer**. Re-run disc-wide, of 34 equal-count twin pairs:
**18 exact X-mirror, 16 related another way, 0 identical-sharing-a-buffer, 0
unrelated** — and the regression test no longer needs its `n206` exception.

❔ Left open: whether `n206_01`/`n206_02` *should* be a mirrored pair at all. The
container holds two direct copies **and** two mirrored ones
(`0x33b7754`, `0x342e284`); our twins take the two direct copies, which is
self-consistent but unverified — `n206` appears in no captured stage.

### ✅ Fixed: grouped pools emitted sub-meshes with out-of-range indices

Chasing whether the "buffer not covered" gate is well founded turned up a real
defect instead. `examples/coverage_audit.rs` measures, for every decoded
sub-mesh, how many tail vertices its indices never reference:

| unreferenced tail vertices | sub-meshes |
|---|---|
| **0** (indices reach the last vertex exactly) | **8 586** |
| 1–3 (inside the gate's ±4 tolerance) | 48 |
| 9, 80 (`f102_break.dat`, `f104_break.dat` in `ptc_pack.xpr`) | 2 |
| **negative — indices point PAST the vertex buffer** | **18** |

The first row answers the original question: real geometry covers its pool
**exactly**, so under-coverage is good evidence of a wrong candidate and the gate
stands as written.

The last row is the defect. `anchor_grouped_meshes` reads sub-meshes *before* the
pivot unconditionally — deliberately, since a tiny flat lead part legitimately
fails the quality gates — but that also skipped the **index-range** check, which
is not a quality question. Eighteen sub-meshes disc-wide were emitted with
indices reaching up to **364 vertices past the end** of their own buffer, which
any renderer would fault on or draw as garbage. Pre-pivot sub-meshes are now
required to be in range (quality gates still relaxed); out-of-range ones are
dropped and the pool's remaining parts are kept.

Everything else holds: 6 069/6 294 decoded, cross-container inconsistency 56, the
capture truth table still 46/46 claimed with 0 unclaimed, suite green. The vertex
total falls by 1 546 — exactly the garbage that is no longer emitted.

**The same reasoning finished the job.** The two remaining outliers were the only
decoded blocks that did not cover their pool: `f102_break.dat` had a sub-mesh
declared 414 vertices whose indices stopped at 404, and `f104_break.dat` one
declared 160 whose indices stopped at 79 — both reading a *neighbouring* block's
index buffer against the wrong declaration. These are grouped `.dat` composites
in `ptc_pack.xpr` whose marker lists (9 and 10 entries) clearly do not map 1:1
onto the stored blocks; only 2 of 9 and 4 of 10 sub-meshes ever decoded. Applying
the **coverage** requirement to pre-pivot sub-meshes as well drops exactly those
two mismatched pieces and keeps the rest.

Every decoded sub-mesh on the disc now covers its own vertex pool: **8 580 at
slack 0, 49 within the ±4 tolerance, none beyond it, none negative.** Coverage,
consistency and the capture oracle are all unchanged by the tightening.

### ✅ Fixed: winding consistency replaces the connectivity heuristic

With every miss attributed (below), **connectivity accounted for 153 of 225** —
by far the largest blocker, and the one gate already known to reject a
capture-proven block. So it was tested against the alternative the decoder
already trusts elsewhere: **winding consistency**, `max(na, 1−na)`, where a real
mesh sits at ≈1.0 or ≈0.0 and a mis-carve lands near 0.5. It is an objective
topology test; the edge-ratio is a shape heuristic.

Swapping them (connectivity inert, winding gating the pad-0 path):

| winding floor | resources decoded | shared inconsistent | capture oracle |
|---|---|---|---|
| — (connectivity 0.42, previous default) | 6 069 | 56 | 46/46 |
| 0.60 | 6 214 | 53 | — |
| **0.70 (now)** | **6 212** | **39** | **46/46** |
| 0.80 | 5 770 | 1 | — |
| 0.85 | 5 770 | 0 | 46/46 |

**0.70 dominates the previous default on both axes** — 143 more resources decode
*and* 17 fewer shared resources disagree across containers — with the capture
oracle unchanged at 46/46 claimed, 0 unclaimed, the twin invariant still clean
(18 exact mirrors, 22 related, 0 collapses, 0 unrelated across 40 pairs, up from
34), and every decoded sub-mesh still covering its pool. So it ships, and the
connectivity cap drops to an inert 1.0, kept as a knob and a backstop.

The cliff at 0.80 is recorded rather than taken: it buys **perfect** cross-container
consistency (0 inconsistent) for 442 resources. Consistency is the weaker witness
— a systematic mis-anchor is consistent — so paying that much coverage for it is
not obviously right, and both points are one env var apart
(`XBG7_PAD0_CONSISTENCY`, `XBG7_EDGE_CAP`) for anyone who wants the conservative
end.

