feat(viewer): stage thumbnail grid, per-model albedo, camera zoom fix

Route stage containers to stage_models (decode both single + stage paths, keep
whichever yields more geometry) — fixes stages showing the dummy-root "black
cube". Present a stage's sub-models as a thumbnail grid: each recentred and
uniformly scaled to a fixed cell so a 1-unit prop and a 3000-unit structure are
equally visible; skybox-plane / stray-anchor models (>20000u) are culled. Each
sub-model is textured by matching its name to the container's `_col` albedo.

Fix the orbit camera: zoom radius was hard-capped at 50u with fixed clip planes,
trapping the camera inside multi-thousand-unit stages. Zoom limits and near/far
planes now scale to the framed model.

Adds docs/HANDOFF-stage-mesh-2026-07-12.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

crates/sylpheed-cli/src/main.rs

@@ -97,6 +97,12 @@ enum Commands {
         #[command(subcommand)]
         cmd: PakCommands,
     },
+
+    /// XBG7 mesh tools (inspect / headless render to PNG)
+    Mesh {
+        #[command(subcommand)]
+        cmd: MeshCommands,
+    },
 }
 
 #[derive(Subcommand)]
@@ -118,6 +124,37 @@ enum PakCommands {
     },
 }
 
+#[derive(Subcommand)]
+enum MeshCommands {
+    /// Print the decoded sub-models of an XBG7 container (`.xpr`)
+    Info {
+        /// Path to the `.xpr` model / stage container
+        file: PathBuf,
+    },
+    /// Headless-render the decoded mesh(es) to a shaded PNG (software rasterizer)
+    Render {
+        /// Path to the `.xpr` model / stage container
+        file: PathBuf,
+        /// Output PNG path
+        output: PathBuf,
+        /// Image size in pixels (square)
+        #[arg(long, default_value_t = 900)]
+        size: u32,
+        /// Camera yaw in degrees
+        #[arg(long, default_value_t = 35.0)]
+        yaw: f32,
+        /// Camera pitch in degrees
+        #[arg(long, default_value_t = 22.0)]
+        pitch: f32,
+        /// Force the stage grid layout even for single models
+        #[arg(long)]
+        row: bool,
+        /// Only render sub-models whose name contains this substring
+        #[arg(long)]
+        only: Option<String>,
+    },
+}
+
 #[derive(Subcommand)]
 enum TextureCommands {
     /// Print information about a texture file
@@ -155,6 +192,12 @@ async fn main() -> Result<()> {
             TextureCommands::Info { file }           => cmd_texture_info(&file),
             TextureCommands::Export { file, output } => cmd_texture_export(&file, &output),
         },
+        Commands::Mesh { cmd } => match cmd {
+            MeshCommands::Info { file } => cmd_mesh_info(&file),
+            MeshCommands::Render { file, output, size, yaw, pitch, row, only } => {
+                cmd_mesh_render(&file, &output, size, yaw, pitch, row, only)
+            }
+        },
         Commands::Pak { cmd } => match cmd {
             PakCommands::List { pak, idxd_only } => cmd_pak_list(&pak, idxd_only),
             PakCommands::Dump { pak, hash }      => cmd_pak_dump(&pak, &hash),
@@ -357,6 +400,273 @@ fn cmd_texture_export(file: &Path, output: &Path) -> Result<()> {
     Ok(())
 }
 
+// ── mesh info / render ──────────────────────────────────────────────────────
+
+/// Decode a container and return its sub-models the same way the viewer routes:
+/// a single model (weapon / prop) OR a stage's many sub-models — whichever
+/// yields more geometry.
+fn decode_models(bytes: &[u8]) -> Vec<sylpheed_formats::mesh::Xbg7Model> {
+    use sylpheed_formats::mesh::Xbg7Model;
+    let single = Xbg7Model::from_xpr2(bytes)
+        .ok()
+        .filter(|m| !m.meshes.is_empty());
+    let single_verts = single.as_ref().map(|m| m.totals().0).unwrap_or(0);
+    let stage = Xbg7Model::stage_models(bytes);
+    let stage_verts: usize = stage.iter().map(|m| m.totals().0).sum();
+    if !stage.is_empty() && stage_verts > single_verts {
+        stage
+    } else {
+        single.into_iter().collect()
+    }
+}
+
+fn cmd_mesh_info(file: &Path) -> Result<()> {
+    let bytes = std::fs::read(file).with_context(|| format!("Cannot read {}", file.display()))?;
+    let models = decode_models(&bytes);
+    if models.is_empty() {
+        println!("{} no decodable XBG7 geometry", "Mesh:".yellow().bold());
+        return Ok(());
+    }
+    let (mut tv, mut tt) = (0usize, 0usize);
+    println!("{} {}", "Mesh:".green().bold(), file.display());
+    for m in &models {
+        let (v, t) = m.totals();
+        tv += v;
+        tt += t;
+        let mut lo = [f32::MAX; 3];
+        let mut hi = [f32::MIN; 3];
+        for sub in &m.meshes {
+            for p in &sub.positions {
+                for a in 0..3 {
+                    lo[a] = lo[a].min(p[a]);
+                    hi[a] = hi[a].max(p[a]);
+                }
+            }
+        }
+        println!(
+            "  {:16} {:>6} v {:>6} t  bbox [{:.1} {:.1} {:.1}]",
+            m.name,
+            v,
+            t,
+            hi[0] - lo[0],
+            hi[1] - lo[1],
+            hi[2] - lo[2]
+        );
+    }
+    println!(
+        "  {} {} sub-models · {} verts · {} tris",
+        "TOTAL".bold(),
+        models.len(),
+        tv,
+        tt
+    );
+    Ok(())
+}
+
+#[allow(clippy::too_many_arguments)]
+fn cmd_mesh_render(
+    file: &Path,
+    output: &Path,
+    size: u32,
+    yaw: f32,
+    pitch: f32,
+    force_row: bool,
+    only: Option<String>,
+) -> Result<()> {
+    let bytes = std::fs::read(file).with_context(|| format!("Cannot read {}", file.display()))?;
+    let mut models = decode_models(&bytes);
+    if let Some(sub) = &only {
+        models.retain(|m| m.name.contains(sub.as_str()));
+    }
+    if models.is_empty() {
+        anyhow::bail!("no decodable XBG7 geometry in {}", file.display());
+    }
+
+    // ── Build a triangle soup. ──
+    // Single models render centred; multi-model containers (stages) get the
+    // viewer's normalised **thumbnail grid**: each sub-model recentred and
+    // uniformly scaled to a fixed cell, so all are equally visible regardless of
+    // native scale (mirrors `spawn_stage_models`).
+    let multi = models.len() > 1 || force_row;
+    let mut tris: Vec<[[f32; 3]; 3]> = Vec::new();
+    const CELL: f32 = 10.0;
+    const GAP: f32 = 4.0;
+    let grid_pitch = CELL + GAP;
+    let cols = (models.len() as f32).sqrt().ceil().max(1.0) as usize;
+    for (i, m) in models.iter().enumerate() {
+        let mut lo = [f32::MAX; 3];
+        let mut hi = [f32::MIN; 3];
+        for sub in &m.meshes {
+            for p in &sub.positions {
+                for a in 0..3 {
+                    lo[a] = lo[a].min(p[a]);
+                    hi[a] = hi[a].max(p[a]);
+                }
+            }
+        }
+        if lo[0] > hi[0] {
+            continue;
+        }
+        let center = [
+            (lo[0] + hi[0]) * 0.5,
+            (lo[1] + hi[1]) * 0.5,
+            (lo[2] + hi[2]) * 0.5,
+        ];
+        let (scale, cell) = if multi {
+            let extent = (hi[0] - lo[0]).max(hi[1] - lo[1]).max(hi[2] - lo[2]).max(1e-3);
+            let col = i % cols;
+            let row = i / cols;
+            (CELL / extent, [col as f32 * grid_pitch, -(row as f32) * grid_pitch, 0.0])
+        } else {
+            (1.0, [0.0, 0.0, 0.0])
+        };
+        for sub in &m.meshes {
+            for tri in sub.indices.chunks_exact(3) {
+                let (a, b, c) = (tri[0] as usize, tri[1] as usize, tri[2] as usize);
+                if a >= sub.positions.len() || b >= sub.positions.len() || c >= sub.positions.len() {
+                    continue;
+                }
+                let f = |i: usize| {
+                    let p = sub.positions[i];
+                    [
+                        (p[0] - center[0]) * scale + cell[0],
+                        (p[1] - center[1]) * scale + cell[1],
+                        (p[2] - center[2]) * scale + cell[2],
+                    ]
+                };
+                tris.push([f(a), f(b), f(c)]);
+            }
+        }
+    }
+    if tris.is_empty() {
+        anyhow::bail!("no triangles to render");
+    }
+
+    let rgba = rasterize(&tris, size, yaw, pitch);
+    image::save_buffer(output, &rgba, size, size, image::ExtendedColorType::Rgba8)
+        .with_context(|| format!("writing PNG {}", output.display()))?;
+    println!(
+        "{} {} tris → {} ({}×{}, yaw {:.0}° pitch {:.0}°)",
+        "Rendered".green().bold(),
+        tris.len(),
+        output.display().to_string().cyan(),
+        size,
+        size,
+        yaw,
+        pitch,
+    );
+    Ok(())
+}
+
+/// Minimal software rasterizer: orthographic, z-buffered, two-sided Lambert +
+/// headlight shading over a flat grey material on a dark background. Enough to
+/// judge whether recovered geometry is coherent.
+fn rasterize(tris: &[[[f32; 3]; 3]], size: u32, yaw_deg: f32, pitch_deg: f32) -> Vec<u8> {
+    let n = size as usize;
+    let (yaw, pitch) = (yaw_deg.to_radians(), pitch_deg.to_radians());
+    let (cy, sy) = (yaw.cos(), yaw.sin());
+    let (cp, sp) = (pitch.cos(), pitch.sin());
+    // Rotate a world point into view space (yaw about Y, then pitch about X).
+    let view = |p: [f32; 3]| -> [f32; 3] {
+        let x = p[0] * cy + p[2] * sy;
+        let z0 = -p[0] * sy + p[2] * cy;
+        let y = p[1] * cp - z0 * sp;
+        let z = p[1] * sp + z0 * cp;
+        [x, y, z]
+    };
+
+    // View-space bbox → orthographic fit.
+    let mut lo = [f32::MAX; 3];
+    let mut hi = [f32::MIN; 3];
+    for t in tris {
+        for v in t {
+            let q = view(*v);
+            for a in 0..3 {
+                lo[a] = lo[a].min(q[a]);
+                hi[a] = hi[a].max(q[a]);
+            }
+        }
+    }
+    let span = (hi[0] - lo[0]).max(hi[1] - lo[1]).max(1e-3);
+    let scale = (n as f32) * 0.9 / span;
+    let cx = (lo[0] + hi[0]) * 0.5;
+    let cyv = (lo[1] + hi[1]) * 0.5;
+    let to_screen = |q: [f32; 3]| -> (f32, f32, f32) {
+        let sx = (q[0] - cx) * scale + n as f32 * 0.5;
+        let sy = n as f32 * 0.5 - (q[1] - cyv) * scale;
+        (sx, sy, q[2])
+    };
+
+    let mut color = vec![18u8; n * n * 4];
+    for i in 0..n * n {
+        color[i * 4 + 3] = 255;
+    }
+    let mut depth = vec![f32::MAX; n * n];
+    // Light in view space (upper-left-front).
+    let light = {
+        let l = [-0.4f32, 0.6, 0.7];
+        let m = (l[0] * l[0] + l[1] * l[1] + l[2] * l[2]).sqrt();
+        [l[0] / m, l[1] / m, l[2] / m]
+    };
+
+    for t in tris {
+        let v0 = view(t[0]);
+        let v1 = view(t[1]);
+        let v2 = view(t[2]);
+        // Face normal in view space.
+        let e1 = [v1[0] - v0[0], v1[1] - v0[1], v1[2] - v0[2]];
+        let e2 = [v2[0] - v0[0], v2[1] - v0[1], v2[2] - v0[2]];
+        let mut nrm = [
+            e1[1] * e2[2] - e1[2] * e2[1],
+            e1[2] * e2[0] - e1[0] * e2[2],
+            e1[0] * e2[1] - e1[1] * e2[0],
+        ];
+        let nl = (nrm[0] * nrm[0] + nrm[1] * nrm[1] + nrm[2] * nrm[2]).sqrt();
+        if nl < 1e-12 {
+            continue;
+        }
+        nrm = [nrm[0] / nl, nrm[1] / nl, nrm[2] / nl];
+        // Two-sided: diffuse from |n·L|, plus a headlight term from |n.z|.
+        let diff = (nrm[0] * light[0] + nrm[1] * light[1] + nrm[2] * light[2]).abs();
+        let head = nrm[2].abs();
+        let inten = (0.18 + 0.55 * diff + 0.3 * head).min(1.0);
+        let shade = (inten * 210.0) as u8;
+
+        let (ax, ay, az) = to_screen(v0);
+        let (bx, by, bz) = to_screen(v1);
+        let (ccx, ccy, ccz) = to_screen(v2);
+        let minx = ax.min(bx).min(ccx).floor().max(0.0) as usize;
+        let maxx = ax.max(bx).max(ccx).ceil().min(n as f32 - 1.0) as usize;
+        let miny = ay.min(by).min(ccy).floor().max(0.0) as usize;
+        let maxy = ay.max(by).max(ccy).ceil().min(n as f32 - 1.0) as usize;
+        let area = (bx - ax) * (ccy - ay) - (by - ay) * (ccx - ax);
+        if area.abs() < 1e-6 {
+            continue;
+        }
+        for py in miny..=maxy {
+            for px in minx..=maxx {
+                let fx = px as f32 + 0.5;
+                let fy = py as f32 + 0.5;
+                let w0 = ((bx - fx) * (ccy - fy) - (by - fy) * (ccx - fx)) / area;
+                let w1 = ((ccx - fx) * (ay - fy) - (ccy - fy) * (ax - fx)) / area;
+                let w2 = 1.0 - w0 - w1;
+                if w0 < 0.0 || w1 < 0.0 || w2 < 0.0 {
+                    continue;
+                }
+                let z = w0 * az + w1 * bz + w2 * ccz;
+                let idx = py * n + px;
+                if z < depth[idx] {
+                    depth[idx] = z;
+                    color[idx * 4] = shade;
+                    color[idx * 4 + 1] = shade;
+                    color[idx * 4 + 2] = (shade as f32 * 1.02).min(255.0) as u8;
+                }
+            }
+        }
+    }
+    color
+}
+
 /// Software-decode a de-tiled `X360Texture` (mip 0) to tightly-packed RGBA8.
 ///
 /// BCn blocks are decompressed with `texpresso`; uncompressed A8R8G8B8/X8R8G8B8

