Cutscene voice: resolve movies via continuous-stream cue regions

Cracked how the game maps a movie to its voice track and reimplemented it,
fixing wrong/short voices for RT and hokyu cutscenes.

## Mechanism (RE'd from a Canary file-I/O trace, user-confirmed by ear)

The movie voices are ONE continuous XMA stream, physically chunked into
`<lang>\Movie\VOICE_*.slb` (and `\etc\VOICE_D_*.slb`) sound.pak TOC entries
whose boundaries do NOT align with the cutscene cues. A single cue routinely
spans two `.slb` chunks, so a `.slb` does not necessarily hold the track its
name claims (they are off-by-one vs the cues). Each cue ends at an inline
trailer descriptor `(sound_id:u32be, 0x11, …)` whose id repeats at +0x800.

Resolution chain (all derivable on-disc, no guest-code RE):
  1. movie -> cue token         (ADVERTISE_MOVIE manifest VOICETRACK)
  2. cue token -> sound-id       (master sound registry, tables.pak
                                  a2c8c185=eng / b04238e0=jpn, schema 0x13cb84ba)
  3. sound-id -> [start,end)      scan the stream for trailer(id) and its
                                  predecessor; cue audio = the bytes between,
                                  read continuously across chunk boundaries.

Validated: decoded voice length matches each movie within ~0.4s, including
cross-boundary cues (ADV/RT01A/RT01B/RT01C/S00A/S01A + the 5 bound hokyu).

## Changes

- sylpheed-formats/src/movie_voice.rs (new): registry_voice_ids(),
  find_descriptor(), find_descriptor_before() (gap-tolerant predecessor).
  Unit-tested.
- viewer iso_loader.rs: resolve_movie_voice_region() + decode_voice_region()
  (continuous read via existing read_segment_range + to_xma_riffs). Voice/
  audio requests now region-first; a `\Movie\` token that fails region stays
  SILENT (never plays the wrong off-by-one chunk). Removed the old
  movie_is_demo_based suppress hack.
- Anchor tries subdirs Movie/etc/Voice so hokyu `\etc\VOICE_D_*` resolve too.
- Examples resolve_all/validate_cues/hokyu_final document + validate the pipeline.

## Status

- RT / S / ADV cutscene voices: CORRECT (user-confirmed).
- 5 manifest-bound hokyu (VOICE_D_450-454): CORRECT (user-confirmed).

## OPEN (handoff)

13 of 18 hokyu movies have NO manifest VOICETRACK; the game reuses the 5
recordings across stages via a code rule. `hokyu_fallback_token()` currently
guesses by category (ship LS/DS x source carrier-A/tanker-H: LS/A->450,
LS/H->453, DS/A->452, DS/H->454). USER REPORTS THIS IS SOMETIMES WRONG.
- LS/carrier has two takes (450 from s02A, 451 from s09A); the early-vs-late
  split is unknown — unbound LS/A currently all default to 450.
- Definitive fix = Canary `--phase_a_fileio_only` file.read trace of an unbound
  hokyu (e.g. hokyu_LS_s03A) to see which VOICE_D the game actually streams,
  then encode the real rule. Same method that cracked the RT mapping.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Cargo.lock

@@ -1892,6 +1892,25 @@ dependencies = [
  "crossbeam-utils",
 ]
 
+[[package]]
+name = "crossbeam-deque"
+version = "0.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5181e0de7b61eb03a81e347d6dd8797bae9da5146707b51077e2d71a54ec0ceb"
+dependencies = [
+ "crossbeam-epoch",
+ "crossbeam-utils",
+]
+
+[[package]]
+name = "crossbeam-epoch"
+version = "0.9.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2d6914041f254d6e9176c01941b21115dcfb7089e55135a35411081bd106ef3f"
+dependencies = [
+ "crossbeam-utils",
+]
+
 [[package]]
 name = "crossbeam-utils"
 version = "0.8.21"
@@ -4107,6 +4126,26 @@ version = "0.6.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "20675572f6f24e9e76ef639bc5552774ed45f1c30e2951e1e99c59888861c539"
 
+[[package]]
+name = "rayon"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fb39b166781f92d482534ef4b4b1b2568f42613b53e5b6c160e24cfbfa30926d"
+dependencies = [
+ "either",
+ "rayon-core",
+]
+
+[[package]]
+name = "rayon-core"
+version = "1.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "22e18b0f0062d30d4230b2e85ff77fdfe4326feb054b9783a3460d8435c8ab91"
+dependencies = [
+ "crossbeam-deque",
+ "crossbeam-utils",
+]
+
 [[package]]
 name = "redox_syscall"
 version = "0.4.1"
@@ -4587,6 +4626,7 @@ dependencies = [
  "binrw",
  "flate2",
  "futures",
+ "rayon",
  "serde",
  "serde_json",
  "thiserror 2.0.18",