### ✅ Fixed by *looking* at the output: exact pool coverage (2026-08-12)

Rendering the assembled `e106` from `Stage_S02` — something no metric had done —
showed the old slab back: `e106_bdy_03` spanning **600×1600×998**, a blocky mass
beside the hull. The same resource decodes to **276×236×941** in `Stage_S01`,
`Stage_S03` and `Stage_S04`. The tell was already in the data:

| container | anchor | slack | span |
|---|---|---|---|
| `Stage_S01` | `0x3c9b7c` | 0 | 276×236×941 |
| **`Stage_S02`** | `0x12d1d44` | **3** | **600×1600×998** |
| `Stage_S03` | `0x1e07b7c` | 0 | 276×236×941 |
| `Stage_S04` | `0x164337c` | 0 | 276×236×941 |
| `Stage_S06` | `0x1f05298` | 0 | 414×636×1121 |

The coverage gate tolerated up to **three** unreferenced tail vertices, and that
tolerance was hiding a mis-anchor: real blocks reach their pool's last vertex
exactly (8 580 of 8 629). Requiring exact coverage moves `Stage_S02`'s
`e106_bdy_03` to `0x2d35b7c`, slack 0, **276×236×941** — the block three other
containers agree on — and the slab disappears from the render
([`captures/e106-cover-slack-before-after.png`](../captures/e106-cover-slack-before-after.png)).

Cost: **3** resources disc-wide (6 212 → 6 209); cross-container inconsistency
39 → 38; capture oracle unchanged at 46/46 claimed, 0 unclaimed; suite green.

(`Stage_S06` still gives a third answer at slack 0, so `e106_bdy_03` is not fully
settled — but it is now consistent across four of the six containers that carry
it instead of three.)

### The other capital ships are clean — and an attempt to automate the slab check failed

`e106` was the first ship rendered, and it had the defect. The other four in the
`Stage_S02` cast — `f101` (ACROPOLIS), `f105`, `f106`, `e105` — were rendered the
same way and all assemble into coherent hulls with no stray masses.

Automating the check did **not** work. `examples/slab_screen.rs` compares each
base part against its ship's median extent (min-axis, so a legitimately long thin
antenna does not swamp the test) and flags outliers. Run against the decoder
**before and after** the coverage fix, at 4× and at 2.5×, it produces the **same
23 / 58 flags either way** — it never sees `e106_bdy_03`, the very part it was
built for. At 600×1600×998 against a ship median of ~250 the slab sits under
3×, and lowering the factor buys noise, not sensitivity.

What the eye actually used was not scale but **relationship**: a blocky mass
sitting apart from the hull silhouette. A containment test — does a part's box
lie within the envelope its neighbours describe? — is the right numeric analogue
and is not built yet.

The screen is kept anyway, because it is honest about what it does see: two
curiosities that are **legitimate**, not defects — `f002_bdy_22`/`_23` span
519 × **100 000** × 519 (a tether/elevator column, consistent in both containers
that carry it) and `t901_e01_D` is 58×59×3013 (a mast). And `f101_bdy_01` flags
at 6× while being **capture-verified exact** in the truth table — a useful
reminder that a bulky hull is not a bug.

### The consistency figure is mostly bounding boxes — real disagreement is ONE resource

`examples/consensus_check.rs` sharpens the cross-container test: with three or
more copies of a resource, the majority span is the reference and the **minority
names the container that is wrong**, not just "these disagree".

Disc-wide: **89 minority decodes across 477 resources that have a majority** —
and **88 of the 89 are scene composites** (`rou_*` / `e_rou_*`), not drawable
geometry. Inspecting one shows why: `e_rou_e106` decodes a **24-vertex,
12-triangle box** (span 22×22×22), `e_rou_f106` another (745×718×718). A
composite's descriptor carries a **bounding box**, the anchor scan finds it, and
because those boxes are interchangeable-looking the assignment shuffles between
containers. `mesh::scene_world_nodes` identifies them structurally: **1 141 of
the 6 209 decoded resources have scene nodes**, i.e. are composites (478 are
named `rou_`/`e_rou_`).

So the headline number this file has been quoting — cross-container inconsistency
— is **dominated by composite bounding boxes**. Counting only real geometry, the
whole disc now has **one** disagreement:

```
Stage_S25.xpr   e101_wep_01_l   [779, 5769, 5769]  vs the other 3 containers' [206, 545, 545]
```

That is the sharper metric to work against, and the concrete next target. (The
composite boxes are harmless — nothing draws them — but they should be excluded
from any consistency figure quoted in future, and the ignored
`shared_resources_decode_identically_in_every_container` test measures the mixed
population, so its number is not comparable to this one.)