crates/sylpheed-formats/src/lib.rs

@@ -54,6 +54,7 @@ pub mod audio;
 pub use font::FontInfo;
 pub use idxd::{IdxdError, IdxdObject};
 pub use ixud::{Cue, Subtitle};
+pub use mesh::{GameMesh, Xbg7Model};
 pub use ratc::RatcChild;
 pub use t8ad::T8adImage;
 pub use pak::{PakArchive, PakEntry, PakError};

crates/sylpheed-formats/src/mesh.rs

@@ -1,71 +1,808 @@
-//! Mesh format parsing — TO BE REVERSE ENGINEERED.
+//! XBG7 mesh geometry decoder (geometry resources inside XPR2 containers).
 //!
-//! Project Sylpheed uses a completely custom engine with unknown mesh formats.
-//! This module is a scaffold: populate these structs as you discover the
-//! actual binary layout using a hex editor (010 Editor) and Ghidra.
+//! ## Clean-room note
 //!
-//! ## RE Strategy for Meshes
+//! This format was reverse-engineered **purely by static observation of the
+//! retail disc's `hidden/resource3d/*.xpr` files** (hex inspection + geometric
+//! validation of the recovered triangles). No game code was decompiled or
+//! copied. See `docs/re/structures/xbg7-mesh.md` for the evidence log.
 //!
-//! 1. Extract the game files using xdvdfs
-//! 2. Look for files with extensions like .mdl, .mesh, .geo, .obj, .pak
-//! 3. Open them in a hex editor — look for:
-//!    - Repeating patterns of 12 bytes (XYZ float vertices)
-//!    - Groups of 3 uint16s (triangle indices)
-//!    - Header magic bytes
-//! 4. Cross-reference with Ghidra's XEX analysis to find the load functions
+//! ## Where XBG7 lives
 //!
-//! ## Useful tools
-//! - 010 Editor with binary templates
-//! - Noesis (can preview many console formats)
-//! - binrw (this project) for writing the parser once the format is known
+//! Ship / weapon / prop models are `XPR2` containers (see [`crate::texture`]).
+//! Their resource directory holds `TX2D` texture resources **and** one or more
+//! `XBG7` geometry resources.  The `XBG7` *descriptor* (at the resource's
+//! `data_offset`) is a scene/material graph; the actual vertex and index
+//! buffers live in the container's shared data section (from `header_size`).
+//!
+//! ## The "simple" layout decoded here (CONFIRMED)
+//!
+//! For single-stream models (weapons, simple props — 36 of the 166 disc models)
+//! the data section is a straight sequence of sub-meshes, each:
+//!
+//! ```text
+//!   [ index buffer  : idx_count × u16 big-endian ]   triangle list
+//!   [ 12-byte vertex-buffer header (contents undecoded) ]
+//!   [ vertex buffer : vtx_count × stride bytes ]      (declaration-driven)
+//!   (pad to 16 bytes → next sub-mesh)
+//! ```
+//!
+//! The vertex layout is **not fixed** — it comes from a **vertex declaration**
+//! in the descriptor: a table of `{offset, format-code, usage}` triples (usage
+//! `0x00` POSITION `f32×3`, `0x03` NORMAL `f16×4`, `0x05` TEXCOORD `f16×2`).
+//! Models omit UV or use fewer elements, so stride varies (20 = pos+normal,
+//! 24 = pos+normal+uv, …). Each element is read in naive big-endian component
+//! order. Correct alignment is pinned by the recovered normals being exactly
+//! unit-length. **Cross-checked against a Canary GPU vertex-fetch capture**,
+//! which confirmed the primitive type (triangle list), formats, and offsets —
+//! see `docs/re/structures/xbg7-mesh.md`.
+//!
+//! `vtx_count` / `idx_count` come from per-sub-mesh records in the descriptor:
+//! a `[vtx_count:u32][0:u32][idx_count:u32][tail:u32]` tuple (big-endian), read
+//! in file order.  Every index is validated to be `< vtx_count`; if any
+//! sub-mesh fails to carve cleanly the whole model is rejected
+//! ([`MeshError::UnsupportedLayout`]) rather than emitting garbage.
+//!
+//! ## Not yet decoded
+//!
+//! The hero-ship *body* meshes (`DeltaSaber_*.xpr` `f004`, and ~100 other
+//! models) use a more complex **multi-stream** layout — separate position /
+//! attribute streams at descriptor-addressed offsets, quantized positions —
+//! which is not handled here and is cleanly declined.
 
+use crate::texture::{Xpr2Header, Xpr2ResourceEntry};
+use binrw::BinRead;
+use std::io::Cursor;
 use thiserror::Error;
 
 #[derive(Debug, Error)]
 pub enum MeshError {
-    #[error("Unknown mesh magic: {0:?}")]
-    UnknownMagic([u8; 4]),
-    #[error("Unsupported mesh version: {0}")]
-    UnsupportedVersion(u32),
+    #[error("Not an XPR2 container")]
+    NotXpr2,
+    #[error("No XBG7 geometry resource in container")]
+    NoGeometry,
+    #[error("Mesh layout not supported (multi-stream / quantized body mesh)")]
+    UnsupportedLayout,
     #[error("Parse error: {0}")]
-    Parse(String),
+    Parse(#[from] binrw::Error),
 }
 
 /// A decoded 3D mesh ready for Bevy.
-/// Vertex positions, normals, UVs, and indices.
 #[derive(Debug, Default, Clone)]
 pub struct GameMesh {
-    /// Interleaved vertex positions [x, y, z, x, y, z, ...]
+    /// Vertex positions in model space `[x, y, z]`.
     pub positions: Vec<[f32; 3]>,
-    /// Vertex normals [nx, ny, nz, ...]  
+    /// Vertex normals `[nx, ny, nz]` — empty when not stored (compute smooth
+    /// normals from geometry instead).
     pub normals: Vec<[f32; 3]>,
-    /// UV texture coordinates [u, v, ...]
+    /// Texture coordinates `[u, v]`.  **Best-guess channel** (attr halves 0 & 2)
+    /// pending in-game visual confirmation — see the module doc.
     pub uvs: Vec<[f32; 2]>,
-    /// Triangle list indices
+    /// Triangle-list indices (3 per triangle).
     pub indices: Vec<u32>,
-    /// Name of this mesh (if available in the file)
+    /// Sub-mesh / node name from the descriptor, when available.
     pub name: Option<String>,
 }
 