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)]
@@ -116,6 +122,51 @@ enum PakCommands {
         /// Entry name-hash, e.g. `0x7c96296c`
         hash: String,
     },
+    /// Decode every T8aD texture in the pak (direct, RATC-nested, and LSTA
+    /// frames) to PNG — our decoder's output, for A/B against the running game.
+    Textures {
+        /// Path to the `.pak` index
+        pak: PathBuf,
+        /// Output directory for the PNGs (created if missing)
+        output: PathBuf,
+        /// Print per-texture dimensions + tile count.
+        #[arg(long)]
+        verbose: bool,
+    },
+}
+
+#[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,
+        /// Camera distance multiplier (1.0 = framed; <1 zooms in, >1 out)
+        #[arg(long, default_value_t = 1.0)]
+        dist: 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)]
@@ -155,9 +206,18 @@ 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, dist, row, only } => {
+                cmd_mesh_render(&file, &output, size, yaw, pitch, dist, 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),
+            PakCommands::Textures { pak, output, verbose } => {
+                cmd_pak_textures(&pak, &output, verbose)
+            }
         },
     }
 }
@@ -357,6 +417,462 @@ 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::{count_xbg7, Xbg7Model};
+    // Route by container kind, NOT by whichever decoder yields more verts (that
+    // old heuristic let `stage_models`' content-anchoring win on single-model
+    // files, fabricating phantom / duplicate / mis-anchored blocks). A file with
+    // one XBG7 resource is a single model (weapon / prop) → use only the
+    // validated records-based list decode; content-anchoring a single-model file
+    // invents geometry. Many XBG7 resources → a Stage_* collection → anchor them.
+    if count_xbg7(bytes) > 1 {
+        // Multi-resource Stage_* collection → content-anchor every resource
+        // (each anchor now gated by stored-normal agreement).
+        Xbg7Model::stage_models(bytes)
+    } else if let Some(m) = Xbg7Model::from_xpr2(bytes)
+        .ok()
+        .filter(|m| !m.meshes.is_empty())
+    {
+        // Single model the records-based list decode carved (authoritative).
+        vec![m]
+    } else {
+        // Single model from_xpr2 couldn't locate (its sequential carve missed
+        // the block) — fall back to content-anchoring, which finds it by shape.
+        // Use a STRICT winding-consistency gate (0.85): on a single-model file a
+        // mis-anchor is an obvious phantom / spike-mess and must be declined,
+        // unlike the large stage corpus which keeps the ungated path.
+        Xbg7Model::anchor_models(bytes, 0.85)
+    }
+}
+
+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]
+        );
+        // Per-sub-mesh integrity diagnostics: degenerate triangles (a zero-area
+        // "hole"), vertices referenced by no triangle (dropped geometry), and the
+        // referenced index range vs the vertex count (short/over reads).
+        for (si, sub) in m.meshes.iter().enumerate() {
+            let nv = sub.positions.len();
+            let mut referenced = vec![false; nv];
+            let (mut degen, mut oob, mut imax) = (0usize, 0usize, 0u32);
+            for tri in sub.indices.chunks_exact(3) {
+                let (a, b, c) = (tri[0], tri[1], tri[2]);
+                imax = imax.max(a).max(b).max(c);
+                if a == b || b == c || a == c {
+                    degen += 1;
+                }
+                for &i in tri {
+                    if (i as usize) < nv {
+                        referenced[i as usize] = true;
+                    } else {
+                        oob += 1;
+                    }
+                }
+            }
+            let unref = referenced.iter().filter(|&&r| !r).count();
+            // Spanning triangles: longest edge ≫ the median (strip-junction spikes).
+            let edge = |a: u32, b: u32| {
+                let (p, q) = (sub.positions[a as usize], sub.positions[b as usize]);
+                ((p[0] - q[0]).powi(2) + (p[1] - q[1]).powi(2) + (p[2] - q[2]).powi(2)).sqrt()
+            };
+            let mut maxedges: Vec<f32> = sub
+                .indices
+                .chunks_exact(3)
+                .map(|t| edge(t[0], t[1]).max(edge(t[1], t[2])).max(edge(t[0], t[2])))
+                .collect();
+            maxedges.sort_by(|a, b| a.partial_cmp(b).unwrap());
+            let median = maxedges.get(maxedges.len() / 2).copied().unwrap_or(1.0).max(1e-6);
+            let spanning = maxedges.iter().filter(|&&e| e > 6.0 * median).count();
+            println!(
+                "      sub{si}: {nv} v, {} tris | degenerate {degen}, unref-verts {unref}, spanning>6×med {spanning}, idx_max {imax}/{}{}",
+                sub.indices.len() / 3,
+                nv.saturating_sub(1),
+                if oob > 0 { format!(", OOB {oob}") } else { String::new() },
+            );
+            // XDUMPVERT=1 → print the first few vertex positions per sub-mesh, for
+            // content-matching a decoded sub-mesh against the GPU draw log.
+            if std::env::var("XDUMPVERT").is_ok() {
+                let mut lo = [f32::MAX; 3];
+                let mut hi = [f32::MIN; 3];
+                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!(
+                    "           bbox X[{:.2}..{:.2}] Y[{:.2}..{:.2}] Z[{:.2}..{:.2}]  ctr({:.2},{:.2},{:.2})",
+                    lo[0], hi[0], lo[1], hi[1], lo[2], hi[2],
+                    (lo[0]+hi[0])/2.0, (lo[1]+hi[1])/2.0, (lo[2]+hi[2])/2.0
+                );
+            }
+        }
+    }
+    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,
+    dist: 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 {
+        // Prefer an exact name match (e.g. `f001` for the neutral ship pose,
+        // excluding the `_rou_f001_mnv*` animation poses that also *contain*
+        // "f001"); fall back to substring when nothing matches exactly.
+        if models.iter().any(|m| m.name == *sub) {
+            models.retain(|m| m.name == *sub);
+        } else {
+            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;
+    // XMIRROR=x|y|z → negate that axis, to test an Xbox(LH)→Bevy(RH) handedness
+    // flip against reference screenshots.
+    let mirror: [f32; 3] = match std::env::var("XMIRROR").ok().as_deref() {
+        Some("x") => [-1.0, 1.0, 1.0],
+        Some("y") => [1.0, -1.0, 1.0],
+        Some("z") => [1.0, 1.0, -1.0],
+        _ => [1.0, 1.0, 1.0],
+    };
+    let mut tris: Vec<[[f32; 3]; 3]> = Vec::new();
+    // XCOLORSUB=1 tints each sub-mesh a distinct colour (to see which sub is
+    // which part / where the "extra fin" comes from). Parallel to `tris`.
+    let color_sub = std::env::var("XCOLORSUB").is_ok();
+    // XONLYSUB=N renders only the N-th global sub-mesh (to isolate one part).
+    let only_sub: Option<usize> = std::env::var("XONLYSUB").ok().and_then(|s| s.parse().ok());
+    let mut tints: Vec<[f32; 3]> = Vec::new();
+    const PALETTE: [[f32; 3]; 8] = [
+        [1.0, 1.0, 1.0],   // sub0 body = white
+        [1.0, 0.35, 0.35], // sub1 red
+        [0.35, 1.0, 0.35], // sub2 green
+        [0.4, 0.55, 1.0],  // sub3 blue
+        [1.0, 0.9, 0.3],   // sub4 yellow
+        [1.0, 0.5, 1.0],   // sub5 magenta
+        [0.3, 1.0, 1.0],   // sub6 cyan
+        [1.0, 0.6, 0.2],   // sub7 orange
+    ];
+    let mut sub_gi = 0usize;
+    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])
+        };
+        // XNODEXFORM=1 applies the XBG7 scene-graph node placement (fins move to
+        // the tail) — to verify the transforms recovered from the graph.
+        let placements = if std::env::var("XNODEXFORM").is_ok() {
+            sylpheed_formats::mesh::node_transforms(&bytes, &m.name)
+        } else {
+            Vec::new()
+        };
+        for (sub_local, sub) in m.meshes.iter().enumerate() {
+            // Every scene-graph instance that draws this sub-mesh (mirrored fin
+            // pair, L/R winglets…); `None` = no graph placement → identity.
+            let mine: Vec<Option<&sylpheed_formats::mesh::NodePlacement>> = {
+                let v: Vec<_> = placements
+                    .iter()
+                    .filter(|p| p.sub_index == sub_local)
+                    .map(Some)
+                    .collect();
+                if v.is_empty() {
+                    vec![None]
+                } else {
+                    v
+                }
+            };
+            // Sub-mesh indices are a triangle list (the decoder has already
+            // expanded the file's triangle strips).
+            let n = sub.positions.len();
+            // XSPANONLY=1 renders ONLY long-edge ("spanning") triangles; XSPANHIDE=1
+            // renders everything EXCEPT them — to see whether the flagged spanning
+            // triangles are real geometry or decode artifacts (phantom sheets).
+            let span_only = std::env::var("XSPANONLY").is_ok();
+            let span_hide = std::env::var("XSPANHIDE").is_ok();
+            let med = {
+                let mut e: Vec<f32> = sub
+                    .indices
+                    .chunks_exact(3)
+                    .filter(|t| (t[0] as usize) < n && (t[1] as usize) < n && (t[2] as usize) < n)
+                    .map(|t| {
+                        let d = |a: u32, b: u32| {
+                            let (p, q) = (sub.positions[a as usize], sub.positions[b as usize]);
+                            ((p[0] - q[0]).powi(2) + (p[1] - q[1]).powi(2) + (p[2] - q[2]).powi(2))
+                                .sqrt()
+                        };
+                        d(t[0], t[1]).max(d(t[1], t[2])).max(d(t[0], t[2]))
+                    })
+                    .collect();
+                e.sort_by(|a, b| a.partial_cmp(b).unwrap());
+                e.get(e.len() / 2).copied().unwrap_or(1.0).max(1e-6)
+            };
+            if let Some(want) = only_sub {
+                if sub_gi != want {
+                    sub_gi += 1;
+                    continue;
+                }
+            }
+            let tint = if color_sub {
+                PALETTE[sub_gi % PALETTE.len()]
+            } else {
+                [1.0, 1.0, 1.0]
+            };
+            for place in &mine {
+                let f = |i: usize| {
+                    let p = place.map(|pl| pl.apply(sub.positions[i])).unwrap_or(sub.positions[i]);
+                    [
+                        (p[0] - center[0]) * scale * mirror[0] + cell[0],
+                        (p[1] - center[1]) * scale * mirror[1] + cell[1],
+                        (p[2] - center[2]) * scale * mirror[2] + cell[2],
+                    ]
+                };
+                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 < n && b < n && c < n {
+                        if span_only || span_hide {
+                            let d = |i: usize, j: usize| {
+                                let (p, q) = (sub.positions[i], sub.positions[j]);
+                                ((p[0] - q[0]).powi(2)
+                                    + (p[1] - q[1]).powi(2)
+                                    + (p[2] - q[2]).powi(2))
+                                .sqrt()
+                            };
+                            let spanning = d(a, b).max(d(b, c)).max(d(a, c)) > 6.0 * med;
+                            if span_only && !spanning {
+                                continue;
+                            }
+                            if span_hide && spanning {
+                                continue;
+                            }
+                        }
+                        tris.push([f(a), f(b), f(c)]);
+                        tints.push(tint);
+                    }
+                }
+            }
+            sub_gi += 1;
+        }
+    }
+    if tris.is_empty() {
+        anyhow::bail!("no triangles to render");
+    }
+
+    let rgba = rasterize(&tris, &tints, size, yaw, pitch, dist);
+    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]],
+    tints: &[[f32; 3]],
+    size: u32,
+    yaw_deg: f32,
+    pitch_deg: f32,
+    dist: 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 * dist.max(1e-3));
+    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 (ti, t) in tris.iter().enumerate() {
+        let tint = tints.get(ti).copied().unwrap_or([1.0, 1.0, 1.0]);
+        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 as f32 * tint[0]).min(255.0) as u8;
+                    color[idx * 4 + 1] = (shade as f32 * tint[1]).min(255.0) as u8;
+                    color[idx * 4 + 2] = (shade as f32 * tint[2] * 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
@@ -523,3 +1039,179 @@ fn cmd_pak_dump(pak: &Path, hash_str: &str) -> Result<()> {
     }
     Ok(())
 }
+
+// ── pak textures ─────────────────────────────────────────────────────────────
+
+/// Turn a child name into a filesystem-safe fragment.
+fn safe_name(s: &str) -> String {
+    s.chars()
+        .map(|c| {
+            if c.is_ascii_alphanumeric() || matches!(c, '.' | '_' | '-') {
+                c
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
+#[derive(Default)]
+struct TexStats {
+    written: usize,
+    skipped: usize,
+}
+
+fn be32_at(b: &[u8], off: usize) -> u32 {
+    u32::from_be_bytes([b[off], b[off + 1], b[off + 2], b[off + 3]])
+}
+
+/// Decode one T8aD slice (whose first bytes are the magic) to PNG. With
+/// `verbose`, prints its dimensions + tile count; the `XDUMPHDR` env var dumps
+/// the raw header (base + offset table) for format RE.
+fn emit_t8ad(
+    slice: &[u8],
+    hash: u32,
+    stem: &str,
+    output: &Path,
+    verbose: bool,
+    stats: &mut TexStats,
+) -> Result<()> {
+    use sylpheed_formats::t8ad;
+
+    if !t8ad::is_t8ad(slice) || slice.len() < 0x40 {
+        return Ok(());
+    }
+    let (w, h, tiles) = (
+        be32_at(slice, 0x14),
+        be32_at(slice, 0x18),
+        be32_at(slice, 0x1c),
+    );
+
+    // Debug: dump the header — 44-byte base + the `tiles`-entry u32 offset table.
+    if std::env::var("XDUMPHDR").is_ok() {
+        println!("\n{stem}  {w}x{h}  tiles={tiles}");
+        print!("  base[0x00..0x2c]:");
+        for i in (0..44).step_by(4) {
+            print!(" {:08x}", be32_at(slice, i));
+        }
+        print!("\n  offsets:");
+        for t in 0..(tiles as usize).min(64) {
+            if 0x2c + t * 4 + 4 <= slice.len() {
+                print!(" {}", be32_at(slice, 0x2c + t * 4));
+            }
+        }
+        println!();
+        return Ok(());
+    }
+
+    if verbose {
+        println!("  {hash:08x} {stem:<34} {w:>4}x{h:<4}  tiles {tiles}");
+    }
+    match t8ad::parse(slice) {
+        Some(img) => {
+            let out =
+                output.join(format!("{hash:08x}_{stem}_{}x{}.png", img.width, img.height));
+            image::save_buffer(
+                &out,
+                &img.rgba,
+                img.width,
+                img.height,
+                image::ExtendedColorType::Rgba8,
+            )
+            .with_context(|| format!("writing PNG {}", out.display()))?;
+            stats.written += 1;
+        }
+        None => stats.skipped += 1,
+    }
+    Ok(())
+}
+
+/// Decode every T8aD in a pak (direct entries, RATC-nested children, LSTA
+/// frames) to PNG — our decoder's exact output — for A/B against the game.
+fn cmd_pak_textures(pak: &Path, output: &Path, verbose: bool) -> Result<()> {
+    use sylpheed_formats::{lsta, ratc, t8ad};
+
+    let arc = PakArchive::open(pak).with_context(|| format!("opening {}", pak.display()))?;
+    std::fs::create_dir_all(output)
+        .with_context(|| format!("creating {}", output.display()))?;
+
+    println!(
+        "{} {} → {}",
+        "Textures".green().bold(),
+        pak.display().to_string().cyan(),
+        output.display().to_string().cyan(),
+    );
+
+    let mut stats = TexStats::default();
+
+    for e in arc.entries() {
+        let payload = match arc.read(e) {
+            Ok(p) => p,
+            Err(_) => continue,
+        };
+        let hash = e.name_hash;
+
+        // Direct T8aD entry.
+        if t8ad::is_t8ad(&payload) {
+            emit_t8ad(&payload, hash, "direct", output, verbose, &mut stats)?;
+            continue;
+        }
+
+        // LSTA sprite list = N inline T8aD frames (walk by magic, emit each).
+        if lsta::is_lsta(&payload) {
+            let mut off = 0usize;
+            let mut idx = 0usize;
+            while let Some(pos) = payload[off..]
+                .windows(4)
+                .position(|w| w == &t8ad::T8AD_MAGIC)
+            {
+                let start = off + pos;
+                let next = payload[start + 4..]
+                    .windows(4)
+                    .position(|w| w == &t8ad::T8AD_MAGIC)
+                    .map(|p| start + 4 + p)
+                    .unwrap_or(payload.len());
+                emit_t8ad(
+                    &payload[start..next],
+                    hash,
+                    &format!("lsta{idx:03}"),
+                    output,
+                    verbose,
+                    &mut stats,
+                )?;
+                idx += 1;
+                off = next;
+            }
+            continue;
+        }
+
+        // RATC bundle: decode its T8aD children (named, e.g. `foo.t32`).
+        if ratc::is_ratc(&payload) {
+            if let Some(children) = ratc::parse(&payload) {
+                for (i, child) in children.iter().enumerate() {
+                    if child.kind != "T8aD" {
+                        continue;
+                    }
+                    let end = (child.offset + child.size).min(payload.len());
+                    if child.offset >= end {
+                        continue;
+                    }
+                    let stem = if child.name.is_empty() {
+                        format!("child{i:03}")
+                    } else {
+                        safe_name(&child.name)
+                    };
+                    emit_t8ad(&payload[child.offset..end], hash, &stem, output, verbose, &mut stats)?;
+                }
+            }
+            continue;
+        }
+    }
+
+    println!(
+        "\n  {} PNG(s) written, {} undecodable (non-tilecount variants — likely DXT)",
+        stats.written.to_string().green(),
+        stats.skipped.to_string().yellow(),
+    );
+    Ok(())
+}