### What the exact-coverage fix actually reached

The fix was justified on one resource (`e106_bdy_03`) and one render. Comparing
whole-disc decodes at `XBG7_COVER_SLACK=4` and `=1` shows what else moved:

- **29 resources changed anchor**, 3 stopped decoding, 6 207 unchanged.
- Of the moved ones that appear in several containers, **0 matched the sibling
  consensus before the fix and 9 match it after** — the fix moved them onto the
  block their own copies agree on, and moved none away from it.
- **22 of the 29 were carrying another resource's geometry** under their own
  name.

The clearest case is a **shift chain** in `Stage_S28`: `n054_bdy_l` was decoding
`n056_bdy_l`'s geometry, and `n056_bdy_l` was decoding `n055_bdy_l`'s — each
resource landing on a neighbour's block, all of them plausible-looking meshes of
the right vertex count. `n054_bdy_l` decoded `(396, 315, 328)` in four containers
where its own copies agree on `(439, 284, 270)`.

That is the failure mode that matters for a reimplementation: not a missing
decode, but a **confidently wrong one under the right name** — and no
count-based metric can see it, because every count is correct. It took the
coverage invariant (real blocks reach their pool's last vertex) to separate them.

(Seven of the 29 carried geometry no resource claims. And the `ptc_pack.xpr`
`eff_*` entries in the list are small effect quads that all share one span, so
their "owner" attribution is weak — noted rather than counted.)

### The containment screen, and a correction: the assembler is NOT at fault

`examples/envelope_screen.rs` is the metric the eye actually used: assemble a
ship, and for each part measure how far its world box protrudes past the box of
all the *other* parts, **per axis** (a bow legitimately extends the long axis, so
protrusion only counts against that axis' own envelope).

It does not flag `e106_bdy_03` under the **pre-fix** decoder, and the reason is
now established. Dumping the pre-fix static assembly shows the shared turret
`e303_wep_01` occupying **1600×2100×4800** around a ~400×400×2000 hull — an
envelope nothing can protrude past.

⚠️ **This was first written up here as a static-assembler defect. That was
wrong.** The composite's nodes are clean — `rou_e303_wep_01_root` carries scale
`1.0` and an orthonormal matrix at `t[±179, 54, 32]` — and under the **current**
decoder the same assembly places the turret as a tidy `49×23×42` box at ±179.
The inflation only appears with `XBG7_COVER_SLACK=4`, i.e. the pre-fix decode:
**the exact-coverage fix repaired the turret as well as `e106_bdy_03`.** The
error came from comparing a deliberately pre-fix render against post-fix
measurements of the same resource.

So the screen's blindness had a mundane cause — one mis-decode hid another by
inflating the envelope — and it is worth keeping as a forward-looking invariant,
with that caveat: it can only see a protruding part when nothing else is inflated.

### The anchor work has plateaued at 98.7 % — state and what is left

Four evidence-driven changes took the decoder from 5 480 to **6 212 of 6 294**
resources (**98.7 %**), capture-verified anchors from 29/46 to **46/46**, and
cross-container inconsistency from 125 to **39**:

1. distinct anchor assignment (single-block, then grouped pools),
2. the connectivity cap 0.28 → 0.42, then replaced entirely by
3. the winding-consistency gate at 0.70, and
4. structural requirements on pre-pivot sub-meshes (index range, pool coverage).

**Two further threshold moves were tested and refuted**: the scale-free
degeneracy test plus a lower extent floor (zero extra resources, inconsistency
39 → 44), and lowering the grouped-pool **pivot** winding floor
(`XBG7_GROUPED_CONSISTENCY`): 0.85 → 0.80 → 0.75 decodes **no more resources**
and leaves inconsistency at 39, while quietly changing which geometry some
grouped models get (the vertex total moves), i.e. strictly worse.

So the remaining **82** misses are not a threshold away. They need a structural
answer of the kind the `.dat` composites already showed — marker lists that do
not map 1:1 onto stored blocks — and the honest next step is a capture of a stage
containing them, not more tuning.