-impl GameMesh {
-    /// Parse a mesh from raw bytes.
+/// A model = the set of sub-meshes recovered from one XPR2 container's first
+/// XBG7 resource, plus the resource's name.
+#[derive(Debug, Default, Clone)]
+pub struct Xbg7Model {
+    pub name: String,
+    pub meshes: Vec<GameMesh>,
+}
+
+impl Xbg7Model {
+    /// Total vertex / triangle counts across all sub-meshes.
+    pub fn totals(&self) -> (usize, usize) {
+        let v = self.meshes.iter().map(|m| m.positions.len()).sum();
+        let t = self.meshes.iter().map(|m| m.indices.len() / 3).sum();
+        (v, t)
+    }
+
+    /// Decode the geometry of the first XBG7 resource in an XPR2 container.
     ///
-    /// TODO: implement once the actual file format is identified.
-    /// Currently returns an error until the format is reverse engineered.
-    pub fn from_bytes(_bytes: &[u8]) -> Result<Vec<Self>, MeshError> {
-        // ┌──────────────────────────────────────────────────────────────┐
-        // │  REVERSE ENGINEERING TODO                                    │
-        // │                                                              │
-        // │  Steps to implement this:                                    │
-        // │  1. Find mesh files in the extraction (look for .mdl etc.)   │
-        // │  2. Identify the file format using identify_format() in vfs  │
-        // │  3. Use 010 Editor to map the binary structure               │
-        // │  4. Add binrw #[derive(BinRead)] structs above               │
-        // │  5. Implement this function to parse and return meshes       │
-        // └──────────────────────────────────────────────────────────────┘
-        Err(MeshError::Parse(
-            "Mesh format not yet reverse engineered. \
-             See RE TODO in src/mesh.rs".to_string()
-        ))
+    /// Returns [`MeshError::UnsupportedLayout`] for models whose data section
+    /// does not carve cleanly under the simple single-stream layout (the
+    /// complex body meshes) — never partial / garbage geometry.
+    pub fn from_xpr2(bytes: &[u8]) -> Result<Self, MeshError> {
+        if bytes.len() < 16 || &bytes[..4] != b"XPR2" {
+            return Err(MeshError::NotXpr2);
+        }
+        let mut cur = Cursor::new(bytes);
+        let header = Xpr2Header::read(&mut cur)?;
+        let mut xbg: Option<Xpr2ResourceEntry> = None;
+        for _ in 0..header.num_resources {
+            let e = Xpr2ResourceEntry::read(&mut cur)?;
+            if &e.type_tag == b"XBG7" {
+                xbg = Some(e);
+                break;
+            }
+        }
+        let xbg = xbg.ok_or(MeshError::NoGeometry)?;
+
+        const DIR_BASE: usize = 0x10;
+        let desc = xbg.data_offset as usize + DIR_BASE;
+        let desc_end = (desc + xbg.descriptor_size as usize).min(bytes.len());
+        if desc >= bytes.len() {
+            return Err(MeshError::UnsupportedLayout);
+        }
+
+        // Resource name (for labelling).
+        let name = read_cstr(bytes, xbg.name_offset as usize + DIR_BASE)
+            .unwrap_or_else(|| "XBG7".to_string());
+
+        // Parse the vertex declaration (element offsets/formats + stride). The
+        // XBG7 layout is NOT fixed-stride — models omit UV or use fewer elements
+        // (stride 20 = pos+normal, stride 24 = pos+normal+uv, …). Confirmed
+        // against a Canary GPU vertex-fetch capture (see docs/re/xbg7-mesh.md).
+        let decl = parse_vertex_decl(&bytes[desc..desc_end])
+            .ok_or(MeshError::UnsupportedLayout)?;
+
+        // Extract the ordered list of sub-mesh (vtx_count, idx_count) records.
+        let records = submesh_records(&bytes[desc..desc_end]);
+        if records.is_empty() {
+            return Err(MeshError::UnsupportedLayout);
+        }
+
+        // Carve the data section sequentially.
+        let base = header.header_size as usize;
+        let mut off = 0usize; // relative to `base`
+        let mut meshes = Vec::new();
+        for (vtx_count, idx_count) in records {
+            // Each sub-mesh block is `[12-byte header][index buffer][vertex
+            // buffer]` — the SAME layout as stage resources (see
+            // `stage_models`). The header must be skipped: reading indices from
+            // the block start instead treats the 12 header bytes as 6 junk
+            // indices (2 leading degenerate triangles — the stray-triangle
+            // artifact) and drops the last 6 real indices. Skipping it leaves the
+            // vertex buffer at the identical offset (`+12 + idx_bytes`), so
+            // coverage is unchanged; only the triangle list is corrected.
+            let ib = base + off + VERTEX_BUFFER_GAP;
+            let ie = ib + idx_count * 2;
+            let vb = ie;
+            let ve = vb + vtx_count * decl.stride;
+            if ve > bytes.len() {
+                return Err(MeshError::UnsupportedLayout);
+            }
+
+            // Indices (u16 BE), validated against the sub-mesh vertex count.
+            let mut indices = Vec::with_capacity(idx_count);
+            for k in 0..idx_count {
+                let i = be16(bytes, ib + k * 2) as u32;
+                if i >= vtx_count as u32 {
+                    return Err(MeshError::UnsupportedLayout);
+                }
+                indices.push(i);
+            }
+
+            // Vertices per the declaration. The .xpr stores each element in
+            // naive big-endian component order (f32 / f16 read at consecutive
+            // offsets) — the GPU's `k8in32` fetch endianness applies to the
+            // rearranged guest-memory copy, not to these file bytes.
+            let mut positions = Vec::with_capacity(vtx_count);
+            let mut normals = Vec::with_capacity(vtx_count);
+            let mut uvs = Vec::with_capacity(vtx_count);
+            let mut normal_len_sum = 0.0f32;
+            for v in 0..vtx_count {
+                let o = vb + v * decl.stride;
+
+                // POSITION: f32×3 big-endian.
+                let p = o + decl.pos_offset;
+                let x = bef(bytes, p);
+                let y = bef(bytes, p + 4);
+                let z = bef(bytes, p + 8);
+                if !(x.is_finite() && y.is_finite() && z.is_finite()) {
+                    return Err(MeshError::UnsupportedLayout);
+                }
+                positions.push([x, y, z]);
+
+                // NORMAL: f16×4 (use xyz).
+                if let Some(no) = decl.normal_offset {
+                    let nb = o + no;
+                    let nx = half(bytes, nb);
+                    let ny = half(bytes, nb + 2);
+                    let nz = half(bytes, nb + 4);
+                    normal_len_sum += (nx * nx + ny * ny + nz * nz).sqrt();
+                    normals.push([nx, ny, nz]);
+                }
+
+                // TEXCOORD: f16×2.
+                if let Some(uo) = decl.uv_offset {
+                    let ub = o + uo;
+                    uvs.push([half(bytes, ub), half(bytes, ub + 2)]);
+                }
+            }
+
+            // Sanity gate: when the declaration has a normal element, a correctly
+            // aligned vertex buffer yields unit-length normals. A mean far from 1
+            // means the layout does not fit (wrong stride / offset) — decline
+            // rather than emit garbage.
+            if decl.normal_offset.is_some() {
+                let mean = normal_len_sum / vtx_count.max(1) as f32;
+                if !(0.5..=2.0).contains(&mean) {
+                    return Err(MeshError::UnsupportedLayout);
+                }
+            }
+
+            meshes.push(GameMesh {
+                positions,
+                normals,
+                uvs,
+                indices,
+                name: None,
+            });
+            off = align16(ve) - base;
+        }
+
+        Ok(Xbg7Model { name, meshes })
+    }
+
+    /// Decode **every** locatable XBG7 geometry resource in a container.
+    ///
+    /// "Stage" containers (`hidden/resource3d/Stage_*.xpr`) are collections of
+    /// many enemy / prop sub-models, each an independent XBG7 resource. Unlike
+    /// the single-stream weapon layout (index buffer immediately followed by its
+    /// vertex buffer), a stage's index buffers and vertex buffers live in
+    /// **separate grouped pools**, and the container stores each resource's
+    /// buffer *sizes* (index count via the marker, vertex count 32 bytes before
+    /// it) but **not** an explicit data offset — the on-disc block layout is a
+    /// separate allocation order we have not reversed.
+    ///
+    /// Rather than guess that order, each resource's `[index buffer][vertex
+    /// buffer]` block is located by **content**: the unique offset in the data
+    /// section where (a) all `index_count` indices are `< vertex_count`, (b) the
+    /// stored normals are unit length, and (c) the resulting triangles are
+    /// non-degenerate with a real spatial extent. This signature is strong
+    /// enough to pin a block unambiguously in a multi-megabyte file. Resources
+    /// that cannot be located and validated this way are **skipped** (never
+    /// emitted as garbage) — including the biggest hero bodies, which use the
+    /// same quantized/complex layout that [`Xbg7Model::from_xpr2`] declines.
+    ///
+    /// Returns one [`Xbg7Model`] per decoded resource (empty if none decode).
+    pub fn stage_models(bytes: &[u8]) -> Vec<Xbg7Model> {
+        let mut out = Vec::new();
+        if bytes.len() < 16 || &bytes[..4] != b"XPR2" {
+            return out;
+        }
+        let mut cur = Cursor::new(bytes);
+        let header = match Xpr2Header::read(&mut cur) {
+            Ok(h) => h,
+            Err(_) => return out,
+        };
+        let data_base = header.header_size as usize;
+        if data_base >= bytes.len() {
+            return out;
+        }
+
+        // ── Collect every XBG7 resource's parameters up front. ──
+        const DIR_BASE: usize = 0x10;
+        struct Res {
+            name: String,
+            index_count: usize,
+            vtx_count: usize,
+            decl: VertexDecl,
+        }
+        let mut resources: Vec<Res> = Vec::new();
+        for _ in 0..header.num_resources {
+            let e = match Xpr2ResourceEntry::read(&mut cur) {
+                Ok(e) => e,
+                Err(_) => break,
+            };
+            if &e.type_tag != b"XBG7" {
+                continue;
+            }
+            let desc = e.data_offset as usize + DIR_BASE;
+            let desc_end = (desc + e.descriptor_size as usize).min(bytes.len());
+            if desc >= bytes.len() || desc_end <= desc {
+                continue;
+            }
+            let d = &bytes[desc..desc_end];
+            let (mk_rel, index_count) = match find_index_marker(d) {
+                Some(m) => m,
+                None => continue,
+            };
+            let decl = match parse_vertex_decl(d) {
+                Some(v) => v,
+                None => continue,
+            };
+            // Total vertex count is stored 32 bytes before the index marker.
+            if mk_rel < 32 {
+                continue;
+            }
+            let vtx_count = be32(d, mk_rel - 32) as usize;
+            if !(3..=400_000).contains(&vtx_count) || index_count < 3 {
+                continue;
+            }
+            // Anchoring relies on the unit-normal signature; skip resources with
+            // no NORMAL element (too ambiguous to pin safely).
+            if decl.normal_offset.is_none() {
+                continue;
+            }
+            let name = read_cstr(bytes, e.name_offset as usize + DIR_BASE)
+                .unwrap_or_else(|| "XBG7".to_string());
+            resources.push(Res {
+                name,
+                index_count,
+                vtx_count,
+                decl,
+            });
+        }
+        if resources.is_empty() {
+            return out;
+        }
+
+        // ── One O(file) pass per distinct stride: find vertex-block *starts*. ──
+        //
+        // Each geometry block is `[12B header][index buffer][vertex buffer]`, and
+        // the blocks are scattered among texture data with no stored offset. But
+        // a vertex buffer is a run of stride-sized records whose NORMAL (f16×4 at
+        // +12) is unit length; a *block start* is the unique offset where that
+        // run begins — the previous stride slot is NOT a unit-normal vertex (it's
+        // index bytes / header). Collecting those starts turns the per-resource
+        // search from O(file) into a scan of a few hundred candidates.
+        let mut strides: Vec<usize> = resources.iter().map(|r| r.decl.stride).collect();
+        strides.sort_unstable();
+        strides.dedup();
+        let mut starts_by_stride: std::collections::BTreeMap<usize, Vec<usize>> =
+            std::collections::BTreeMap::new();
+        for &s in &strides {
+            starts_by_stride.insert(s, vertex_run_starts(bytes, data_base, s));
+        }
+
+        for r in &resources {
+            let starts = &starts_by_stride[&r.decl.stride];
+            if let Some(mesh) = anchor_pool_mesh(
+                bytes,
+                starts,
+                r.index_count,
+                r.vtx_count,
+                &r.decl,
+            ) {
+                out.push(Xbg7Model {
+                    name: r.name.clone(),
+                    meshes: vec![mesh],
+                });
+            }
+        }
+        out
+    }
+}
+
+/// Scan the data section for offsets that begin a `stride`-sized unit-normal
+/// vertex run (NORMAL is `f16×4` at vertex offset +12). A run *start* is an
+/// offset whose normal is unit while the preceding stride slot's is not — i.e.
+/// the first vertex of a buffer, not a mid-buffer position. Returns the sorted
+/// candidate starts (block vertex-buffer offsets).
+fn vertex_run_starts(bytes: &[u8], data_base: usize, stride: usize) -> Vec<usize> {
+    const NRM: usize = 12; // POSITION f32×3 occupies [0,12); NORMAL f16×4 follows
+    let mut starts = Vec::new();
+    if stride < NRM + 8 {
+        return starts;
+    }
+    let is_unit = |o: usize| -> bool {
+        if o + NRM + 6 > bytes.len() {
+            return false;
+        }
+        let nx = half(bytes, o + NRM);
+        let ny = half(bytes, o + NRM + 2);