crates/sylpheed-formats/Cargo.toml

@@ -19,5 +19,11 @@ tracing   = { workspace = true }
 serde     = { workspace = true }
 serde_json = { workspace = true }
 
+# Native-only: parallelise the per-resource XBG7 stage anchoring (hundreds of
+# independent sub-models per container). rayon needs OS threads, so it is not
+# pulled in for the wasm32 build, which falls back to a sequential decode.
+[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
+rayon = "1"
+
 [dev-dependencies]
 tokio = { workspace = true }

crates/sylpheed-formats/examples/hokyu_final.rs

@@ -0,0 +1,36 @@
+use sylpheed_formats::{hash::name_hash, movie_manifest, movie_voice, slb, PakArchive};
+use std::fs;use std::io::{Read,Seek,SeekFrom};use std::process::Command;
+fn rg(disc:&str,g:u64,n:usize)->Vec<u8>{let mut segs=vec![];let mut cum=0u64;for i in 0..5{let p=format!("{disc}/dat/sound.p{i:02}");if let Ok(m)=fs::metadata(&p){segs.push((cum,m.len(),p));cum+=m.len();}}let mut out=vec![];let(mut need,mut pos)=(n,g);for(base,len,path)in &segs{if need==0||pos>=base+len||pos<*base{continue;}let local=pos-base;let take=need.min((len-local)as usize);let mut f=fs::File::open(path).unwrap();f.seek(SeekFrom::Start(local)).unwrap();let mut b=vec![0u8;take];f.read_exact(&mut b).unwrap();out.extend_from_slice(&b);need-=take;pos+=take as u64;}out}
+fn contains(h:&[u8],n:&[u8])->bool{h.windows(n.len()).any(|w|w==n)}
+fn dur(w:&str)->String{let o=Command::new("ffprobe").args(["-v","error","-show_entries","format=duration","-of","default=nw=1:nk=1",w]).output().unwrap();String::from_utf8_lossy(&o.stdout).trim().to_string()}
+fn hokyu_fallback(m:&str)->Option<String>{if !m.starts_with("hokyu_"){return None}let ls=m.contains("_LS_");let ds=m.contains("_DS_");let a=m.ends_with('A');let h=m.ends_with('H');Some(match(ls,ds,a,h){(true,_,true,_)=>"VOICE_D_450",(true,_,_,true)=>"VOICE_D_453",(_,true,true,_)=>"VOICE_D_452",(_,true,_,true)=>"VOICE_D_454",_=>return None}.into())}
+fn main(){
+    let disc=std::env::var("SYLPHEED_DISC").unwrap();
+    let tpak=PakArchive::open(format!("{disc}/dat/tables.pak")).unwrap();
+    let manifest=tpak.entries().iter().find_map(|e|tpak.read(e).ok().filter(|b|movie_manifest::is_manifest(b))).unwrap();
+    let code="eng";
+    let registry=tpak.entries().iter().find_map(|e|tpak.read(e).ok().filter(|b|contains(b,format!("{code}\\Movie\\VOICE_ADV.slb").as_bytes()))).unwrap();
+    let ids=movie_voice::registry_voice_ids(&registry);
+    let stoc=fs::read(format!("{disc}/dat/sound.pak")).unwrap();
+    let entries=PakArchive::parse_toc(&stoc).unwrap();
+    let hokyu:Vec<String>=movie_manifest::parse(&manifest).into_iter().filter(|m|m.movie.starts_with("hokyu_")).map(|m|m.movie).collect();
+    for movie in &hokyu{
+        let bound=movie_manifest::voice_token(&manifest,movie);
+        let token=bound.clone().or_else(||hokyu_fallback(movie));
+        let Some(token)=token else{println!("{movie:16} NO TOKEN"); continue};
+        let src = if bound.is_some(){"bound"}else{"fallback"};
+        let Some(&id)=ids.get(&token) else{println!("{movie:16} {token} not in registry");continue};
+        let Some(anchor)=["Movie","etc","Voice"].iter().find_map(|d|{let h=name_hash(&format!("{code}\\{d}\\{token}.slb"));entries.binary_search_by_key(&h,|e|e.name_hash).ok().map(|i|entries[i].offset as u64)}) else{println!("{movie:16} no anchor");continue};
+        let win_start=(anchor.saturating_sub(2*1024*1024))&!3;
+        let window=rg(&disc,win_start,8*1024*1024);
+        let Some(el)=movie_voice::find_descriptor(&window,id) else{println!("{movie:16} desc not found");continue};
+        let end=win_start+el as u64;
+        let start=movie_voice::find_descriptor(&window,id.wrapping_sub(1)).or_else(||movie_voice::find_descriptor_before(&window,el)).map(|o|win_start+o as u64).filter(|&s|s<end&&end-s<1_500_000).unwrap_or(anchor);
+        let region=rg(&disc,start,(end-start)as usize);
+        let mut riffs=slb::to_xma_riffs(&region); if riffs.is_empty(){riffs=slb::to_xma_riff_best(&region).into_iter().collect();}
+        let mut d="?".into();
+        if let Some(r)=riffs.first(){let xp=format!("/tmp/hf_{movie}.xma.wav");let wp=format!("/tmp/hf_{movie}.wav");fs::write(&xp,r).unwrap();let _=Command::new("ffmpeg").args(["-hide_banner","-v","error","-y","-i",&xp,&wp]).status();d=dur(&wp);}
+        let mv=dur(&format!("{disc}/dat/movie/{movie}.wmv"));
+        println!("{movie:16} {src:8} {token:12} id={id} voice={d}s movie={mv}s");
+    }
+}

crates/sylpheed-formats/examples/manifest_map.rs

@@ -0,0 +1,46 @@
+//! Validation: dump the parsed mission/phase cutscene map from the real manifest.
+use sylpheed_formats::movie_manifest::{self, MovieKind};
+use sylpheed_formats::pak::PakArchive;
+
+fn main() {
+    let disc = std::env::var("SYLPHEED_DISC").expect("set SYLPHEED_DISC");
+    let arc = PakArchive::open(format!("{disc}/dat/tables.pak")).unwrap();
+    let man = arc
+        .entries()
+        .iter()
+        .find_map(|e| arc.read(e).ok().filter(|b| movie_manifest::is_manifest(b)))
+        .expect("manifest");
+    let entries = movie_manifest::parse(&man);
+    let mut counts = [0usize; 5];
+    for e in &entries {
+        counts[match e.kind {
+            MovieKind::System => 0,
+            MovieKind::Intro => 1,
+            MovieKind::Phase => 2,
+            MovieKind::PhaseEnd => 3,
+            MovieKind::Supply => 4,
+        }] += 1;
+        let m = e.mission.map(|x| x.to_string()).unwrap_or_default();
+        let p = e.phase.map(|x| x.to_string()).unwrap_or_default();
+        println!(
+            "{:<22} {:<9} m={:<2} p={:<1} {:<16} {:<14} tel={:<14} sub={}",
+            e.slot,
+            format!("{:?}", e.kind),
+            m,
+            p,
+            e.movie,
+            e.voice_token.as_deref().unwrap_or("-"),
+            e.telop.as_deref().unwrap_or("-"),
+            e.subtitle.as_deref().unwrap_or("-"),
+        );
+    }
+    eprintln!(
+        "TOTAL {} entries — System={} Intro={} Phase={} PhaseEnd={} Supply={}",
+        entries.len(),
+        counts[0],
+        counts[1],
+        counts[2],
+        counts[3],
+        counts[4]
+    );
+}

