re: a detector for silent XBG7 mis-decodes -- 125 of 681 shared resources fail it

Following the e106 slab to its cause. e303_wep_01 decodes from 14 containers: in
11 it is a 49x23x42 turret with organic vertices, in Stage_S02/S08/S26 the same
resource -- identical 172 verts and 330 indices -- decodes to 1600x2100x4800 of
axis-aligned box corners with (0,0,1) normals. The anchor scan found a different
buffer that happens to share both counts, so every size-based check passes. That
is the "declined for location, not format" risk succeeding wrongly instead of
declining.

The detector needs no ground truth: a resource shared by several containers must
decode to the same bounds. 681 resources appear in >=2 containers and 125 decode
to different bounds while reporting identical vertex/triangle counts -- a lower
bound, since a resource wrong everywhere is invisible to it. _rou_f401 is
62x25x10 in 16 containers and 4738x3147x4738 in 2; _rou_e011_wep05 gives four
different spans across 8.

Majority vote would resolve 104 of the 125 (14 are exact 50/50 splits) and agrees
with ground truth in the one case that has a render and a runtime capture behind
it. Recorded as a heuristic, not applied: for _rou_e302_base_break the majority
is the LARGER span and nothing yet says which is right. Flag, do not silently
rewrite geometry on a vote.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/re/INDEX.md

@@ -19,7 +19,7 @@ Promote to a prose `structures/…md` file when a format needs behavioural notes
 | LSTA sprite list | ✅ | `sylpheed-formats/src/lsta.rs` | A display list of inline elements: **T8aD sprites and `PRMD` primitives**. The `count` at `0x04` is **exact and counts both** — `count == T8aD + PRMD` for **64/64** lists on the disc, which retires the old "a few entries disagree" note (it compared sprites against a total including primitives). **All 1 281 sprite frames decode** after the T8aD rectangle-list fix |
 | IXUD subtitle | 🟡/✅ | `sylpheed-formats/src/ixud.rs` + [movie link](movie-subtitle-link.md) | timed cues. **The movie↔subtitle↔voice link is solved — statically**, from the movie config record in `tables.pak` (schema `0x067025b9`), not from the running game as this row previously assumed: [101 movies mapped](captures/movie-subtitle-voice-map.csv), 94 with subtitles, 83 with voice, 21 with a telop overlay. 93 of 94 subtitle refs resolve in the language paks; **`SUBTITLE_S12B.tbl` is missing from all six languages** — a dangling reference on the disc. Naming is `SUBTITLE_<base>.tbl` / `VOICE_<base>` with six documented exceptions. The record's ~104 **script ids** are ❔ — positional pairing drifts by three because the IDXD pool dedupes repeated values |
 | Fonts (ttf/otf/ttc) | ✅ | `sylpheed-formats/src/font.rs` | standard OpenType, parsed via ttf-parser |
-| XBG7 mesh | 🟡/❔ | `sylpheed-formats/src/mesh.rs` + `tests/mesh_disc.rs` ([xbg7](structures/xbg7-mesh.md)) | weapons/props: declaration-driven variable stride (36 models), GPU-confirmed. **Stage containers: 5662 sub-models across 22 stages** via content-anchored grouped pools (`stage_models`). **Declined set now measured**: 6 294 resources, **5 480 decode (87.1 %), 814 declined** in 31/166 containers — and it is *not* "a few quantized hero bodies" but 492 `e*`, 142 `f*`, 73 `n*`, 23 `eff*` plus `*_dead` variants. The descriptor does **not** declare decodability (word[2] is a sub-mesh count; decoded and declined appear at every value), so the gap is the anchor scan, not an unread format flag — see [xbg7](structures/xbg7-mesh.md) |
+| XBG7 mesh | 🟡/❔ | `sylpheed-formats/src/mesh.rs` + `tests/mesh_disc.rs` ([xbg7](structures/xbg7-mesh.md)) | weapons/props: declaration-driven variable stride (36 models), GPU-confirmed. **Stage containers: 5662 sub-models across 22 stages** via content-anchored grouped pools (`stage_models`). **Declined set now measured**: 6 294 resources, **5 480 decode (87.1 %), 814 declined** in 31/166 containers — and it is *not* "a few quantized hero bodies" but 492 `e*`, 142 `f*`, 73 `n*`, 23 `eff*` plus `*_dead` variants. The descriptor does **not** declare decodability (word[2] is a sub-mesh count; decoded and declined appear at every value), so the gap is the anchor scan, not an unread format flag — see [xbg7](structures/xbg7-mesh.md). **Silent mis-decodes also exist and are now measurable**: a resource shared across containers must decode to the same bounds, and **125 of 681 shared resources fail that check** with identical vertex/triangle counts — the decoder picked a different buffer of the same size. A majority vote across containers would resolve 104 of them (14 are 50/50), but that is 🟡 unvalidated beyond the one case with a render and a capture behind it |
 | Capital-ship part placement | 🟡 | `sylpheed-formats/src/ship.rs` (static) + [runtime capture](ship-placement-runtime-capture.md) | hull placement static-exact; external parts approximate statically. **Runtime capture** (Canary F10 → VS-constant WorldView) gives ground truth — validated on `e106` destroyer; not yet baked into the viewer |
 | Weapon fields defaulted on disc | ✅ | [runtime struct](structures/weapon-struct-runtime.md) · [DATA SHEET route](weapon-datasheet-runtime.md) | **Solved.** Canary maps guest RAM into `/dev/shm`, so the parsed `Weapon`/`Shell` objects are readable live; their layout is solved against disc ground truth (zero contradictions over 100+ records). All 126 weapons, exact numbers, no story progress needed — [4 393 values](captures/weapon-runtime-fields.csv) the disc does not carry. Supersedes the letter-bucket limit of the DATA SHEET route, which now serves as the independent cross-check |
 | Unit (craft/vessel) fields defaulted on disc | ✅/🟡 | [runtime struct](structures/unit-struct-runtime.md) | The parsed `unit\UN_*.tbl` definition object, vtable `0x820af844`, ≥`0x380` bytes, one per unit — **discovered, not assumed** (`unit_discover.py`), and distinguished from the spawned-entity class `0x820af030` by being one-per-ID and byte-constant within a run. Across runs only pointer words move — `--crosscheck` proves **no reported field offset is run-dependent** (two words, `+0x2c8`/`+0x2d0`, are stage-dependent and remain unidentified). 27 fields ✅ (21 units, 7 runs); the `Maneuver` block is **schema declaration order, 4 bytes/field, base `0x9c` with a two-slot gap after `AA_Roll_Min`** (29 anchors, 0 conflicts), which also pins 5 fields *no* disc record ever values. Angles are **radians at runtime, degrees on disc**. Unlike weapons, unit definitions are instantiated **per stage**, so coverage (21/110) grows by visiting missions — but a defaulted field is **not** a global constant: `Size_Y` provably inherits `Size_X` (7 independent units, 6 distinct values), and three more sibling rules are recorded ❔, recovering 65 values in units never visited — [values](captures/unit-runtime-fields.csv) |

docs/re/structures/xbg7-mesh.md

@@ -289,3 +289,52 @@ 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.