+        let nz = half(bytes, o + NRM + 4);
+        let l = (nx * nx + ny * ny + nz * nz).sqrt();
+        (0.85..=1.15).contains(&l)
+    };
+    // Vertex buffers begin on 4-byte boundaries in practice; step 4.
+    let end = bytes.len().saturating_sub(NRM + 6);
+    let mut o = data_base;
+    while o <= end {
+        if is_unit(o) && (o < data_base + stride || !is_unit(o - stride)) {
+            starts.push(o);
+        }
+        o += 4;
+    }
+    starts
+}
+
+/// Locate a stage resource's `[index buffer][vertex buffer]` block among the
+/// precomputed vertex-run `starts` (see [`vertex_run_starts`]) and decode it, or
+/// return `None` if no candidate validates. A candidate `vb` is accepted when
+/// the `index_count` indices ending just before it are all `< vtx_count`,
+/// reference (nearly) all vertices, and produce non-degenerate triangles with a
+/// real spatial extent — a signature strong enough to pin the block.
+fn anchor_pool_mesh(
+    bytes: &[u8],
+    starts: &[usize],
+    index_count: usize,
+    vtx_count: usize,
+    decl: &VertexDecl,
+) -> Option<GameMesh> {
+    let stride = decl.stride;
+    let idx_bytes = index_count * 2;
+    let vtx_bytes = vtx_count.checked_mul(stride)?;
+
+    for &vb in starts {
+        // The index buffer sits immediately before the vertex buffer.
+        if vb < idx_bytes {
+            continue;
+        }
+        let ib = vb - idx_bytes;
+        if vb + vtx_bytes > bytes.len() {
+            continue;
+        }
+
+        // ── Full index validation: every index in range, uses ~all vertices. ──
+        let mut max_idx = 0u32;
+        let mut ok = true;
+        for k in 0..index_count {
+            let i = be16(bytes, ib + k * 2) as u32;
+            if i >= vtx_count as u32 {
+                ok = false;
+                break;
+            }
+            max_idx = max_idx.max(i);
+        }
+        if !ok || (max_idx as usize) + 4 < vtx_count {
+            continue;
+        }
+
+        // ── Triangle quality: finite, non-degenerate, real spatial extent. ──
+        let pos = decl.pos_offset;
+        let mut lo = [f32::MAX; 3];
+        let mut hi = [f32::MIN; 3];
+        let mut degenerate = 0usize;
+        let mut sampled = 0usize;
+        let mut edge_sum = 0.0f32;
+        let tris = index_count / 3;
+        let tstep = (tris / 96).max(1);
+        let mut t = 0;
+        let mut bad = false;
+        while t < tris {
+            let mut p = [[0.0f32; 3]; 3];
+            for (c, pc) in p.iter_mut().enumerate() {
+                let vi = be16(bytes, ib + (3 * t + c) * 2) as usize;
+                let base = vb + vi * stride + pos;
+                for (a, slot) in pc.iter_mut().enumerate() {
+                    let x = bef(bytes, base + a * 4);
+                    if !x.is_finite() || x.abs() > 1.0e6 {
+                        bad = true;
+                        break;
+                    }
+                    *slot = x;
+                    lo[a] = lo[a].min(x);
+                    hi[a] = hi[a].max(x);
+                }
+                if bad {
+                    break;
+                }
+            }
+            if bad {
+                break;
+            }
+            let u = [p[1][0] - p[0][0], p[1][1] - p[0][1], p[1][2] - p[0][2]];
+            let w = [p[2][0] - p[0][0], p[2][1] - p[0][1], p[2][2] - p[0][2]];
+            let cx = [
+                u[1] * w[2] - u[2] * w[1],
+                u[2] * w[0] - u[0] * w[2],
+                u[0] * w[1] - u[1] * w[0],
+            ];
+            if 0.5 * (cx[0] * cx[0] + cx[1] * cx[1] + cx[2] * cx[2]).sqrt() < 1.0e-9 {
+                degenerate += 1;
+            }
+            // Sum the triangle's three edge lengths (for the connectivity check).
+            let e3 = [p[2][0] - p[1][0], p[2][1] - p[1][1], p[2][2] - p[1][2]];
+            edge_sum += (u[0] * u[0] + u[1] * u[1] + u[2] * u[2]).sqrt()
+                + (w[0] * w[0] + w[1] * w[1] + w[2] * w[2]).sqrt()
+                + (e3[0] * e3[0] + e3[1] * e3[1] + e3[2] * e3[2]).sqrt();
+            sampled += 1;
+            t += tstep;
+        }
+        if bad {
+            continue;
+        }
+        let extent = (hi[0] - lo[0]).max(hi[1] - lo[1]).max(hi[2] - lo[2]);
+        if extent < 0.5 || sampled == 0 || degenerate * 10 > sampled * 3 {
+            continue; // too flat, or >30% degenerate → not this block
+        }
+        // Connectivity check: a correctly-anchored mesh has triangle edges that
+        // are SMALL relative to its overall size (~0.05–0.15 of the bbox
+        // diagonal). A wrong anchor / cross-wired index buffer connects distant
+        // vertices, so its mean edge spans a large fraction of the model (a spiky
+        // mess). Reject those — try another candidate or decline.
+        let diag = ((hi[0] - lo[0]).powi(2) + (hi[1] - lo[1]).powi(2) + (hi[2] - lo[2]).powi(2))
+            .sqrt()
+            .max(1e-6);
+        let mean_edge = edge_sum / (sampled as f32 * 3.0);
+        if mean_edge / diag > 0.28 {
+            continue;
+        }
+
+        // ── Accepted: read the full mesh. ──
+        return Some(read_pool_mesh(bytes, ib, vb, index_count, vtx_count, decl));
+    }
+    None
+}
+
+/// Read positions / normals / uvs / indices for an anchored stage block.
+fn read_pool_mesh(
+    bytes: &[u8],
+    ib: usize,
+    vb: usize,
+    index_count: usize,
+    vtx_count: usize,
+    decl: &VertexDecl,
+) -> GameMesh {
+    let stride = decl.stride;
+    let indices: Vec<u32> = (0..index_count)
+        .map(|k| be16(bytes, ib + k * 2) as u32)
+        .collect();
+
+    let mut positions = Vec::with_capacity(vtx_count);
+    let mut normals = Vec::with_capacity(vtx_count);
+    let mut uvs = Vec::with_capacity(vtx_count);
+    for v in 0..vtx_count {
+        let o = vb + v * stride;
+        let p = o + decl.pos_offset;
+        positions.push([bef(bytes, p), bef(bytes, p + 4), bef(bytes, p + 8)]);
+        if let Some(no) = decl.normal_offset {
+            let nb = o + no;
+            normals.push([half(bytes, nb), half(bytes, nb + 2), half(bytes, nb + 4)]);
+        }
+        if let Some(uo) = decl.uv_offset {
+            let ub = o + uo;
+            uvs.push([half(bytes, ub), half(bytes, ub + 2)]);
+        }
+    }
+    GameMesh {
+        positions,
+        normals,
+        uvs,
+        indices,
+        name: None,
+    }
+}
+
+// ── Layout constants ────────────────────────────────────────────────────────
+
+/// Size of the header that precedes each sub-mesh block's index buffer
+/// (`[12-byte header][index buffer][vertex buffer]`). Contents not yet decoded.
+const VERTEX_BUFFER_GAP: usize = 12;
+
+// ── Vertex declaration ───────────────────────────────────────────────────────
+
+/// The vertex layout for one XBG7 resource, parsed from the descriptor's
+/// declaration table (shared by all its sub-meshes).
+struct VertexDecl {
+    /// Bytes per vertex.
+    stride: usize,
+    /// Byte offset of the POSITION element (`f32×3`) within a vertex.
+    pos_offset: usize,
+    /// Byte offset of the NORMAL element (`f16×4`), if present.
+    normal_offset: Option<usize>,
+    /// Byte offset of the TEXCOORD element (`f16×2`), if present.
+    uv_offset: Option<usize>,
+}
+
+/// Known element format codes → element size in bytes (from the GPU capture:
+/// POSITION `f32×3`, NORMAL `f16×4`, TEXCOORD `f16×2`).
+fn decl_code_size(code: u32) -> Option<usize> {
+    match code {
+        0x2A_23B9 => Some(12), // f32×3  (POSITION)
+        0x1A_2360 => Some(8),  // f16×4  (NORMAL)
+        0x2C_235F => Some(4),  // f16×2  (TEXCOORD)
+        _ => None,
+    }
+}
+
+/// Parse the XBG7 vertex declaration: a table of `{offset:u32, code:u32,
+/// usage<<16:u32}` big-endian triples that follows the `(index_bytes,
+/// index_count)` marker, terminated by an `offset == 0x00FF0000` /
+/// `code == 0xFFFFFFFF` sentinel. Usage codes: `0` POSITION, `3` NORMAL,
+/// `5` TEXCOORD. Stride is the max element extent; unknown element sizes are
+/// inferred from the next element's offset.
+/// Locate the `(index_bytes, index_count)` marker in a descriptor: the first
+/// big-endian pair where `index_bytes == index_count * 2` and `index_count` is a
+/// positive multiple of 3. Returns `(rel_offset, index_count)`.
+fn find_index_marker(desc: &[u8]) -> Option<(usize, usize)> {
+    let mut rel = 0usize;
+    while rel + 40 <= desc.len() {
+        let a = be32(desc, rel);
+        let c = be32(desc, rel + 4);
+        if c >= 3 && c % 3 == 0 && c < 400_000 && a == c * 2 {
+            return Some((rel, c as usize));
+        }
+        rel += 4;
+    }
+    None
+}
+
+fn parse_vertex_decl(desc: &[u8]) -> Option<VertexDecl> {
+    let mk = find_index_marker(desc)?.0;
+
+    // Read declaration triples.
+    let mut elems: Vec<(usize, u32, u32)> = Vec::new(); // (offset, code, usage)
+    let mut r = mk + 8;
+    for _ in 0..16 {
+        if r + 12 > desc.len() {
+            break;
+        }
+        let off = be32(desc, r);
+        let code = be32(desc, r + 4);
+        let usage = be32(desc, r + 8) >> 16;
+        if off == 0x00FF_0000 || code == 0xFFFF_FFFF {
+            break;
+        }
+        if off as usize > 0x1000 {
+            break; // out-of-range offset — not a real element
+        }
+        elems.push((off as usize, code & 0x00FF_FFFF, usage));
+        r += 12;
+    }
+    if elems.is_empty() {
+        return None;
+    }
+
+    let mut stride = 0usize;
+    for (i, &(off, code, _)) in elems.iter().enumerate() {
+        let size = decl_code_size(code).unwrap_or_else(|| {
+            if i + 1 < elems.len() {
+                elems[i + 1].0.saturating_sub(off)
+            } else {
+                4
+            }
+        });
+        stride = stride.max(off + size);
+    }
+    if stride == 0 || stride > 256 {
+        return None;
+    }
+
+    let pos_offset = elems
+        .iter()
+        .find(|&&(_, c, u)| c == 0x2A_23B9 || u == 0)
+        .map(|&(o, _, _)| o)
+        .unwrap_or(0);
+    let normal_offset = elems.iter().find(|&&(_, _, u)| u == 3).map(|&(o, _, _)| o);
+    let uv_offset = elems.iter().find(|&&(_, _, u)| u == 5).map(|&(o, _, _)| o);
+
+    Some(VertexDecl {
+        stride,
+        pos_offset,
+        normal_offset,
+        uv_offset,
+    })
+}
+
+// ── Descriptor sub-mesh record scan ─────────────────────────────────────────
+
+/// Scan an XBG7 descriptor for the ordered list of per-sub-mesh
+/// `(vtx_count, idx_count)` records.
+///
+/// The record is a big-endian tuple `[vtx:u32][0:u32][idx:u32][tail:u32]` with
+/// `3 ≤ vtx ≤ 65535`, the second word zero, `idx` a positive multiple of 3, and
+/// a small non-zero `tail`.  Found by a sliding 4-byte scan (records are not on
+/// a fixed stride in the scene graph).
+fn submesh_records(desc: &[u8]) -> Vec<(usize, usize)> {
+    let mut out = Vec::new();
+    if desc.len() < 16 {
+        return out;
+    }
+    let mut rel = 0usize;
+    while rel + 16 <= desc.len() {
+        let a = be32(desc, rel);
+        let z = be32(desc, rel + 4);
+        let c = be32(desc, rel + 8);
+        let t = be32(desc, rel + 12);
+        if (3..=65535).contains(&a)
+            && z == 0
+            && c >= 3
+            && c <= 200_000
+            && c % 3 == 0
+            && (1..=64).contains(&t)
+        {
+            out.push((a as usize, c as usize));
+            rel += 16; // consume the record
+        } else {
+            rel += 4;
+        }
+    }
+    out
+}
+
+// ── Little primitive readers ────────────────────────────────────────────────
+
+#[inline]
+fn align16(x: usize) -> usize {
+    (x + 15) & !15
+}
+#[inline]
+fn be16(b: &[u8], o: usize) -> u16 {
+    u16::from_be_bytes([b[o], b[o + 1]])
+}
+#[inline]
+fn be32(b: &[u8], o: usize) -> u32 {
+    u32::from_be_bytes([b[o], b[o + 1], b[o + 2], b[o + 3]])
+}
+#[inline]
+fn bef(b: &[u8], o: usize) -> f32 {
+    f32::from_be_bytes([b[o], b[o + 1], b[o + 2], b[o + 3]])
+}
+/// Big-endian IEEE-754 half → f32.
+#[inline]
+fn half(b: &[u8], o: usize) -> f32 {
+    f16_to_f32(be16(b, o))
+}
+
+/// Minimal IEEE-754 binary16 → binary32 (no external dep).
+fn f16_to_f32(h: u16) -> f32 {
+    let sign = (h >> 15) & 1;
+    let exp = (h >> 10) & 0x1F;
+    let mant = h & 0x3FF;
+    let bits: u32 = match exp {
+        0 if mant == 0 => (sign as u32) << 31, // ±0
+        0 => {
+            // subnormal → normalize
+            let mut e: i32 = -1;
+            let mut m = mant as u32;
+            loop {
+                e += 1;
+                m <<= 1;
+                if m & 0x400 != 0 {
+                    break;
+                }
+            }
+            let exp32 = (127 - 15 - e) as u32;
+            ((sign as u32) << 31) | (exp32 << 23) | ((m & 0x3FF) << 13)
+        }
+        0x1F => ((sign as u32) << 31) | (0xFF << 23) | ((mant as u32) << 13), // Inf/NaN
+        _ => {
+            let exp32 = (exp as i32 - 15 + 127) as u32;
+            ((sign as u32) << 31) | (exp32 << 23) | ((mant as u32) << 13)
+        }
+    };
+    f32::from_bits(bits)
+}
+
+fn read_cstr(b: &[u8], o: usize) -> Option<String> {
+    if o >= b.len() {
+        return None;
+    }
+    let end = b[o..].iter().position(|&c| c == 0).map(|p| o + p)?;
+    if end == o {
+        return None;
+    }
+    Some(String::from_utf8_lossy(&b[o..end]).into_owned())
+}
+
+// ── Tests ───────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn half_roundtrip_known_values() {
+        assert_eq!(f16_to_f32(0x3C00), 1.0); // 1.0
+        assert_eq!(f16_to_f32(0x0000), 0.0); // +0
+        assert_eq!(f16_to_f32(0xBC00), -1.0); // -1.0
+        assert_eq!(f16_to_f32(0x4000), 2.0); // 2.0
+        assert!((f16_to_f32(0x3800) - 0.5).abs() < 1e-6); // 0.5
+    }
+
+    #[test]
+    fn submesh_record_scan_finds_tuple() {
+        // [vtx=215][0][idx=1092][tail=4]
+        let mut d = vec![0u8; 32];
+        d[0..4].copy_from_slice(&215u32.to_be_bytes());
+        d[8..12].copy_from_slice(&1092u32.to_be_bytes());
+        d[12..16].copy_from_slice(&4u32.to_be_bytes());
+        let recs = submesh_records(&d);
+        assert_eq!(recs, vec![(215, 1092)]);
+    }
+
+    #[test]
+    fn rejects_non_xpr2() {
+        assert!(matches!(
+            Xbg7Model::from_xpr2(b"NOPEnotacontainerXXXXXXXX"),
+            Err(MeshError::NotXpr2)
+        ));
     }
 }