crates/sylpheed-formats/examples/resolve_all.rs

@@ -0,0 +1,36 @@
+use sylpheed_formats::{hash::name_hash, movie_manifest, movie_voice, slb, PakArchive};
+use std::fs;use std::io::{Read,Seek,SeekFrom};use std::process::Command;
+fn rg(disc:&str,g:u64,n:usize)->Vec<u8>{let mut segs=vec![];let mut cum=0u64;for i in 0..5{let p=format!("{disc}/dat/sound.p{i:02}");if let Ok(m)=fs::metadata(&p){segs.push((cum,m.len(),p));cum+=m.len();}}let mut out=vec![];let(mut need,mut pos)=(n,g);for(base,len,path)in &segs{if need==0||pos>=base+len||pos<*base{continue;}let local=pos-base;let take=need.min((len-local)as usize);let mut f=fs::File::open(path).unwrap();f.seek(SeekFrom::Start(local)).unwrap();let mut b=vec![0u8;take];f.read_exact(&mut b).unwrap();out.extend_from_slice(&b);need-=take;pos+=take as u64;}out}
+fn contains(h:&[u8],n:&[u8])->bool{h.windows(n.len()).any(|w|w==n)}
+fn dur(w:&str)->String{let o=Command::new("ffprobe").args(["-v","error","-show_entries","format=duration","-of","default=nw=1:nk=1",w]).output().unwrap();String::from_utf8_lossy(&o.stdout).trim().to_string()}
+fn main(){
+    let disc=std::env::var("SYLPHEED_DISC").unwrap();
+    let tpak=PakArchive::open(format!("{disc}/dat/tables.pak")).unwrap();
+    let manifest=tpak.entries().iter().find_map(|e|tpak.read(e).ok().filter(|b|movie_manifest::is_manifest(b))).unwrap();
+    let code="eng";
+    let registry=tpak.entries().iter().find_map(|e|tpak.read(e).ok().filter(|b|contains(b,format!("{code}\\Movie\\VOICE_ADV.slb").as_bytes()))).unwrap();
+    let ids=movie_voice::registry_voice_ids(&registry);
+    let stoc=fs::read(format!("{disc}/dat/sound.pak")).unwrap();
+    let entries=PakArchive::parse_toc(&stoc).unwrap();
+    for movie in ["ADV","RT01A","RT01B","RT01C_1","S00A","S01A","hokyu_LS_s02A","hokyu_LS_s09A","hokyu_LS_s02H","hokyu_DS_s07H"]{
+        let Some(token)=movie_manifest::voice_token(&manifest,movie) else { println!("{movie:16} no token"); continue };
+        let Some(&id)=ids.get(&token) else { println!("{movie:16} token {token} not in registry"); continue };
+        let Some(anchor)=["Movie","etc","Voice"].iter().find_map(|d|{let h=name_hash(&format!("{code}\\{d}\\{token}.slb"));entries.binary_search_by_key(&h,|e|e.name_hash).ok().map(|i|entries[i].offset as u64)}) else { println!("{movie:16} no anchor for {token}"); continue };
+        let win_start=(anchor.saturating_sub(2*1024*1024))&!3;
+        let window=rg(&disc,win_start,8*1024*1024);
+        let Some(end_local)=movie_voice::find_descriptor(&window,id) else { println!("{movie:16} id {id}: desc NOT FOUND"); continue };
+        let end=win_start+end_local as u64;
+        let start=movie_voice::find_descriptor(&window,id.wrapping_sub(1))
+            .or_else(||movie_voice::find_descriptor_before(&window,end_local))
+            .map(|o|win_start+o as u64)
+            .filter(|&s|s<end && end-s<1_500_000)
+            .unwrap_or(anchor);
+        let region=rg(&disc,start,(end-start) as usize);
+        let mut riffs=slb::to_xma_riffs(&region);
+        if riffs.is_empty(){ riffs=slb::to_xma_riff_best(&region).into_iter().collect(); }
+        let mut d="?".into();
+        if let Some(r)=riffs.first(){ let xp=format!("/tmp/ra2_{movie}.xma.wav");let wp=format!("/tmp/ra2_{movie}.wav");fs::write(&xp,r).unwrap();let _=Command::new("ffmpeg").args(["-hide_banner","-v","error","-y","-i",&xp,&wp]).status();d=dur(&wp);}
+        let mv=dur(&format!("{disc}/dat/movie/{movie}.wmv"));
+        println!("{movie:16} id={id:5} region={}KB riffs={} voice={d}s movie={mv}s", (end-start)/1024, riffs.len());
+    }
+}

crates/sylpheed-formats/examples/validate_cues.rs

@@ -0,0 +1,32 @@
+use sylpheed_formats::slb;
+use std::fs;use std::io::{Read,Seek,SeekFrom};use std::process::Command;
+fn rg(disc:&str,g:u64,n:usize)->Vec<u8>{let mut segs=vec![];let mut cum=0u64;for i in 0..5{let p=format!("{disc}/dat/sound.p{i:02}");if let Ok(m)=fs::metadata(&p){segs.push((cum,m.len(),p));cum+=m.len();}}let mut out=vec![];let(mut need,mut pos)=(n,g);for(base,len,path)in &segs{if need==0||pos>=base+len||pos<*base{continue;}let local=pos-base;let take=need.min((len-local)as usize);let mut f=fs::File::open(path).unwrap();f.seek(SeekFrom::Start(local)).unwrap();let mut b=vec![0u8;take];f.read_exact(&mut b).unwrap();out.extend_from_slice(&b);need-=take;pos+=take as u64;}out}
+fn dur(w:&str)->String{let o=Command::new("ffprobe").args(["-v","error","-show_entries","format=duration:stream=channels","-of","default=nw=1:nk=1",w]).output().unwrap();String::from_utf8_lossy(&o.stdout).replace('\n'," ")}
+fn movdur(disc:&str,m:&str)->String{ dur(&format!("{disc}/dat/movie/{m}")) }
+fn main(){
+    let disc=std::env::var("SYLPHEED_DISC").unwrap();
+    // descriptor trailer global offsets (from scan) => cue N data = [desc(N-1)..desc(N)]
+    // (id, end_off, movie)
+    let cues=[(1600u32,437044592u64,"ADV.wmv"),(1601,437345648,"RT01A.wmv"),(1602,437712240,"RT01B.wmv"),
+              (1603,438080880,"RT01C_1.wmv"),(1604,438451568,"RT01C_2.wmv"),(1605,438789488,"RT02A.wmv")];
+    for k in 1..cues.len(){
+        let (id,end,mov)=cues[k];
+        let start=cues[k-1].1;
+        let region=rg(&disc,start,(end-start) as usize + 12000); // +tail to include full data before next trailer
+        let riffs=slb::to_xma_riffs(&region);
+        let mut total=0.0f32; let mut parts=vec![];
+        for (j,r) in riffs.iter().enumerate(){
+            let xp=format!("/tmp/cue{id}_{j}.xma.wav"); let wp=format!("/tmp/cue{id}_{j}.wav");
+            fs::write(&xp,r).unwrap();
+            let _=Command::new("ffmpeg").args(["-hide_banner","-v","error","-y","-i",&xp,&wp]).status();
+            let d=dur(&wp); parts.push(d.clone());
+            total+=d.split_whitespace().next().unwrap_or("0").parse::<f32>().unwrap_or(0.0);
+        }
+        // spanning?
+        let span_adv_end=437547264u64;
+        let spans = start < span_adv_end && end > span_adv_end || (start/1_000 != end/1_000);
+        println!("cue {id} ({mov}): region[{start}..{end}] {} bytes, {} riff(s), Σ={:.1}s  | movie={}  parts={:?}",
+            end-start, riffs.len(), total, movdur(&disc,mov), parts);
+    }
+    println!("\nWAVs at /tmp/cue16XX_*.wav  (listen: RT01B=1602, RT01C_1=1603, RT01C_2=1604)");
+}

crates/sylpheed-formats/src/lib.rs

@@ -50,10 +50,23 @@ pub mod mesh;
 // Audio parsing — scaffold for XMA handling
 pub mod audio;
 
+// Movie cutscene subtitles — the movie → track → text chain.
+pub mod movie_subtitle;
+
+// XACT `.slb` sound banks (voice / music) → decodable XMA1 RIFF.
+pub mod slb;
+
+// The movie manifest: authoritative movie → subtitle → voice index.
+pub mod movie_manifest;
+
+pub mod movie_voice;
+
 /// Re-export the most commonly used types at the crate root.
 pub use font::FontInfo;
 pub use idxd::{IdxdError, IdxdObject};
 pub use ixud::{Cue, Subtitle};
+pub use movie_subtitle::{SubCue, SubLang};
+pub use mesh::{GameMesh, Xbg7Model};
 pub use ratc::RatcChild;
 pub use t8ad::T8adImage;
 pub use pak::{PakArchive, PakEntry, PakError};