### Where the remaining 82 misses stand — and a refuted fix

With the winding gate shipped, coverage is **6 212 / 6 294 = 98.7 %** and only
**82** resources never decode (was 225). Re-attributed:

| furthest gate reached | count |
|---|---|
| degenerate / implausible positions (`extent < 0.5`, >30 % degenerate) | 42 |
| winding consistency | 31 |
| buffer not covered by indices | 9 |
| connectivity | 0 (inert) |

**The biggest bucket is not the blocker** — which is exactly the caveat this
attribution carries. Both of its thresholds are *absolute*, which on a format
with no unit convention is a scale assumption: an area test of `< 1e-9` calls
every triangle of a small object degenerate (`g005` spans 0.346 units and scored
7 of 8), and `extent < 0.5` rejects it outright. Replacing the area test with a
**scale-free collinearity** test (`|u × w| < 1e-6·|u|·|w|`, i.e. sin of the angle
between the edges) is the principled version — and measured on this disc it
decodes **no more resources at all**, while raising cross-container inconsistency
39 → 44. Lowering the extent floor to 0.05 adds **two**.

So the fix that the histogram appeared to point at is refuted: those 42 are
resources where some *wrong* candidate reached that gate, not where the true
block was rejected. Both are kept as knobs (`XBG7_REL_DEGEN`, `XBG7_MIN_EXTENT`),
neither is the default, and the measurement is recorded so the next reader does
not re-derive it.

### Coverage has a denominator now, and the misses have a cause breakdown

Coverage has been quoted as "resources decoded" with no total. `examples/undecoded.rs`
supplies both by enumerating the XBG7 directory of every container:

**6 294 XBG7 resources on the disc — 6 069 decode (96.4 %), 225 are searched and
missed, 0 lack a usable descriptor.**

`examples/gate_histogram.rs` then asks, for each miss, **which gate the best
candidate reached** before being rejected (`mesh::debug_best_rejection`):

| furthest gate reached | count | example |
|---|---|---|
| connectivity (mean edge / diagonal) | **120** | `g004`: 0.724 > cap 0.42 |
| grouped pool — placed by a different path, not analysed here | 74 | `t170` (2 sub-meshes) |
| degenerate / implausible positions | 15 | `g005`: extent 0.346 (min 0.5), 7/8 degenerate |
| winding consistency | 9 | `e007_bdy_01_l`: 0.667 < 0.85 |
| buffer not covered by indices | 7 | `e101_bdy_02_d`: indices reach 13 171 of 15 430 |

⚠️ **Read this as a work-list, not a verdict.** "Furthest gate reached" is taken
over *all* candidates, and a wrong candidate can pass more gates than the true
block — so this says where to look, not what is broken. What it does establish is
that after the cap move to 0.42, **connectivity is still the single largest
blocker** (53 % of single-block misses), and that a third of the misses are
grouped-pool resources that need the pivot path analysed on its own terms.

The two smallest buckets are the interesting ones for a fix that cannot go wrong:
`extent < 0.5` rejects genuinely tiny props (`g005` spans 0.346), and "buffer not
covered" fires when the index buffer addresses only part of a large vertex pool —
which is exactly what a **sub-range draw** looks like, and the capture's
`indices=` field already showed the engine issuing those.

### The twin invariant, checked disc-wide (2026-08-12)

The capture gave a rule that needs no capture to apply: a `…_01`/`…_02` pair of
equal vertex count should decode to **mirrored** geometry, never to the same
buffer. `examples/twin_mirror_audit.rs` applies it to all 166 containers:

| twin pairs of equal vertex count | 34 |
|---|---|
| exact X-mirror | **18** |
| related another way (Y/Z mirror, or the same cloud in another vertex order) | 15 |
| identical — a collapse | **1** |
| unrelated — no relation at all | **0** |

Two calibration notes, because the first run of this audit got both wrong.
Comparing quantised keys **exactly** reported four false "unrelated" pairs
(`e101_eng_01/_02`): the halves are authored, not bit-negated, so they differ in
the last digits — a tolerance is required. And a mirrored pair may be stored in a
**different vertex order**, so the multiset has to be compared mirrored as well
as directly. With both fixed, nothing on the disc is unrelated.

The single collapse is `n206_01`/`n206_02` (`Stage_S08`), both anchored at
`0x33b6e54` while the container holds a second direct copy at `0x342d984` and
mirrors at `0x33b7754`/`0x342e284`. It survives because both twins are
**grouped-pool** resources (4 sub-meshes), and distinct assignment excludes that
path — so this is the concrete next target, and the fix direction is to extend
distinctness across grouped models.

`tests/mesh_consistency_disc.rs::twin_pairs_do_not_share_a_buffer` locks this in:
no twin pair may share a buffer, with `n206` the one asserted exception.

Not settled: `e106_brg_01_b_02` ≡ `e106_brg_01_l` (51 verts). A second 51-vertex
`vbase` exists in the logs but is **not** from this container, and the container
holds three near-identical 51-vertex runs, so the pair has no oracle yet.
`n006_01A` ≡ `n006_01B` shows a single `vbase` in all three logs — consistent
with real reuse, but equally with only one of the two being on screen.