crates/sylpheed-formats/src/texture.rs

@@ -221,6 +221,45 @@ impl X360Texture {
     /// 3. Read its GPUTEXTURE_FETCH_CONSTANT (GPUFC) at descriptor +0x18
     /// 4. De-tile the pixel data (if tiled) → linear layout
     pub fn from_xpr2(bytes: &[u8]) -> Result<Self, TextureError> {
+        // XPR files are frequently PACKS of many textures; XPR_RES_INDEX picks
+        // the Nth texture resource (default 0) for RE/browse validation.
+        let want = std::env::var("XPR_RES_INDEX")
+            .ok()
+            .and_then(|v| v.trim().parse::<usize>().ok())
+            .unwrap_or(0);
+        Self::from_xpr2_index(bytes, want)
+    }
+
+    /// List the names of the texture resources (`TX2D` / `TXCM`) in an XPR2
+    /// container, in directory order — the same order [`from_xpr2_index`]
+    /// selects by. Non-texture resources (e.g. `XBG7`) are skipped.
+    pub fn texture_names(bytes: &[u8]) -> Vec<String> {
+        use std::io::Cursor;
+        let mut cur = Cursor::new(bytes);
+        let Ok(header) = Xpr2Header::read(&mut cur) else {
+            return Vec::new();
+        };
+        let mut names = Vec::new();
+        for _ in 0..header.num_resources {
+            let Ok(e) = Xpr2ResourceEntry::read(&mut cur) else {
+                break;
+            };
+            if e.is_texture() || e.is_cubemap() {
+                const DIR_BASE: usize = 0x10;
+                let no = e.name_offset as usize + DIR_BASE;
+                let name = bytes
+                    .get(no..)
+                    .and_then(|s| s.iter().position(|&c| c == 0).map(|p| &s[..p]))
+                    .map(|s| String::from_utf8_lossy(s).into_owned())
+                    .unwrap_or_default();
+                names.push(name);
+            }
+        }
+        names
+    }
+
+    /// Decode the `want`-th texture resource (`TX2D` / `TXCM`, directory order).
+    pub fn from_xpr2_index(bytes: &[u8], want: usize) -> Result<Self, TextureError> {
         use std::io::Cursor;
         let mut cur = Cursor::new(bytes);
 
@@ -236,12 +275,6 @@ impl X360Texture {
         // Select a texture resource. TX2D = 2D texture; TXCM = cubemap
         // (skybox / backdrop) — same 52-byte descriptor + GPUFC layout, but the
         // pixel section holds 6 faces. For a preview we decode face 0.
-        // XPR files are frequently PACKS of many textures; XPR_RES_INDEX picks
-        // the Nth texture resource (default 0) for RE/browse validation.
-        let want = std::env::var("XPR_RES_INDEX")
-            .ok()
-            .and_then(|v| v.trim().parse::<usize>().ok())
-            .unwrap_or(0);
         let tex_entry = entries.iter()
             .filter(|e| e.is_texture() || e.is_cubemap())
             .nth(want)

crates/sylpheed-formats/tests/mesh_disc.rs

@@ -0,0 +1,234 @@
+//! Integration test: decode XBG7 geometry from REAL `.xpr` model files.
+//!
+//! Uses loose files extracted from the retail disc (models live in
+//! `hidden/resource3d/*.xpr`). Skipped unless the directory is found; point
+//! `SYLPHEED_RES3D` at it to override.
+//!
+//! Run: `cargo test -p sylpheed-formats --test mesh_disc -- --ignored --nocapture`
+
+use std::path::PathBuf;
+use sylpheed_formats::mesh::Xbg7Model;
+
+fn res3d_dir() -> Option<PathBuf> {
+    if let Ok(p) = std::env::var("SYLPHEED_RES3D") {
+        let p = PathBuf::from(p);
+        if p.is_dir() {
+            return Some(p);
+        }
+    }
+    let default =
+        PathBuf::from("/home/fabi/RE Project Sylpheed/sylph_extract/hidden/resource3d");
+    default.is_dir().then_some(default)
+}
+
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn weapon_model_decodes_to_expected_geometry() {
+    let Some(dir) = res3d_dir() else {
+        eprintln!("SKIP: resource3d dir not found (set SYLPHEED_RES3D)");
+        return;
+    };
+
+    // rou_f001_wep_00 = the player ship's first weapon: 1 sub-mesh,
+    // 215 vertices, 364 triangles (verified by hex analysis).
+    let bytes = std::fs::read(dir.join("rou_f001_wep_00.xpr")).unwrap();
+    let model = Xbg7Model::from_xpr2(&bytes).expect("weapon must decode");
+    assert_eq!(model.meshes.len(), 1);
+    let (v, t) = model.totals();
+    assert_eq!(v, 215, "vertex count");
+    assert_eq!(t, 364, "triangle count");
+
+    let m = &model.meshes[0];
+    assert_eq!(m.positions.len(), 215);
+    assert_eq!(m.uvs.len(), 215);
+    assert_eq!(m.normals.len(), 215);
+    assert_eq!(m.indices.len(), 1092);
+    // every index in range
+    assert!(m.indices.iter().all(|&i| (i as usize) < m.positions.len()));
+    // positions are real geometry within the model's ~2-unit bbox
+    let ys: Vec<f32> = m.positions.iter().map(|p| p[1]).collect();
+    let span = ys.iter().cloned().fold(f32::MIN, f32::max)
+        - ys.iter().cloned().fold(f32::MAX, f32::min);
+    assert!(span > 1.0 && span < 10.0, "y-span {span} out of range");
+
+    // Correct vertex alignment ⇒ normals are unit-length (the pin for the
+    // +12 vertex offset) and UVs land in a sane texture range.
+    let mean_nlen: f32 = m
+        .normals
+        .iter()
+        .map(|n| (n[0] * n[0] + n[1] * n[1] + n[2] * n[2]).sqrt())
+        .sum::<f32>()
+        / m.normals.len() as f32;
+    assert!((mean_nlen - 1.0).abs() < 0.05, "mean |normal| {mean_nlen} ≠ 1");
+    assert!(
+        m.uvs.iter().all(|uv| uv[0] > -0.1 && uv[0] < 2.0 && uv[1] > -0.1 && uv[1] < 2.0),
+        "UVs out of expected [0,1]-ish range"
+    );
+}
+
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn declaration_driven_decode_covers_expected_model_count() {
+    let Some(dir) = res3d_dir() else {
+        return;
+    };
+    let mut ok = 0usize;
+    let mut total = 0usize;
+    for entry in std::fs::read_dir(&dir).unwrap() {
+        let path = entry.unwrap().path();
+        if path.extension().and_then(|e| e.to_str()) != Some("xpr") {
+            continue;
+        }
+        total += 1;
+        let bytes = std::fs::read(&path).unwrap();
+        if let Ok(model) = Xbg7Model::from_xpr2(&bytes) {
+            if !model.meshes.is_empty() {
+                // Every decoded model must be self-consistent (indices in range,
+                // and unit normals where present).
+                for m in &model.meshes {
+                    assert!(
+                        m.indices.iter().all(|&i| (i as usize) < m.positions.len()),
+                        "{path:?}: index out of range"
+                    );
+                }
+                ok += 1;
+            }
+        }
+    }
+    eprintln!("XBG7 declaration-driven decode: {ok}/{total} models");
+    // Variable-stride declaration parsing lifted coverage vs the old fixed
+    // stride-24 decoder (25 → 36 fully-validated models; multi-sub-mesh models
+    // whose later sub-mesh offset isn't yet handled are still declined whole).
+    assert!(ok >= 35, "coverage regressed: only {ok}/{total} decoded");
+}
+
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn complex_body_mesh_is_declined_not_garbage() {
+    let Some(dir) = res3d_dir() else {
+        return;
+    };
+    // The hero-ship body uses the multi-stream layout we do not decode; it must
+    // be cleanly rejected, never returned as partial geometry.
+    let bytes = std::fs::read(dir.join("DeltaSaber_A.xpr")).unwrap();
+    match Xbg7Model::from_xpr2(&bytes) {
+        Err(_) => {}                       // expected: declined
+        Ok(m) => {
+            // If it ever does decode, it must at least be self-consistent.
+            for mesh in &m.meshes {
+                assert!(mesh.indices.iter().all(|&i| (i as usize) < mesh.positions.len()));
+            }
+        }
+    }
+}
+
+/// Stage containers decode multiple enemy/prop sub-models via content anchoring.
+/// Prints coverage; asserts the known-good Stage_S10 meshes decode with sane geometry.
+#[test]
+#[ignore]
+fn stage_models_decode() {
+    use sylpheed_formats::mesh::Xbg7Model;
+    let dir = std::env::var("SYLPHEED_RES3D").unwrap_or_else(|_| {
+        "/home/fabi/RE Project Sylpheed/sylph_extract/hidden/resource3d".to_string()
+    });
+    let path = format!("{dir}/Stage_S10.xpr");
+    let bytes = std::fs::read(&path).expect("read Stage_S10");
+    let models = Xbg7Model::stage_models(&bytes);
+    for m in &models {
+        let (v, t) = m.totals();
+        let mut lo = [f32::MAX; 3];
+        let mut hi = [f32::MIN; 3];
+        for sub in &m.meshes {
+            for p in &sub.positions {
+                for a in 0..3 {
+                    lo[a] = lo[a].min(p[a]);
+                    hi[a] = hi[a].max(p[a]);
+                }
+            }
+        }
+        let ext = [hi[0] - lo[0], hi[1] - lo[1], hi[2] - lo[2]];
+        println!(
+            "  {:16} v={v} t={t} bbox=[{:.1},{:.1},{:.1}]",
+            m.name, ext[0], ext[1], ext[2]
+        );
+    }
+    // Known: e003 (main enemy, ~1400 tris) must be among the decoded models.
+    let e003 = models.iter().find(|m| m.name == "e003").expect("e003 decoded");
+    let (v, t) = e003.totals();
+    assert_eq!(v, 2383, "e003 vertex count");
+    assert!(t > 1400, "e003 triangle count {t}");
+    assert!(models.len() >= 3, "at least 3 stage sub-models, got {}", models.len());
+}
+
+/// Timing/coverage sweep across all stage containers (manual; release recommended).
+#[test]
+#[ignore]
+fn stage_models_sweep() {
+    use sylpheed_formats::mesh::Xbg7Model;
+    use std::time::Instant;
+    let dir = std::env::var("SYLPHEED_RES3D").unwrap_or_else(|_| {
+        "/home/fabi/RE Project Sylpheed/sylph_extract/hidden/resource3d".to_string()
+    });
+    let mut names: Vec<_> = std::fs::read_dir(&dir)
+        .unwrap()
+        .filter_map(|e| e.ok().map(|e| e.file_name().into_string().unwrap()))
+        .filter(|n| n.starts_with("Stage_") && n.ends_with(".xpr"))
+        .collect();
+    names.sort();
+    let mut tot = 0usize;
+    for n in &names {
+        let bytes = std::fs::read(format!("{dir}/{n}")).unwrap();
+        let t0 = Instant::now();
+        let models = Xbg7Model::stage_models(&bytes);
+        let dt = t0.elapsed().as_millis();
+        tot += models.len();
+        println!("{n:16} {:>4} MB  {:>3} models  {:>5} ms", bytes.len()/1_000_000, models.len(), dt);
+    }
+    println!("TOTAL stage sub-models decoded: {tot}");
+}
+
+/// Quality audit: for a big stage, verify decoded models are real geometry
+/// (bounded bbox, low full-mesh degeneracy) rather than false anchors.
+#[test]
+#[ignore]
+fn stage_models_quality_audit() {
+    use sylpheed_formats::mesh::Xbg7Model;
+    let dir = std::env::var("SYLPHEED_RES3D").unwrap_or_else(|_| {
+        "/home/fabi/RE Project Sylpheed/sylph_extract/hidden/resource3d".to_string()
+    });
+    let bytes = std::fs::read(format!("{dir}/Stage_S07.xpr")).unwrap();
+    let models = Xbg7Model::stage_models(&bytes);
+    let (mut small, mut mid, mut huge, mut dupnames) = (0, 0, 0, 0);
+    let mut seen = std::collections::HashSet::new();
+    let mut worst_deg = 0.0f32;
+    for m in &models {
+        if !seen.insert(m.name.clone()) { dupnames += 1; }
+        let mut lo = [f32::MAX; 3]; let mut hi = [f32::MIN; 3];
+        let mut deg = 0usize; let mut tot = 0usize;
+        for sub in &m.meshes {
+            for p in &sub.positions { for a in 0..3 { lo[a]=lo[a].min(p[a]); hi[a]=hi[a].max(p[a]); } }
+            for tri in sub.indices.chunks_exact(3) {
+                let (a,b,c)=(tri[0] as usize,tri[1] as usize,tri[2] as usize);
+                if a>=sub.positions.len()||b>=sub.positions.len()||c>=sub.positions.len(){continue;}
+                let pa=sub.positions[a]; let pb=sub.positions[b]; let pc=sub.positions[c];
+                let u=[pb[0]-pa[0],pb[1]-pa[1],pb[2]-pa[2]];
+                let v=[pc[0]-pa[0],pc[1]-pa[1],pc[2]-pa[2]];
+                let cx=[u[1]*v[2]-u[2]*v[1],u[2]*v[0]-u[0]*v[2],u[0]*v[1]-u[1]*v[0]];
+                if 0.5*(cx[0]*cx[0]+cx[1]*cx[1]+cx[2]*cx[2]).sqrt()<1e-9 { deg+=1; }
+                tot+=1;
+            }
+        }
+        let ext = (hi[0]-lo[0]).max(hi[1]-lo[1]).max(hi[2]-lo[2]);
+        let df = if tot>0 { deg as f32/tot as f32 } else {1.0};
+        worst_deg = worst_deg.max(df);
+        if ext < 200.0 { small += 1; } else if ext < 5000.0 { mid += 1; } else { huge += 1; }
+    }
+    println!("S07: {} models | small(<200u)={} mid={} huge(>5k)={} | dup-names={} | worst full-mesh degeneracy={:.1}%",
+        models.len(), small, mid, huge, dupnames, worst_deg*100.0);
+    // Each XBG7 resource must anchor to a *distinct* block (no collisions), the
+    // bulk must be bounded-scale geometry, and none may be mostly-degenerate.
+    assert_eq!(dupnames, 0, "no two resources should anchor to the same block");
+    assert!(small + mid > models.len() * 9 / 10, "≥90% bounded-scale geometry");
+    assert!(huge < models.len() / 20, "few huge (skybox-plane) models");
+    assert!(worst_deg < 0.35, "no model should be mostly-degenerate");
+}