crates/sylpheed-formats/src/movie_manifest.rs

@@ -0,0 +1,428 @@
+//! The movie manifest: the game's authoritative **mission → phase → cutscene**
+//! table (statically reversed). Beyond the movie→voice binding, it defines the
+//! whole cutscene *structure*: which video (+ subtitle, + on-screen text overlay,
+//! + voice) plays at each mission phase, and of what kind (story intro, in-mission
+//! phase cutscene, resupply, …).
+//!
+//! ## On-disc structure (`dat/tables.pak`, entry hash `0x5b983a08`)
+//!
+//! One `IDXD` table (`Z1`+zlib block), schema `0x067025b9`. Unlike a flat IDXD
+//! object, this one is an **array of records**, and its string pool is laid out
+//! as **two consecutive arrays**:
+//!   1. the **slot KEYS** — `LOGO1`, `ADVERTISE_MOVIE`, `STAFF_ROLL`, then one
+//!      per cutscene: `MS<NN><X>`, `STAGE<NN>_PHASE0<N>`, `STAGE<NN>_PHASE_END…`,
+//!      `S<NN>_SUPPLY_<ACROPOLIS|TANKER>`, in play order;
+//!   2. the **value RECORDS**, in the same order — each a run led by `<movie>.wmv`
+//!      and (optionally) a `<pak>+…​.prt` overlay (field `TELOP`), a
+//!      `<pak>+SUBTITLE_<movie>.tbl` (field `SUBTITLE`), and a `VOICE_<token>`
+//!      (field `VOICETRACK`). A few slot keys are **valueless** (e.g. stages 5 and
+//!      12 have no resupply video) — those slots have no record and are skipped.
+//!
+//! The **slot key is the semantic role** — this is how the game separates a
+//! cutscene voice from in-mission radio chatter (structurally, by which slot/bank,
+//! not one boolean): `MS*` = story intro (`VOICE_S*`), `STAGE*_PHASE*` = radio-box
+//! cutscene (`VOICE_RT*`), `S*_SUPPLY_*` = resupply (`VOICE_D_45x` or none).
+//! Resupply videos are **reused across stages** (`S04_SUPPLY_ACROPOLIS` →
+//! `hokyu_DS_s02A`), so the mapping is genuinely a table, not derivable from names.
+//!
+//! The voice token is likewise **not** always `VOICE_<movie>`: most movies use
+//! `VOICE_<movie>` (in `<lang>\Movie\`), but the resupply cutscenes bind to
+//! in-mission radio clips like `VOICE_D_450` (in `<lang>\etc\`), and some have no
+//! voice at all. This manifest is the ground truth.
+//!
+//! ## Alignment
+//!
+//! Slot keys and value records would be a 1:1 positional zip but for the valueless
+//! slots. We recover the gaps without decoding the binary node/index region by
+//! using the fully-derivable **`MS<NN><X>` → `S<NN><X>` anchors**: a non-`MS` slot
+//! whose record would be an `MS` target actually belongs to the upcoming `MS`
+//! slot, so the non-`MS` slot is a gap. (Decoding the node region would make this
+//! exact and also unlock the defaulted numeric fields in other IDXD tables.)
+
+use crate::slb::VoiceLang;
+
+/// The role a cutscene slot plays in the mission flow, from its slot key.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum MovieKind {
+    /// `LOGO*`, `ADVERTISE_MOVIE`, `STAFF_ROLL` — boot/system movies.
+    System,
+    /// `MS<NN><X>` — story intro movie (camera on the speaker; stereo `VOICE_S*`).
+    Intro,
+    /// `STAGE<NN>_PHASE0<N>` — in-mission phase cutscene (radio box; `VOICE_RT*`).
+    Phase,
+    /// `STAGE<NN>_PHASE_END[_0N]` — phase-end cutscene.
+    PhaseEnd,
+    /// `S<NN>_SUPPLY_<ACROPOLIS|TANKER>` — resupply (hokyu) cutscene.
+    Supply,
+}
+
+/// One cutscene slot's manifest row: its semantic slot plus the bound assets.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct MovieEntry {
+    /// The slot key, e.g. `STAGE01_PHASE01`, `MS00A`, `S02_SUPPLY_ACROPOLIS`.
+    /// Empty when parsed from a blob without the slot-key array (fallback).
+    pub slot: String,
+    /// The slot's role in the mission flow (intro vs. phase cutscene vs. supply …).
+    pub kind: MovieKind,
+    /// Mission/stage number (`STAGE<NN>` / `MS<NN>` / `S<NN>_SUPPLY`), if any.
+    pub mission: Option<u8>,
+    /// Phase number for [`MovieKind::Phase`]/[`MovieKind::PhaseEnd`], if any.
+    pub phase: Option<u8>,
+    /// Movie basename without extension, e.g. `S13A`, `RT01A`, `hokyu_DS_s13A`.
+    pub movie: String,
+    /// Voice bank basename (`VOICE_S13A`, `VOICE_D_450`), or `None` for no voice.
+    pub voice_token: Option<String>,
+    /// Subtitle table basename (`SUBTITLE_S13A.tbl`), pak prefix stripped, if any.
+    pub subtitle: Option<String>,
+    /// On-screen text-overlay (`TELOP`) basename (`pwrt01.prt`), if any.
+    pub telop: Option<String>,
+}
+
+/// Does this blob look like the movie manifest? (`IDXD` whose string pool
+/// carries movie + subtitle + voice references.) Used to locate it in
+/// `tables.pak` without hardcoding a hash.
+pub fn is_manifest(bytes: &[u8]) -> bool {
+    bytes.len() >= 4
+        && &bytes[0..4] == b"IDXD"
+        && contains(bytes, b".wmv")
+        && contains(bytes, b"VOICE_")
+        && contains(bytes, b"SUBTITLE_")
+}
+
+/// The field-type keys that appear once in the value region as a schema template.
+const FIELD_KEYS: [&str; 4] = ["MOVIE", "VOICETRACK", "SUBTITLE", "TELOP"];
+
+/// Parse the manifest into per-slot cutscene rows, in play order.
+///
+/// Each returned [`MovieEntry`] carries a movie; valueless slots (e.g. stage 5's
+/// absent resupply) are omitted. When the two-array slot/record layout is present
+/// (the real manifest) every row is tagged with its authoritative slot/kind/
+/// mission/phase; otherwise (a blob without the key array) the kind is inferred
+/// from the movie name and `slot` is empty.
+pub fn parse(bytes: &[u8]) -> Vec<MovieEntry> {
+    let toks = ascii_runs(bytes, 3);
+
+    // Locate the slot-key array (`LOGO1` … first `<movie>.wmv`) and the value
+    // region (from the first `.wmv` onward). Fall back to a keyless scrape.
+    let first_wmv = toks.iter().position(|t| t.ends_with(".wmv"));
+    let key_start = toks.iter().position(|t| t == "LOGO1");
+    let (slot_keys, value_toks): (Vec<&str>, &[String]) = match (key_start, first_wmv) {
+        (Some(k), Some(w)) if k < w => (
+            toks[k..w].iter().map(String::as_str).collect(),
+            &toks[w..],
+        ),
+        _ => (Vec::new(), toks.as_slice()),
+    };
+
+    // Build the value records (drop the schema-template field keys).
+    let mut records: Vec<MovieEntry> = Vec::new();
+    for t in value_toks.iter().filter(|t| !FIELD_KEYS.contains(&t.as_str())) {
+        if let Some(name) = t.strip_suffix(".wmv") {
+            records.push(MovieEntry {
+                slot: String::new(),
+                kind: kind_from_movie(name),
+                mission: None,
+                phase: None,
+                movie: name.to_string(),
+                voice_token: None,
+                subtitle: None,
+                telop: None,
+            });
+        } else if let Some(rec) = records.last_mut() {
+            if let Some(v) = t.strip_prefix("VOICE_") {
+                if rec.voice_token.is_none() {
+                    rec.voice_token = Some(format!("VOICE_{v}"));
+                }
+            } else if t.ends_with(".tbl") {
+                rec.subtitle.get_or_insert_with(|| after_plus(t));
+            } else if t.ends_with(".prt") {
+                rec.telop.get_or_insert_with(|| after_plus(t));
+            }
+        }
+    }
+
+    // Attach slot keys to records, recovering valueless-slot gaps via MS anchors.
+    if !slot_keys.is_empty() {
+        let ms_targets: std::collections::HashSet<String> = slot_keys
+            .iter()
+            .filter_map(|k| ms_target(k))
+            .collect();
+        let mut j = 0usize;
+        for slot in &slot_keys {
+            if j >= records.len() {
+                break;
+            }
+            // A gap: this slot's "record" is actually the next MS slot's movie.
+            let would_be = &records[j].movie;
+            let is_gap = match ms_target(slot) {
+                // MS slot: only binds if its record is the expected `S<NN><X>`.
+                Some(expected) => *would_be != expected,
+                // Non-MS slot: a gap iff the record belongs to an upcoming MS slot.
+                None => ms_targets.contains(would_be),
+            };
+            if is_gap {
+                continue;
+            }
+            let (kind, mission, phase) = classify_slot(slot);
+            let rec = &mut records[j];
+            rec.slot = (*slot).to_string();
+            rec.kind = kind;
+            rec.mission = mission;
+            rec.phase = phase;
+            j += 1;
+        }
+    }
+
+    records
+}
+
+/// If `slot` is an `MS<NN><X>` intro slot, the movie it must bind to (`S<NN><X>`).
+fn ms_target(slot: &str) -> Option<String> {
+    let rest = slot.strip_prefix("MS")?;
+    // `<NN><X>`: two digits then a letter suffix.
+    if rest.len() >= 3 && rest[..2].bytes().all(|b| b.is_ascii_digit()) {
+        Some(format!("S{rest}"))
+    } else {
+        None
+    }
+}
+
+/// Semantic role + mission/phase from a slot key.
+fn classify_slot(slot: &str) -> (MovieKind, Option<u8>, Option<u8>) {
+    if let Some(rest) = slot.strip_prefix("MS") {
+        if rest.len() >= 2 {
+            return (MovieKind::Intro, rest[..2].parse().ok(), None);
+        }
+    }
+    if let Some(rest) = slot.strip_prefix("STAGE") {
+        let nn: String = rest.chars().take_while(|c| c.is_ascii_digit()).collect();
+        if let Ok(m) = nn.parse::<u8>() {
+            if slot.contains("PHASE_END") {
+                // trailing `_NN` if present (`STAGE12_PHASE_END_02`)
+                let phase = slot.rsplit('_').next().and_then(|x| x.parse().ok());
+                return (MovieKind::PhaseEnd, Some(m), phase);
+            }
+            // `STAGE<NN>_PHASE0<N>`
+            let phase = rest[nn.len()..]
+                .strip_prefix("_PHASE")
+                .and_then(|p| p.parse().ok());
+            return (MovieKind::Phase, Some(m), phase);
+        }
+    }
+    if slot.contains("_SUPPLY") {
+        if let Some(rest) = slot.strip_prefix('S') {
+            let nn: String = rest.chars().take_while(|c| c.is_ascii_digit()).collect();
+            return (MovieKind::Supply, nn.parse().ok(), None);
+        }
+    }
+    (MovieKind::System, None, None)
+}
+
+/// Fallback kind from a movie basename (used when no slot-key array is present).
+fn kind_from_movie(movie: &str) -> MovieKind {
+    if movie.starts_with("hokyu_") {
+        MovieKind::Supply
+    } else if movie.starts_with("RT") {
+        MovieKind::Phase
+    } else if movie.len() >= 3
+        && movie.starts_with('S')
+        && movie[1..3].bytes().all(|b| b.is_ascii_digit())
+    {
+        MovieKind::Intro
+    } else {
+        MovieKind::System
+    }
+}
+
+/// The part of a `<pak>+<name>` reference after the `+` (or the whole string).
+fn after_plus(s: &str) -> String {
+    s.rsplit('+').next().unwrap_or(s).to_string()
+}
+
+/// The voice bank basename bound to `movie`, if the manifest lists one.
+///
+/// Returns `None` both when the movie is absent and when it is present with no
+/// voice — callers that need to distinguish should use [`parse`].
+pub fn voice_token(bytes: &[u8], movie: &str) -> Option<String> {
+    parse(bytes)
+        .into_iter()
+        .find(|e| e.movie == movie)
+        .and_then(|e| e.voice_token)
+}
+
+/// Resolve a movie to its full `sound.pak` voice-entry name for `lang`, using the
+/// manifest for the *token* and `sounds.tbl` for the token's *directory* (Movie /
+/// etc / Voice differ per token). `None` = the movie has no voice-over.
+pub fn resolve_voice_entry(
+    manifest: &[u8],
+    sounds_tbl: &[u8],
+    movie: &str,
+    lang: VoiceLang,
+) -> Option<String> {
+    let token = voice_token(manifest, movie)?;
+    resolve_token_path(sounds_tbl, &token, lang)
+}
+
+/// Find the full `<lang>\…\<token>.slb` path for a voice `token` in a decompressed
+/// `sounds.tbl` (the token's subdir — `Movie`, `etc`, `Voice` — is not fixed).
+pub fn resolve_token_path(sounds_tbl: &[u8], token: &str, lang: VoiceLang) -> Option<String> {
+    let prefix = format!("{}\\", lang.code_pub());
+    let suffix = format!("\\{token}.slb");
+    ascii_runs(sounds_tbl, 6)
+        .into_iter()
+        .find(|r| r.starts_with(&prefix) && r.ends_with(&suffix))
+}
+
+fn contains(hay: &[u8], needle: &[u8]) -> bool {
+    hay.windows(needle.len()).any(|w| w == needle)
+}
+
+/// Collect printable-ASCII runs of at least `min` chars.
+fn ascii_runs(bytes: &[u8], min: usize) -> Vec<String> {
+    let mut out = Vec::new();
+    let mut cur = String::new();
+    for &b in bytes {
+        if (0x20..=0x7e).contains(&b) {
+            cur.push(b as char);
+        } else {
+            if cur.len() >= min {
+                out.push(std::mem::take(&mut cur));
+            } else {
+                cur.clear();
+            }
+        }
+    }
+    if cur.len() >= min {
+        out.push(cur);
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn synth() -> Vec<u8> {
+        // Minimal IDXD string pool: three movies — a standard one, a hokyu bound
+        // to a VOICE_D radio clip, and a hokyu with no voice.
+        let mut b = b"IDXD".to_vec();
+        b.extend_from_slice(&[0, 0, 0, 0x69, 0, 0]); // binary header (as on disc)
+        for s in [
+            "S13A.wmv",
+            "eng.pak+SUBTITLE_S13A.tbl",
+            "VOICE_S13A",
+            "hokyu_LS_s02A.wmv",
+            "eng.pak+SUBTITLE_hokyu_LS_s02A.tbl",
+            "VOICE_D_450",
+            "hokyu_DS_s13A.wmv",
+            "eng.pak+SUBTITLE_hokyu_DS_s13A.tbl",
+        ] {
+            b.extend_from_slice(s.as_bytes());
+            b.push(0);
+        }
+        b
+    }
+
+    #[test]
+    fn groups_movies_and_binds_voice() {
+        // Keyless fallback path: movie + voice + subtitle + inferred kind.
+        let m = parse(&synth());
+        assert_eq!(m.len(), 3);
+        assert_eq!(m[0].movie, "S13A");
+        assert_eq!(m[0].voice_token.as_deref(), Some("VOICE_S13A"));
+        assert_eq!(m[0].kind, MovieKind::Intro);
+        assert_eq!(m[0].subtitle.as_deref(), Some("SUBTITLE_S13A.tbl"));
+        assert_eq!(m[1].voice_token.as_deref(), Some("VOICE_D_450"));
+        assert_eq!(m[1].kind, MovieKind::Supply);
+        assert_eq!(m[2].voice_token, None, "hokyu_DS_s13A has no voice-over");
+        assert_eq!(m[2].kind, MovieKind::Supply);
+    }
+
+    /// Build a two-array (keys-then-records) manifest like the real one.
+    fn synth_two_array(keys: &[&str], records: &[&[&str]]) -> Vec<u8> {
+        let mut b = b"IDXD".to_vec();
+        b.extend_from_slice(&[0, 0, 0, 0x05, 0, 0]);
+        for k in keys {
+            b.extend_from_slice(k.as_bytes());
+            b.push(0);
+        }
+        for rec in records {
+            for s in *rec {
+                b.extend_from_slice(s.as_bytes());
+                b.push(0);
+            }
+        }
+        b
+    }
+
+    #[test]
+    fn two_array_slots_kind_mission_phase() {
+        let blob = synth_two_array(
+            &["LOGO1", "MS01A", "STAGE01_PHASE01", "STAGE01_PHASE_END_02"],
+            &[
+                &["logo1.wmv", "MOVIE"],
+                &["S01A.wmv", "eng.pak+SUBTITLE_S01A.tbl", "VOICE_S01A"],
+                &["RT01A.wmv", "eng.pak+pwrt01.prt", "VOICE_RT01A", "VOICETRACK"],
+                &["RT01C_2.wmv", "VOICE_RT01C_2"],
+            ],
+        );
+        let m = parse(&blob);
+        assert_eq!(m.len(), 4);
+        assert_eq!((m[0].slot.as_str(), m[0].kind), ("LOGO1", MovieKind::System));
+        assert_eq!(
+            (m[1].slot.as_str(), m[1].kind, m[1].mission),
+            ("MS01A", MovieKind::Intro, Some(1))
+        );
+        assert_eq!(
+            (m[2].slot.as_str(), m[2].kind, m[2].mission, m[2].phase),
+            ("STAGE01_PHASE01", MovieKind::Phase, Some(1), Some(1))
+        );
+        assert_eq!(m[2].telop.as_deref(), Some("pwrt01.prt"));
+        assert_eq!(
+            (m[3].slot.as_str(), m[3].kind, m[3].phase),
+            ("STAGE01_PHASE_END_02", MovieKind::PhaseEnd, Some(2))
+        );
+    }
+
+    #[test]
+    fn valueless_supply_slot_is_skipped_via_ms_anchor() {
+        // S05_SUPPLY has no record; its "record" (S06A) belongs to the next MS slot.
+        let blob = synth_two_array(
+            &["LOGO1", "MS05A", "S05_SUPPLY_ACROPOLIS", "MS06A"],
+            &[
+                &["logo1.wmv"],
+                &["S05A.wmv", "VOICE_S05A"],
+                &["S06A.wmv", "VOICE_S06A"],
+            ],
+        );
+        let m = parse(&blob);
+        assert_eq!(m.len(), 3, "the valueless supply slot yields no entry");
+        assert_eq!((m[1].slot.as_str(), m[1].movie.as_str()), ("MS05A", "S05A"));
+        // S06A binds to MS06A, NOT to the intervening valueless supply slot.
+        assert_eq!((m[2].slot.as_str(), m[2].movie.as_str()), ("MS06A", "S06A"));
+        assert!(m.iter().all(|e| e.slot != "S05_SUPPLY_ACROPOLIS"));
+    }
+
+    #[test]
+    fn resolves_token_directory_from_sounds_tbl() {
+        // sounds.tbl places the two tokens in different subdirs.
+        let mut tbl = Vec::new();
+        for s in ["eng\\Movie\\VOICE_S13A.slb", "eng\\etc\\VOICE_D_450.slb"] {
+            tbl.extend_from_slice(s.as_bytes());
+            tbl.push(0);
+        }
+        let man = synth();
+        assert_eq!(
+            resolve_voice_entry(&man, &tbl, "S13A", VoiceLang::English).as_deref(),
+            Some("eng\\Movie\\VOICE_S13A.slb")
+        );
+        assert_eq!(
+            resolve_voice_entry(&man, &tbl, "hokyu_LS_s02A", VoiceLang::English).as_deref(),
+            Some("eng\\etc\\VOICE_D_450.slb")
+        );
+        assert_eq!(
+            resolve_voice_entry(&man, &tbl, "hokyu_DS_s13A", VoiceLang::English),
+            None
+        );
+    }
+}