crates/sylpheed-viewer/src/camera.rs

@@ -32,6 +32,10 @@ pub struct OrbitCamera {
     pub orbit_sensitivity: f32,
     pub zoom_sensitivity: f32,
     pub pan_sensitivity: f32,
+    /// Zoom limits — set when framing a model so large scenes can be pulled back
+    /// far enough (the old fixed 0.5..50 cap trapped the camera inside big stages).
+    pub min_radius: f32,
+    pub max_radius: f32,
 }
 
 impl Default for OrbitCamera {
@@ -44,6 +48,8 @@ impl Default for OrbitCamera {
             orbit_sensitivity: 0.005,
             zoom_sensitivity: 0.3,
             pan_sensitivity: 0.003,
+            min_radius: 0.05,
+            max_radius: 500.0,
         }
     }
 }
@@ -54,19 +60,26 @@ fn spawn_camera(mut commands: Commands) {
 
     commands.spawn((
         Camera3d::default(),
+        // Wide near/far so both tiny weapons and multi-thousand-unit stages fit;
+        // the planes are re-scaled to the zoom distance each frame (below).
+        Projection::Perspective(PerspectiveProjection {
+            near: 0.05,
+            far: 100_000.0,
+            ..default()
+        }),
         transform,
         orbit,
     ));
 }
 
 fn orbit_camera(
-    mut query: Query<(&mut OrbitCamera, &mut Transform)>,
+    mut query: Query<(&mut OrbitCamera, &mut Transform, &mut Projection)>,
     mouse_buttons: Res<ButtonInput<MouseButton>>,
     keys: Res<ButtonInput<KeyCode>>,
     mut mouse_motion: EventReader<MouseMotion>,
     mut scroll: EventReader<MouseWheel>,
 ) {
-    let Ok((mut cam, mut transform)) = query.get_single_mut() else { return };
+    let Ok((mut cam, mut transform, mut projection)) = query.get_single_mut() else { return };
 
     let mut delta_motion = Vec2::ZERO;
     for ev in mouse_motion.read() {
@@ -99,13 +112,20 @@ fn orbit_camera(
 
     // Zoom (scroll)
     cam.radius -= scroll_delta * cam.zoom_sensitivity * cam.radius;
-    cam.radius = cam.radius.clamp(0.5, 50.0);
+    cam.radius = cam.radius.clamp(cam.min_radius, cam.max_radius);
 
     // Reset (R key)
     if keys.just_pressed(KeyCode::KeyR) {
         *cam = OrbitCamera::default();
     }
 
+    // Keep the clip planes proportional to the zoom distance so depth precision
+    // stays usable across scales (tiny prop → whole stage grid).
+    if let Projection::Perspective(p) = projection.as_mut() {
+        p.near = (cam.radius * 0.02).clamp(0.02, 50.0);
+        p.far = (cam.radius * 50.0).max(2000.0);
+    }
+
     *transform = orbit_transform(&cam);
 }
 

crates/sylpheed-viewer/src/iso_loader.rs

@@ -221,6 +221,21 @@ pub struct SkyboxPreview {
     pub info: String,
 }
 
+/// Marker for entities spawned to preview an XBG7 3D model, so they can all be
+/// despawned when a new file is selected.
+#[derive(Component)]
+pub struct PreviewModel;
+
+/// The currently-previewed XBG7 3D model. When `active`, the central panel is
+/// drawn transparent so the real Bevy scene (mesh + orbit camera) shows through.
+#[derive(Resource, Default)]
+pub struct ModelPreview {
+    pub active: bool,
+    pub name: String,
+    /// Summary line: sub-meshes / vertices / triangles / texture.
+    pub info: String,
+}
+
 /// The currently-open `.wmv` cutscene, shown in the central panel with a
 /// transport bar. Registered unconditionally (the decode/audio engine below is
 /// native-only); on wasm `active` stays false and `.wmv` shows the info panel.
@@ -459,6 +474,7 @@ impl Plugin for IsoLoaderPlugin {
             .init_resource::<TextPreview>()
             .init_resource::<PakView>()
             .init_resource::<SkyboxPreview>()
+            .init_resource::<ModelPreview>()
             .init_resource::<VideoPreview>();
 
         #[cfg(not(target_arch = "wasm32"))]
@@ -1423,6 +1439,378 @@ fn free_video(
     *video = VideoPreview::default();
 }
 
+/// Build Bevy meshes from a decoded [`Xbg7Model`], texture them with the model's
+/// albedo (`_col`) map, spawn them into the scene tagged [`PreviewModel`], and
+/// frame the orbit camera on the combined bounding box.
+#[cfg(not(target_arch = "wasm32"))]
+#[allow(clippy::too_many_arguments)]
+fn spawn_model_preview(
+    model: &sylpheed_formats::mesh::Xbg7Model,
+    xpr_bytes: &[u8],
+    commands: &mut Commands,
+    meshes: &mut Assets<Mesh>,
+    materials: &mut Assets<StandardMaterial>,
+    images: &mut Assets<Image>,
+    model_preview: &mut ModelPreview,
+    orbit: &mut Query<&mut crate::camera::OrbitCamera>,
+) {
+    use bevy::render::mesh::{Indices, PrimitiveTopology};
+    use sylpheed_formats::texture::X360Texture;
+
+    // ── Albedo texture: prefer the `_col` map, else the first texture. ──
+    let names = X360Texture::texture_names(xpr_bytes);
+    let col_idx = names
+        .iter()
+        .position(|n| n.ends_with("_col"))
+        .or(if names.is_empty() { None } else { Some(0) });
+    let (base_color_texture, tex_label) = match col_idx {
+        Some(i) => match X360Texture::from_xpr2_index(xpr_bytes, i)
+            .ok()
+            .and_then(|t| crate::asset_loader::x360_texture_to_bevy_image(t).ok())
+        {
+            Some(img) => (Some(images.add(img)), names[i].clone()),
+            None => (None, "none".to_string()),
+        },
+        None => (None, "none".to_string()),
+    };
+
+    let material = materials.add(StandardMaterial {
+        base_color: Color::srgb(0.8, 0.8, 0.82),
+        base_color_texture,
+        perceptual_roughness: 0.7,
+        metallic: 0.1,
+        // Xbox winding / our handedness may disagree — show both sides so a
+        // model is never invisible from the "back".
+        cull_mode: None,
+        double_sided: true,
+        ..default()
+    });
+
+    // ── Combined bounding box for camera framing. ──
+    let mut lo = Vec3::splat(f32::MAX);
+    let mut hi = Vec3::splat(f32::MIN);
+
+    // Parent entity so the whole model can be despawned / transformed as one.
+    let root = commands
+        .spawn((
+            PreviewModel,
+            Transform::default(),
+            Visibility::default(),
+            Name::new(model.name.clone()),
+        ))
+        .id();
+
+    let mut total_v = 0usize;
+    let mut total_t = 0usize;
+    for sub in &model.meshes {
+        if sub.positions.is_empty() || sub.indices.is_empty() {
+            continue;
+        }
+        total_v += sub.positions.len();
+        total_t += sub.indices.len() / 3;
+        for p in &sub.positions {
+            lo = lo.min(Vec3::from_array(*p));
+            hi = hi.max(Vec3::from_array(*p));
+        }
+
+        let positions: Vec<[f32; 3]> = sub.positions.clone();
+        let uvs: Vec<[f32; 2]> = if sub.uvs.len() == sub.positions.len() {
+            sub.uvs.clone()
+        } else {
+            vec![[0.0, 0.0]; sub.positions.len()]
+        };
+        // Prefer the model's own normals; fall back to computed smooth normals.
+        let normals = if sub.normals.len() == sub.positions.len() {
+            sub.normals.clone()
+        } else {
+            compute_smooth_normals(&positions, &sub.indices)
+        };
+
+        let mut mesh = Mesh::new(
+            PrimitiveTopology::TriangleList,
+            RenderAssetUsages::default(),
+        );
+        mesh.insert_attribute(Mesh::ATTRIBUTE_POSITION, positions);
+        mesh.insert_attribute(Mesh::ATTRIBUTE_NORMAL, normals);
+        mesh.insert_attribute(Mesh::ATTRIBUTE_UV_0, uvs);
+        mesh.insert_indices(Indices::U32(sub.indices.clone()));
+
+        let child = commands
+            .spawn((
+                PreviewModel,
+                Mesh3d(meshes.add(mesh)),
+                MeshMaterial3d(material.clone()),
+                Transform::default(),
+            ))
+            .id();
+        commands.entity(root).add_child(child);
+    }
+
+    // ── Frame the orbit camera on the model. ──
+    let center = if lo.x <= hi.x { (lo + hi) * 0.5 } else { Vec3::ZERO };
+    let extent = if lo.x <= hi.x {
+        (hi - lo).length().max(0.001)
+    } else {
+        5.0
+    };
+    if let Ok(mut cam) = orbit.get_single_mut() {
+        cam.focus = center;
+        cam.radius = extent * 1.4;
+        cam.min_radius = (extent * 0.02).max(0.02);
+        cam.max_radius = (extent * 20.0).max(50.0);
+    }
+
+    model_preview.active = true;
+    model_preview.name = model.name.clone();
+    model_preview.info = format!(
+        "XBG7 model · {} sub-mesh(es) · {} verts · {} tris · albedo: {}",
+        model.meshes.len(),
+        total_v,
+        total_t,
+        tex_label,
+    );
+}
+
+/// Spawn every decoded sub-model of a **stage** container as a side-by-side row
+/// (a "cast sheet"): the stage's enemy / prop models are separate objects with no
+/// recovered scene transforms, so each is recentred on its own origin and laid
+/// out along +X, spaced by its width. The orbit camera frames the whole row.
+#[cfg(not(target_arch = "wasm32"))]
+#[allow(clippy::too_many_arguments)]
+fn spawn_stage_models(
+    models: &[sylpheed_formats::mesh::Xbg7Model],
+    xpr_bytes: &[u8],
+    commands: &mut Commands,
+    meshes: &mut Assets<Mesh>,
+    materials: &mut Assets<StandardMaterial>,
+    images: &mut Assets<Image>,
+    model_preview: &mut ModelPreview,
+    orbit: &mut Query<&mut crate::camera::OrbitCamera>,
+) {
+    use bevy::render::mesh::{Indices, PrimitiveTopology};
+    use sylpheed_formats::texture::X360Texture;
+
+    // Per-sub-model albedo: match each model's name to its `_col` texture
+    // (e.g. `e003` → `e003_bdy_col`, `e005` → `e005_01_col`). Decode each once
+    // and cache by texture index; fall back to a neutral tint when no `_col`
+    // channel matches. Colours are the same "unverified" item as the 2D textures.
+    let tex_names = X360Texture::texture_names(xpr_bytes);
+    let neutral = materials.add(StandardMaterial {
+        base_color: Color::srgb(0.72, 0.74, 0.78),
+        perceptual_roughness: 0.65,
+        metallic: 0.1,
+        cull_mode: None,
+        double_sided: true,
+        ..default()
+    });
+    let mut tex_cache: std::collections::HashMap<usize, Handle<StandardMaterial>> =
+        std::collections::HashMap::new();
+    let material_for = |model_name: &str,
+                        materials: &mut Assets<StandardMaterial>,
+                        images: &mut Assets<Image>,
+                        cache: &mut std::collections::HashMap<usize, Handle<StandardMaterial>>|
+     -> Handle<StandardMaterial> {
+        // Core name = strip leading `_`/`rou_` and trailing `_l` LOD suffix.
+        let core = model_name
+            .trim_start_matches('_')
+            .trim_start_matches("rou_")
+            .trim_end_matches("_l");
+        let idx = tex_names.iter().position(|n| {
+            n.ends_with("_col") && (n.starts_with(core) || n.contains(core))
+        });
+        let Some(i) = idx else { return neutral.clone() };
+        if let Some(h) = cache.get(&i) {
+            return h.clone();
+        }
+        let handle = X360Texture::from_xpr2_index(xpr_bytes, i)
+            .ok()
+            .and_then(|t| crate::asset_loader::x360_texture_to_bevy_image(t).ok())
+            .map(|img| {
+                materials.add(StandardMaterial {
+                    base_color: Color::WHITE,
+                    base_color_texture: Some(images.add(img)),
+                    perceptual_roughness: 0.7,
+                    metallic: 0.1,
+                    cull_mode: None,
+                    double_sided: true,
+                    ..default()
+                })
+            })
+            .unwrap_or_else(|| neutral.clone());
+        cache.insert(i, handle.clone());
+        handle
+    };
+
+    let root = commands
+        .spawn((
+            PreviewModel,
+            Transform::default(),
+            Visibility::default(),
+            Name::new("stage"),
+        ))
+        .id();
+
+    // A stage holds up to ~450 sub-models at wildly different true scales (a
+    // 1-unit prop next to a 3000-unit structure) with no recovered scene
+    // transforms. Laying them end-to-end made a mile-wide strip the camera
+    // couldn't escape. Instead present a **thumbnail grid**: each model is
+    // recentred and uniformly scaled to a fixed cell, so all are equally visible
+    // and browsable regardless of native size. Grid spans the XY plane.
+    const CELL: f32 = 10.0; // target on-screen size of each thumbnail
+    const GAP: f32 = 4.0;
+    let pitch = CELL + GAP;
+
+    // Extent above which a sub-model is a skybox plane (300 k-unit cloud quad) or
+    // a stray-vertex mis-anchor rather than a real object — dropped from the grid.
+    const MAX_EXTENT: f32 = 20_000.0;
+    let mut drawn: Vec<&sylpheed_formats::mesh::Xbg7Model> = Vec::new();
+    for model in models {
+        let mut lo = Vec3::splat(f32::MAX);
+        let mut hi = Vec3::splat(f32::MIN);
+        let mut has_geo = false;
+        for sub in &model.meshes {
+            if sub.positions.is_empty() || sub.indices.is_empty() {
+                continue;
+            }
+            has_geo = true;
+            for p in &sub.positions {
+                lo = lo.min(Vec3::from_array(*p));
+                hi = hi.max(Vec3::from_array(*p));
+            }
+        }
+        if has_geo && (hi - lo).max_element() <= MAX_EXTENT {
+            drawn.push(model);
+        }
+    }
+    let cols = (drawn.len() as f32).sqrt().ceil().max(1.0) as usize;
+
+    let mut total_v = 0usize;
+    let mut total_t = 0usize;
+    for (i, model) in drawn.iter().enumerate() {
+        // Per-model bbox → recentre + uniform scale to the cell.
+        let mut lo = Vec3::splat(f32::MAX);
+        let mut hi = Vec3::splat(f32::MIN);
+        for sub in &model.meshes {
+            for p in &sub.positions {
+                lo = lo.min(Vec3::from_array(*p));
+                hi = hi.max(Vec3::from_array(*p));
+            }
+        }
+        if lo.x > hi.x {
+            continue;
+        }
+        let center = (lo + hi) * 0.5;
+        let extent = (hi - lo).max_element().max(1e-3);
+        let scale = CELL / extent;
+
+        // Grid cell centre (row-major, growing down in −Y).
+        let col = i % cols;
+        let row = i / cols;
+        let cell_pos = Vec3::new(col as f32 * pitch, -(row as f32) * pitch, 0.0);
+
+        // One entity per model carries the placement + normalising scale; its
+        // sub-meshes are recentred to the model's own origin underneath it.
+        let model_root = commands
+            .spawn((
+                PreviewModel,
+                Transform::from_translation(cell_pos).with_scale(Vec3::splat(scale)),
+                Visibility::default(),
+                Name::new(model.name.clone()),
+            ))
+            .id();
+        commands.entity(root).add_child(model_root);
+
+        let material = material_for(&model.name, materials, images, &mut tex_cache);
+
+        for sub in &model.meshes {
+            if sub.positions.is_empty() || sub.indices.is_empty() {
+                continue;
+            }
+            total_v += sub.positions.len();
+            total_t += sub.indices.len() / 3;
+
+            let positions: Vec<[f32; 3]> =
+                sub.positions.iter().map(|p| {
+                    [p[0] - center.x, p[1] - center.y, p[2] - center.z]
+                }).collect();
+            let uvs: Vec<[f32; 2]> = if sub.uvs.len() == sub.positions.len() {
+                sub.uvs.clone()
+            } else {
+                vec![[0.0, 0.0]; sub.positions.len()]
+            };
+            let normals = if sub.normals.len() == sub.positions.len() {
+                sub.normals.clone()
+            } else {
+                compute_smooth_normals(&positions, &sub.indices)
+            };
+
+            let mut mesh =
+                Mesh::new(PrimitiveTopology::TriangleList, RenderAssetUsages::default());
+            mesh.insert_attribute(Mesh::ATTRIBUTE_POSITION, positions);
+            mesh.insert_attribute(Mesh::ATTRIBUTE_NORMAL, normals);
+            mesh.insert_attribute(Mesh::ATTRIBUTE_UV_0, uvs);
+            mesh.insert_indices(Indices::U32(sub.indices.clone()));
+
+            let child = commands
+                .spawn((
+                    PreviewModel,
+                    Mesh3d(meshes.add(mesh)),
+                    MeshMaterial3d(material.clone()),
+                    Transform::default(),
+                ))
+                .id();
+            commands.entity(model_root).add_child(child);
+        }
+    }
+
+    // Frame the whole grid.
+    let rows = drawn.len().div_ceil(cols);
+    let grid_w = cols as f32 * pitch;
+    let grid_h = rows as f32 * pitch;
+    let center = Vec3::new(grid_w * 0.5 - pitch * 0.5, -(grid_h * 0.5 - pitch * 0.5), 0.0);
+    let diag = (grid_w * grid_w + grid_h * grid_h).sqrt().max(CELL);
+    if let Ok(mut cam) = orbit.get_single_mut() {
+        cam.focus = center;
+        cam.radius = diag * 0.75;
+        cam.min_radius = CELL * 0.05;
+        cam.max_radius = diag * 3.0;
+    }
+
+    model_preview.active = true;
+    model_preview.name = "stage".to_string();
+    model_preview.info = format!(
+        "Stage container · {} sub-models (thumbnail grid, each scaled to fit) · \
+         {} verts · {} tris (content-anchored; hero bodies with quantized layout \
+         are skipped)",
+        drawn.len(),
+        total_v,
+        total_t,
+    );
+}
+
+/// Per-vertex smooth normals = normalized sum of incident face normals.
+#[cfg(not(target_arch = "wasm32"))]
+fn compute_smooth_normals(positions: &[[f32; 3]], indices: &[u32]) -> Vec<[f32; 3]> {
+    let mut acc = vec![Vec3::ZERO; positions.len()];
+    for tri in indices.chunks_exact(3) {
+        let (a, b, c) = (tri[0] as usize, tri[1] as usize, tri[2] as usize);
+        if a >= positions.len() || b >= positions.len() || c >= positions.len() {
+            continue;
+        }
+        let pa = Vec3::from_array(positions[a]);
+        let pb = Vec3::from_array(positions[b]);
+        let pc = Vec3::from_array(positions[c]);
+        let n = (pb - pa).cross(pc - pa);
+        acc[a] += n;
+        acc[b] += n;
+        acc[c] += n;
+    }
+    acc.into_iter()
+        .map(|n| n.normalize_or_zero().to_array())
+        .map(|n| if n == [0.0, 0.0, 0.0] { [0.0, 1.0, 0.0] } else { n })
+        .collect()
+}
+
 /// Converts pending raw bytes into a Bevy `Image`, registers it with egui,
 /// and populates `TexturePreview` / `SkyboxPreview` / `FileInfo`.
 #[cfg(not(target_arch = "wasm32"))]
@@ -1437,6 +1825,12 @@ fn apply_loaded_texture(
     mut file_info: ResMut<FileInfo>,
     mut images: ResMut<Assets<Image>>,
     mut contexts: bevy_egui::EguiContexts,
+    mut commands: Commands,
+    mut meshes: ResMut<Assets<Mesh>>,
+    mut materials: ResMut<Assets<StandardMaterial>>,
+    mut model_preview: ResMut<ModelPreview>,
+    old_models: Query<Entity, With<PreviewModel>>,
+    mut orbit: Query<&mut crate::camera::OrbitCamera>,
 ) {
     if !pending.ready {
         return;
@@ -1456,6 +1850,12 @@ fn apply_loaded_texture(
     *text_preview = TextPreview::default();
     *pak_view = PakView::default();
 
+    // Despawn any previously-previewed 3D model.
+    for e in old_models.iter() {
+        commands.entity(e).despawn_recursive();
+    }
+    *model_preview = ModelPreview::default();
+
     // Populate FileInfo for the status / info panel.
     file_info.name = path
         .split('/')
@@ -1488,6 +1888,51 @@ fn apply_loaded_texture(
         return; // Not a texture — FileInfo is sufficient
     }
 
+    // 3D model? XPR2 containers that hold XBG7 geometry (ship / weapon / prop
+    // models under `hidden/resource3d/`) preview as an orbitable mesh, textured
+    // with their albedo (`_col`) map. Complex multi-stream body meshes decline
+    // cleanly (Err) and fall through to the 2D texture preview below.
+    // A container may decode as a single model (weapons / props, via the
+    // single-block layout) OR as a stage — a collection of many enemy / prop
+    // sub-models. A stage's *first* XBG7 is often a tiny dummy "root" node that
+    // the single-model path decodes into a degenerate cube, so we can't just try
+    // it first and stop. Decode both and keep whichever yields more geometry.
+    use sylpheed_formats::mesh::Xbg7Model;
+    let single = Xbg7Model::from_xpr2(&bytes)
+        .ok()
+        .filter(|m| !m.meshes.is_empty());
+    let single_verts = single.as_ref().map(|m| m.totals().0).unwrap_or(0);
+
+    let stage = Xbg7Model::stage_models(&bytes);
+    let stage_verts: usize = stage.iter().map(|m| m.totals().0).sum();
+
+    if !stage.is_empty() && stage_verts > single_verts {
+        spawn_stage_models(
+            &stage,
+            &bytes,
+            &mut commands,
+            &mut meshes,
+            &mut materials,
+            &mut images,
+            &mut model_preview,
+            &mut orbit,
+        );
+        return;
+    }
+    if let Some(model) = single {
+        spawn_model_preview(
+            &model,
+            &bytes,
+            &mut commands,
+            &mut meshes,
+            &mut materials,
+            &mut images,
+            &mut model_preview,
+            &mut orbit,
+        );
+        return;
+    }
+
     // World cubemap (`TXCM`) → decode all 6 faces and show a labelled grid.
     // Returns `None` for ordinary 2D textures, which fall through below.
     match sylpheed_formats::texture::X360Texture::cube_faces_from_xpr2(&bytes) {

crates/sylpheed-viewer/src/ui.rs

@@ -10,7 +10,7 @@ use bevy::prelude::*;
 use bevy_egui::{egui, EguiContexts};
 
 use crate::iso_loader::{
-    FileInfo, FileSelected, ImageRgba, IsoState, PakContent, PakView, RequestOpenDir,
+    FileInfo, FileSelected, ImageRgba, IsoState, ModelPreview, PakContent, PakView, RequestOpenDir,
     RequestOpenIso, SkyboxPreview, TextPreview, TexturePreview, VideoPreview, IsoLoaderSystemSet,
 };
 use crate::ViewerState;
@@ -132,6 +132,7 @@ fn draw_viewer_ui(
     text_preview: Res<TextPreview>,
     mut pak_view: ResMut<PakView>,
     skybox: Res<SkyboxPreview>,
+    model: Res<ModelPreview>,
     mut video: ResMut<VideoPreview>,
     file_info: Res<FileInfo>,
     mut open_iso_events: EventWriter<RequestOpenIso>,
@@ -251,7 +252,30 @@ fn draw_viewer_ui(
     });
 