crates/sylpheed-formats/src/movie_subtitle.rs

@@ -0,0 +1,454 @@
+//! Movie cutscene subtitles: the movie → track → text chain.
+//!
+//! Reverse-engineered statically (see `docs/re/structures/movie-subtitles.md`).
+//! A cutscene's on-screen captions are assembled from three places on the disc:
+//!
+//! 1. `dat/movie/<lang>.pak` holds one **timing track** per movie, keyed by
+//!    [`track_key`] = `name_hash("subtitle_<movie>.tbl")`. Each track is a
+//!    `Z1`+zlib block that decompresses to an **IXUD** container whose UTF-16**LE**
+//!    payload is `SUBTITLE  MSG_DEMO_<demo>  <mm:ss.cc>  …` — i.e. *which*
+//!    demo-message shows *when* (radio / resupply movies), or inline text.
+//! 2. `dat/GP_MAIN_GAME_<L>.pak` holds the **caption text**: more `Z1`+zlib IXUD
+//!    blocks where each line is stored as `text` immediately followed by its key
+//!    `MSG_DEMO_<demo>_<page>_<line>` (multi-line captions split across lines).
+//!
+//! [`load`] joins the two into timed [`SubCue`]s.
+//!
+//! Note: the IXUD payload here is UTF-16 **little-endian** (verified on the real
+//! disc); this is deliberately separate from the older big-endian [`crate::ixud`]
+//! presenter, which targets a different (uncompressed) variant.
+
+use std::collections::BTreeMap;
+
+use crate::hash::name_hash;
+use crate::pak::PakArchive;
+
+/// Subtitle language. `pak_code` selects `dat/movie/<code>.pak`; `game_code`
+/// selects `dat/GP_MAIN_GAME_<code>.pak` (the caption text).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum SubLang {
+    English,
+    Japanese,
+    German,
+    French,
+    Spanish,
+    Italian,
+}
+
+impl SubLang {
+    /// All languages, in menu order.
+    pub const ALL: [SubLang; 6] = [
+        SubLang::English,
+        SubLang::Japanese,
+        SubLang::German,
+        SubLang::French,
+        SubLang::Spanish,
+        SubLang::Italian,
+    ];
+
+    /// `dat/movie/<code>.pak` stem (the timing tracks + caption font).
+    pub fn pak_code(self) -> &'static str {
+        match self {
+            SubLang::English => "eng",
+            SubLang::Japanese => "jpn",
+            SubLang::German => "deu",
+            SubLang::French => "fra",
+            SubLang::Spanish => "esp",
+            SubLang::Italian => "ita",
+        }
+    }
+
+    /// `dat/GP_MAIN_GAME_<code>.pak` suffix (the caption text pack).
+    pub fn game_code(self) -> &'static str {
+        match self {
+            SubLang::English => "E",
+            SubLang::Japanese => "J",
+            SubLang::German => "D",
+            SubLang::French => "F",
+            SubLang::Spanish => "S",
+            SubLang::Italian => "I",
+        }
+    }
+
+    /// Human label for a UI selector. Kept to Latin script so it renders in a
+    /// default (non-CJK) UI font; Japanese is romanized for the same reason.
+    pub fn label(self) -> &'static str {
+        match self {
+            SubLang::English => "English",
+            SubLang::Japanese => "Japanese",
+            SubLang::German => "Deutsch",
+            SubLang::French => "Français",
+            SubLang::Spanish => "Español",
+            SubLang::Italian => "Italiano",
+        }
+    }
+}
+
+/// One timed caption line.
+#[derive(Debug, Clone, PartialEq)]
+pub struct SubCue {
+    pub start: f32,
+    pub end: Option<f32>,
+    pub text: String,
+}
+
+/// The `<lang>.pak` TOC key for a movie's subtitle timing track.
+pub fn track_key(movie_basename: &str) -> u32 {
+    name_hash(&format!("subtitle_{movie_basename}.tbl"))
+}
+
+/// Load and resolve a movie's timed captions in `lang`.
+///
+/// `lang_pak` is `dat/movie/<lang>.pak` (+segments); `text_pak` is
+/// `dat/GP_MAIN_GAME_<L>.pak` (+segments). Returns the cues in start order, or an
+/// empty vec if the movie has no timed track (e.g. story movies whose text is a
+/// pre-rendered title card).
+pub fn load(movie_basename: &str, lang_pak: &PakArchive, text_pak: &PakArchive) -> Vec<SubCue> {
+    let Some(Ok(track)) = lang_pak.read_by_hash(track_key(movie_basename)) else {
+        return Vec::new();
+    };
+    if !is_ixud(&track) {
+        return Vec::new();
+    }
+    let raw = parse_track(&track);
+    if raw.is_empty() {
+        return Vec::new();
+    }
+    // Only build the (large) text table if the track actually references demos.
+    let needs_text = raw.iter().any(|(t, _, _)| demo_ref(t).is_some());
+    let text = if needs_text {
+        build_demo_text(text_pak)
+    } else {
+        BTreeMap::new()
+    };
+
+    let mut cues = Vec::new();
+    for (token, start, end) in raw {
+        let resolved = match demo_ref(&token) {
+            Some(demo) => match text.get(&demo) {
+                Some(lines) => lines.join("\n"),
+                None => continue, // demo id with no text on disc → skip
+            },
+            None => clean(&token), // already inline text
+        };
+        if resolved.trim().is_empty() {
+            continue;
+        }
+        cues.push(SubCue {
+            start,
+            end,
+            text: resolved,
+        });
+    }
+    cues.sort_by(|a, b| a.start.partial_cmp(&b.start).unwrap_or(std::cmp::Ordering::Equal));
+    cues
+}
+
+/// The `MSG_DEMO_<id>` ids a movie's timing track references, in track order
+/// (empty for inline-text or absent tracks). Language-independent — the same
+/// demo ids appear in every `<lang>.pak`. Used to share a voice clip between
+/// movies that play the same demo line (the resupply cutscenes reuse one clip
+/// per line while only the video differs).
+pub fn track_demo_ids(lang_pak: &PakArchive, movie_basename: &str) -> Vec<u32> {
+    let Some(Ok(track)) = lang_pak.read_by_hash(track_key(movie_basename)) else {
+        return Vec::new();
+    };
+    if !is_ixud(&track) {
+        return Vec::new();
+    }
+    parse_track(&track)
+        .iter()
+        .filter_map(|(t, _, _)| demo_ref(t))
+        .collect()
+}
+
+/// Per-line `(demo id, start seconds)` for a movie's timing track — the radio-box
+/// voice schedule. Each `MSG_DEMO_<id>` reference takes the start time of the
+/// timing that follows it (a caption may list two demo lines under one timing;
+/// both then share that start). Empty for inline-text / absent tracks (those are
+/// story movies whose voice is the single manifest `VOICETRACK`, not per-line).
+pub fn track_voice_cues(lang_pak: &PakArchive, movie_basename: &str) -> Vec<(u32, f32)> {
+    let Some(Ok(track)) = lang_pak.read_by_hash(track_key(movie_basename)) else {
+        return Vec::new();
+    };
+    if !is_ixud(&track) {
+        return Vec::new();
+    }
+    let toks = utf16le_tokens(&track);
+    let start = toks
+        .iter()
+        .position(|t| t.ends_with("SUBTITLE"))
+        .map(|i| i + 1)
+        .unwrap_or(0);
+    let mut out = Vec::new();
+    let mut pending: Vec<u32> = Vec::new();
+    for tok in &toks[start..] {
+        match parse_timing(tok) {
+            Some((s, _)) => {
+                for d in pending.drain(..) {
+                    out.push((d, s));
+                }
+            }
+            None => {
+                if let Some(d) = demo_ref(tok) {
+                    pending.push(d);
+                }
+            }
+        }
+    }
+    out
+}
+
+/// Build `demo id → ordered caption lines` from a `GP_MAIN_GAME_<L>` pack.
+///
+/// Scans every IXUD block, pairing each text token with the immediately
+/// following `MSG_DEMO_<demo>_<page>_<line>` key, and groups the lines per demo
+/// ordered by (page, line).
+pub fn build_demo_text(text_pak: &PakArchive) -> BTreeMap<u32, Vec<String>> {
+    // demo -> (page,line) -> text
+    let mut by_demo: BTreeMap<u32, BTreeMap<(u32, u32), String>> = BTreeMap::new();
+    for entry in text_pak.entries() {
+        let Ok(bytes) = text_pak.read(entry) else {
+            continue;
+        };
+        if !is_ixud(&bytes) {
+            continue;
+        }
+        let toks = utf16le_tokens(&bytes);
+        for w in toks.windows(2) {
+            if let Some((demo, page, line)) = text_key(&w[1]) {
+                // Pair only when the preceding token is real text — not another
+                // MSG_DEMO key (bare keys are also serialized consecutively in the
+                // record directory, which would otherwise masquerade as text).
+                if demo_ref(&w[0]).is_none()
+                    && text_key(&w[0]).is_none()
+                    && !w[0].trim().is_empty()
+                {
+                    by_demo
+                        .entry(demo)
+                        .or_default()
+                        .insert((page, line), clean(&w[0]));
+                }
+            }
+        }
+    }
+    by_demo
+        .into_iter()
+        .map(|(d, m)| (d, m.into_values().collect()))
+        .collect()
+}
+
+/// Parse a timing track's IXUD payload into `(token, start, end)` triples, where
+/// `token` is either a `MSG_DEMO_<d>` reference or an inline caption string.
+///
+/// A single on-screen caption may be stored as **several consecutive text
+/// tokens** followed by one timing (the source hard-splits a multi-line caption
+/// into one token per line — e.g. `"Look at it father"` + `"& beautiful isn't
+/// it"` share the `01:14.80-01:17.60` timing). We therefore accumulate every
+/// text token seen since the last timing and join them with a newline when the
+/// timing arrives, instead of pairing strictly 1:1 (which silently dropped every
+/// line but the last of a multi-line caption).
+fn parse_track(ixud: &[u8]) -> Vec<(String, f32, Option<f32>)> {
+    let toks = utf16le_tokens(ixud);
+    let start = toks
+        .iter()
+        .position(|t| t.ends_with("SUBTITLE"))
+        .map(|i| i + 1)
+        .unwrap_or(0);
+    let rest = &toks[start..];
+    let mut out = Vec::new();
+    let mut pending: Vec<&str> = Vec::new();
+    for tok in rest {
+        match parse_timing(tok) {
+            Some((s, e)) => {
+                if !pending.is_empty() {
+                    out.push((pending.join("\n"), s, e));
+                    pending.clear();
+                }
+            }
+            None => pending.push(tok),
+        }
+    }
+    out
+}
+
+fn is_ixud(b: &[u8]) -> bool {
+    b.len() >= 4 && &b[0..4] == b"IXUD"
+}
+
+/// Normalize a caption string: the source stores line breaks as the literal
+/// two-character escape `\n`; turn it into a real newline and trim.
+fn clean(s: &str) -> String {
+    s.replace("\\n", "\n").trim().to_string()
+}
+
+/// `MSG_DEMO_<d>` → `Some(d)` (reference token, no page/line).
+fn demo_ref(t: &str) -> Option<u32> {
+    let rest = t.strip_prefix("MSG_DEMO_")?;
+    if rest.contains('_') {
+        return None; // that's a text key (has page/line), not a bare reference
+    }
+    rest.parse().ok()
+}
+
+/// `MSG_DEMO_<d>_<page>_<line>` → `Some((d,page,line))` (text key).
+fn text_key(t: &str) -> Option<(u32, u32, u32)> {
+    let rest = t.strip_prefix("MSG_DEMO_")?;
+    let mut it = rest.split('_');
+    let d = it.next()?.parse().ok()?;
+    let p = it.next()?.parse().ok()?;
+    let l = it.next()?.parse().ok()?;
+    if it.next().is_some() {
+        return None;
+    }
+    Some((d, p, l))
+}
+
+/// Extract UTF-16**LE** text runs from a blob, **independent of byte alignment**.
+///
+/// The IXUD string pool packs records at odd byte offsets, so a fixed 2-byte
+/// stride from the block start misreads every character. Instead we scan a
+/// sliding window, decoding each 16-bit LE unit and keeping runs of "text-like"
+/// code points (ASCII, Latin-1 supplement — the German umlauts / accented
+/// Latin — and Latin Extended-A). A non-text unit ends the run and we advance by
+/// one byte to re-lock onto the correct parity of the next run. Accepting the
+/// Latin ranges (not just ASCII) is what keeps `müsste`, `Français`, `español`
+/// intact — earlier the accented char *and its predecessor* were dropped.
+///
+/// CJK (Japanese) is deliberately out of range: those units also occur as noise
+/// at the wrong parity, and the UI font can't render them anyway.
+fn utf16le_tokens(bytes: &[u8]) -> Vec<String> {
+    let mut tokens = Vec::new();
+    let mut cur = String::new();
+    let mut i = 0;
+    while i + 1 < bytes.len() {
+        let u = u16::from_le_bytes([bytes[i], bytes[i + 1]]);
+        if is_text_unit(u) {
+            if let Some(c) = char::from_u32(u as u32) {
+                cur.push(c);
+            }
+            i += 2;
+        } else {
+            if cur.chars().count() >= 2 {
+                tokens.push(std::mem::take(&mut cur));
+            } else {
+                cur.clear();
+            }
+            i += 1;
+        }
+    }
+    if cur.chars().count() >= 2 {
+        tokens.push(cur);
+    }
+    tokens
+}
+
+/// Whether a UTF-16 unit is a printable Latin text character we keep in a run.
+/// Excludes the C0/C1 control blocks (incl. 0x7F–0x9F) so binary noise breaks
+/// runs instead of joining them.
+fn is_text_unit(u: u16) -> bool {
+    matches!(u, 0x20..=0x7e | 0xa0..=0xff | 0x100..=0x17f)
+}
+
+/// Parse `MM:SS.ss` or a `start-end` range into seconds.
+fn parse_timing(s: &str) -> Option<(f32, Option<f32>)> {
+    let one = |p: &str| -> Option<f32> {
+        let (mm, ss) = p.trim().split_once(':')?;
+        let m: f32 = if mm.trim().is_empty() {
+            0.0
+        } else {
+            mm.trim().parse().ok()?
+        };
+        let sec: f32 = ss.trim().parse().ok()?;
+        Some(m * 60.0 + sec)
+    };
+    match s.split_once('-') {
+        Some((a, b)) => Some((one(a)?, Some(one(b)?))),
+        None => Some((one(s)?, None)),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn track_key_matches_disc() {
+        // Verified against dat/movie/eng.pak.
+        assert_eq!(track_key("S00A"), 0x6F2D_9663);
+        assert_eq!(track_key("hokyu_DS_s02A"), 0x3662_B1F8);
+        assert_eq!(track_key("RT01C_1"), 0x756F_69FB);
+    }
+
+    #[test]
+    fn demo_and_text_keys() {
+        assert_eq!(demo_ref("MSG_DEMO_192"), Some(192));
+        assert_eq!(demo_ref("MSG_DEMO_604_000_00"), None);
+        assert_eq!(text_key("MSG_DEMO_604_000_01"), Some((604, 0, 1)));
+        assert_eq!(text_key("MSG_DEMO_192"), None);
+    }
+
+    #[test]
+    fn keeps_latin1_accents() {
+        // "müsste" must survive intact (umlaut + its neighbour), not collapse to
+        // "sste". Encode as UTF-16LE at an ODD start offset to exercise the
+        // alignment-independent scan.
+        let mut b = vec![0xAAu8]; // 1 junk byte → strings start at an odd offset
+        for s in ["müsste", "Français", "español"] {
+            for u in s.encode_utf16() {
+                b.extend_from_slice(&u.to_le_bytes());
+            }
+            b.extend_from_slice(&[0, 0]);
+        }
+        let toks = utf16le_tokens(&b);
+        assert!(toks.contains(&"müsste".to_string()), "got {toks:?}");
+        assert!(toks.contains(&"Français".to_string()), "got {toks:?}");
+        assert!(toks.contains(&"español".to_string()), "got {toks:?}");
+    }
+
+    /// Build a synthetic UTF-16LE IXUD and check the LE token walk + timing pair.
+    #[test]
+    fn parses_le_track() {
+        let mut b = b"IXUD".to_vec();
+        b.extend_from_slice(&[0, 1, 0, 0, 0x70, 0x3E, 0xC8, 0x6C]);
+        for t in ["SUBTITLE", "MSG_DEMO_5", "00:01.50", "Hello", "00:03.00"] {
+            for u in t.encode_utf16() {
+                b.extend_from_slice(&u.to_le_bytes());
+            }
+            b.extend_from_slice(&[0, 0]);
+        }
+        let raw = parse_track(&b);
+        assert_eq!(raw.len(), 2);
+        assert_eq!(raw[0], ("MSG_DEMO_5".to_string(), 1.5, None));
+        assert_eq!(raw[1], ("Hello".to_string(), 3.0, None));
+    }
+
+    /// Two text tokens before a single timing = one two-line caption; the first
+    /// line must NOT be dropped (the real S13A `Look at it father` regression).
+    #[test]
+    fn joins_multiline_caption() {
+        let mut b = b"IXUD".to_vec();
+        b.extend_from_slice(&[0, 1, 0, 0, 0x70, 0x3E, 0xC8, 0x6C]);
+        for t in [
+            "SUBTITLE",
+            "That is such magnificent power.",
+            "01:09.70-01:12.20",
+            "Look at it father",
+            "& beautiful isn't it",
+            "01:14.80-01:17.60",
+        ] {
+            for u in t.encode_utf16() {
+                b.extend_from_slice(&u.to_le_bytes());
+            }
+            b.extend_from_slice(&[0, 0]);
+        }
+        let raw = parse_track(&b);
+        assert_eq!(raw.len(), 2);
+        assert_eq!(raw[0].0, "That is such magnificent power.");
+        assert_eq!(
+            raw[1].0, "Look at it father\n& beautiful isn't it",
+            "both lines of the caption must be kept"
+        );
+        assert_eq!((raw[1].1, raw[1].2), (74.8, Some(77.6)));
+    }
+}