     // ── Central area ──────────────────────────────────────────────────────
-    if browser.selected.is_some() {
+    if browser.selected.is_some() && model.active && !browser.loading {
+        // 3D model: draw the central panel TRANSPARENT so the real Bevy scene
+        // (mesh + orbit camera) shows through, with just an info overlay on top.
+        egui::CentralPanel::default()
+            .frame(egui::Frame::none())
+            .show(ctx, |ui| {
+                egui::Frame::none()
+                    .fill(egui::Color32::from_black_alpha(160))
+                    .inner_margin(egui::Margin::same(8.0))
+                    .rounding(egui::Rounding::same(4.0))
+                    .show(ui, |ui| {
+                        ui.heading(&model.name);
+                        ui.label(&model.info);
+                        ui.colored_label(
+                            egui::Color32::from_gray(160),
+                            "3D preview · LMB orbit · RMB pan · scroll zoom · R reset",
+                        );
+                        ui.colored_label(
+                            egui::Color32::from_gray(140),
+                            "position + normal + UV from the XBG7 vertex declaration",
+                        );
+                    });
+            });
+    } else if browser.selected.is_some() {
         egui::CentralPanel::default().show(ctx, |ui| {
             if browser.loading {
                 // A file was just clicked — show immediate feedback while the

docs/HANDOFF-stage-mesh-2026-07-12.md

@@ -0,0 +1,94 @@
+# Handoff — XBG7 stage-container meshes, unified layout, mesh renderer (2026-07-12)
+
+## Summary
+
+This session extended the XBG7 mesh work from single weapons/props to the big
+**stage containers**, fixed a long-standing weapon decode artifact, added a
+headless mesh renderer for self-verification, and reworked the viewer's stage
+presentation. Companion Xenia-Canary GPU instrumentation (the `log_draws`
+draw-logger used as the ground-truth oracle) is on the `instrument-current`
+branch of the Canary fork.
+
+## What landed (Reborn — `feature/ipfb-idxd-parser`)
+
+### 1. Stage containers decode — `Xbg7Model::stage_models`
+`hidden/resource3d/Stage_S*.xpr` are **collections of up to ~450 enemy/prop
+sub-models**, not single meshes. Each resource is a block
+`[12-byte header][index buffer][vertex buffer]`; the index count is the
+descriptor's `(index_bytes, index_count)` marker and the vertex count is the
+`u32` 32 bytes before it. The blocks are **scattered among the container's
+texture data with no stored offset**, so each is located by **content**: one
+`O(file)` pass per stride finds vertex-buffer starts (unit NORMAL at vertex +12
+whose previous stride slot isn't — a block boundary), then each resource is
+pinned to the candidate where its indices validate and produce non-degenerate,
+well-connected triangles. Fast (≤ ~4 s on a 70 MB stage), unambiguous.
+
+- **Coverage: ~4993 sub-models across the 22 stages** (content-anchored).
+- A **connectivity gate** (mean triangle edge ≤ 0.28× bbox diagonal) rejects
+  mis-anchored blocks that would render as spiky messes.
+
+### 2. Weapon layout fix (unified with stages)
+Weapons use the **same** `[12-byte header][index][vertex]` block layout. The old
+decoder read the index buffer 12 bytes too early → 2 leading degenerate
+triangles (the recurring "stray white triangle" artifact) + a lost tail of 6
+real indices. Now the header is skipped; vertex offsets are unchanged, so
+coverage stays 36/166 and the triangle lists are correct. Verified on
+wep_00/03/04 via the renderer.
+
+### 3. Headless mesh renderer — `sylpheed-cli mesh {info,render}`
+A software rasterizer (orthographic, z-buffered, two-sided shading) that writes a
+PNG. Lets anyone (including the assistant) verify recovered geometry without the
+GUI. `--only <substr>` filters sub-models; `--yaw/--pitch/--size/--row` control
+the view. Stage containers render as a normalised thumbnail grid.
+
+### 4. Viewer UX
+- Stage files show a **thumbnail grid**: each sub-model recentred + uniformly
+  scaled to a fixed cell (a 1-unit prop and a 3000-unit structure are equally
+  visible), skybox-plane / stray-anchor models (>20000 u) culled.
+- Stage sub-models are **textured** by matching name → `_col` albedo.
+- **Camera fixed**: zoom was hard-capped at 50 u (camera trapped inside big
+  stages); now zoom limits + clip planes scale to the framed model.
+- Dispatch decodes both single-model and stage paths and keeps whichever yields
+  more geometry (fixes stages previously showing the dummy-root "black cube").
+
+## State / how to verify
+
+```bash
+# formats + disc tests (needs the extracted disc)
+SYLPHEED_RES3D=/path/to/hidden/resource3d \
+  cargo test -p sylpheed-formats --test mesh_disc -- --ignored --nocapture
+
+# render any model / stage to a PNG
+cargo run --release -p sylpheed-cli -- mesh render \
+  .../resource3d/Stage_S10.xpr /tmp/s10.png --size 1000
+cargo run --release -p sylpheed-cli -- mesh render \
+  .../resource3d/rou_f001_wep_00.xpr /tmp/wep.png
+
+# viewer
+cargo run -p sylpheed-viewer     # File → Open Directory → the extracted disc
+```
+
+## Known limitations / next steps
+
+- **Multi-submesh main bodies with separate vertex pools** still mis-decode
+  (e.g. `Stage_S10` `e005`: a coherent fighter overlaid with wrong spike
+  triangles). `e003` works only because its two submeshes *share* one pool. Same
+  unsolved class as the **DeltaSaber hero body** (`f004`, still declined). Fix
+  needs per-submesh vertex-base RE. The LODs of these bodies (e.g. `e005_l`)
+  decode cleanly.
+- **The per-resource data offset is not stored** — blocks are found by content.
+  Reversing the container's allocation order would give exact offsets + 100%
+  coverage but is a larger effort.
+- **Texture colour correctness** remains the known "unverified" dynamic-RE item
+  (channel order / sRGB) for both 2D textures and model albedos.
+- The viewer's stage decode runs on the main thread (~10 s on a 70 MB stage in a
+  debug build); move off-thread if the load hitch matters.
+
+## Canary oracle (`instrument-current` on the Xenia-Canary fork)
+
+`command_processor.cc::LogDrawForRE` (cvar `log_draws`) dumps each distinct
+draw's primitive type, index buffer, vertex declaration, and sample vertex
+positions to `xenia_re_draws.log`. Used to confirm the XBG7 layout (triangle
+list, f32×3 position + f16×4 normal + f16×2 uv). Build with the memory-safe
+flags (`CC=clang-19 CXX=clang++-19 CMAKE_BUILD_PARALLEL_LEVEL=3 … -j 3`); run one
+emulator process at a time.

docs/re/INDEX.md

@@ -14,11 +14,12 @@ Promote to a prose `structures/…md` file when a format needs behavioural notes
 | name-hash (TOC keys) | ✅ | `sylpheed-formats/src/hash.rs` | Barrett-reduction hash; recovers original paths |
 | IDXD object/table | ✅ | `sylpheed-formats/src/idxd.rs` | self-describing; ship/weapon stats verified vs known values |
 | XPR2 texture + cubemap | 🟡 | `sylpheed-formats/src/texture.rs` | de-tile + A8R8G8B8; **colours unverified** (dynamic item) |
-| T8aD 2D texture | 🟡 | `sylpheed-formats/src/t8ad.rs` | ~85% decode; **colours unverified**; ~15% variants deferred |
+| T8aD 2D texture | 🟡 | `sylpheed-formats/src/t8ad.rs` | ~85% decode; **colours ✅ CONFIRMED** ([k8888](structures/texture-color-k8888.md)); ~15% variants deferred |
 | RATC bundle | 🟡 | `sylpheed-formats/src/ratc.rs` | child listing confirmed; one level deep |
 | LSTA sprite list | 🟡 | `sylpheed-formats/src/lsta.rs` | inline T8aD frames |
 | IXUD subtitle | 🟡 | `sylpheed-formats/src/ixud.rs` | timed cues; **movie↔track link unknown** (dynamic item) |
 | 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`). Quantized hero bodies (DeltaSaber `f004`) still declined |
 
 ## Functions / code paths
 

docs/re/structures/texture-color-k8888.md

@@ -21,18 +21,30 @@ is sampled as **raw UNORM bytes — apply no sRGB/gamma decode.**
 → *Implication:* if the viewer/engine applies an sRGB→linear (or linear→sRGB) step to these, that is
 wrong. Show the bytes as-is (raw RGBA8), gamma only where the game explicitly requests it.
 
-### ❔ HYPOTHESIS — exact source **channel order** (needs measurement, do not infer)
-Our decoders currently assume on-disk **A8R8G8B8** and reorder `[a,r,g,b] → [r,g,b,a]`. Canary
-proves the *host* target is R8G8B8A8 identity-swizzled, but the *guest→host* byte order also depends
-on the texture's runtime **endian** field (e.g. `8in32`), which is set by the game at fetch time and
-**cannot be read statically** from the format table. So whether our reorder matches is unverified.
-→ *What would confirm/refute it:* one **measured** decoded texel — either capture Canary's
-framebuffer over a texture with distinct R≠G≠B pixels of known colour, or patch Canary's already-
-present-but-unwired `TextureDump` (`src/xenia/gpu/texture_dump.cc`) to dump the decoded texture and
-diff bytes against our decode. **Measure, don't guess** — a symmetric colour (white logo) can't
-disambiguate channel order.
+### ✅ CONFIRMED — on-disk order is **ARGB** (alpha first); `[a,r,g,b]→[r,g,b,a]` is correct
+Settled by two independent lines that agree, without an emulator run:
+1. **Measured on the disc data** — across 789 real T8aD textures (GP_HANGAR_ARSENAL, GP_MAIN_GAME_E,
+   DefTables), the byte position that is most-often exactly `0xFF` (the opaque-alpha signature of UI
+   art) is **byte 0 in 767/789 (97%)** — tf1 385/0, tf2 382/12. So texel byte 0 = alpha ⇒ on-disk
+   layout is **A,R,G,B**. (An earlier bimodality test tied byte0/byte3 only because colour channels
+   are also `0x00`-heavy from black backgrounds; the `==0xFF` discriminator isolates alpha cleanly.)
+2. **Cross-check vs our Canary-validated renderer** — `xenia-rs`'s `decode_k8888_tiled` (the path
+   behind the user-confirmed M1 splash / M2 video colours) nets memory-`ARGB` → endian-swap →
+   `swap(0,2)` = `[A,R,G,B]→[R,G,B,A]`, identical to ours. Since that renderer was confirmed against
+   Canary's output, Canary's ground truth is transitively in this chain.
+
+Net: our channel order and the linear/UNORM interpretation are both **correct** for the ~85% RGBA
+variants. A live Canary framebuffer capture would be redundant confirmation, not a new signal.
+
+## Remaining / adjacent
+- **XPR2** shares the ARGB→RGBA reorder + the linear/UNORM finding, but its **tiling** (de-tile) is a
+  separate path; if XPR2 skyboxes still look odd it's tiling or viewer display-gamma, not channel order.
+- **Viewer display gamma:** decode is correct; if egui shows these too dark/bright it's how egui
+  interprets the RGBA8 (sRGB-encoded vs linear) at display time — a rendering nit, not a decode bug.
 
 ## Evidence log
-- `2026-07-11` — static read of Canary `src/xenia/gpu/vulkan/vulkan_texture_cache.cc` host-format
-  table + confirmed no `R8G8B8A8_SRGB` entry exists. → sRGB/UNORM claim `CONFIRMED`; channel-order
-  claim remains `HYPOTHESIS` pending a measured texel.
+- `2026-07-11` — static read of Canary `vulkan_texture_cache.cc` host-format table; no `R8G8B8A8_SRGB`
+  entry exists → sRGB/UNORM claim `CONFIRMED`.
+- `2026-07-11` — measured `==0xFF` alpha dominance over 789 disc T8aD textures (byte0 = alpha 97%) +
+  cross-checked vs `xenia-rs decode_k8888_tiled` (the Canary-validated M1/M2 path). → channel-order
+  claim promoted `HYPOTHESIS → CONFIRMED`.

docs/re/structures/xbg7-mesh.md

@@ -0,0 +1,183 @@
+# 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}`.
+- 2026-07-12 — `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**.