crates/sylpheed-formats/src/movie_voice.rs

@@ -0,0 +1,142 @@
+//! Cutscene-voice resolution: movie → sound-id → continuous-stream byte region.
+//!
+//! Reverse-engineered 2026-07-22 (Canary file-I/O trace, user-confirmed by ear).
+//! The movie voices form **one continuous XMA stream**, physically chunked into
+//! `<lang>\Movie\VOICE_*.slb` `sound.pak` TOC entries whose boundaries do **not**
+//! align with the cutscene cues — a single cue routinely spans two `.slb` chunks,
+//! so a `.slb` does not necessarily hold the track its name claims. Each cue's
+//! audio ends at an inline **trailer descriptor** `(sound_id: u32be, 0x11, …)`
+//! whose id repeats at +0x800. Cue N's audio = the stream bytes
+//! `[descriptor(N-1) .. descriptor(N)]`, read continuously across chunk
+//! boundaries. Verified: decoded voice length matches each movie within ~0.3 s,
+//! including cross-boundary cues (RT01B/RT01C span `VOICE_ADV`→`VOICE_RT01A`→…).
+//!
+//! Resolution chain:
+//!   1. movie → cue name (`ADVERTISE_MOVIE` manifest `VOICETRACK`, e.g. `VOICE_RT01A`)
+//!   2. cue name → sound-id (master sound registry, [`registry_voice_ids`], → 1601)
+//!   3. sound-id → `[start, end]` global offsets via [`find_descriptor`] over the
+//!      `sound.pNN` stream (the caller supplies the bytes + the physical anchor).
+
+use std::collections::HashMap;
+
+/// Parse the master sound registry — the large per-language `tables.pak` IDXD
+/// entry (schema `0x13cb84ba`) that carries the `<lang>\Movie\VOICE_*.slb` paths —
+/// into `cue-name → sound-id` from its `NAME\0 ID\0` string pool
+/// (e.g. `VOICE_ADV`→1600, `VOICE_RT01A`→1601).
+pub fn registry_voice_ids(registry: &[u8]) -> HashMap<String, u32> {
+    let mut toks: Vec<&[u8]> = Vec::new();
+    let mut start = 0usize;
+    for i in 0..registry.len() {
+        if registry[i] == 0 {
+            if i > start {
+                toks.push(&registry[start..i]);
+            }
+            start = i + 1;
+        }
+    }
+    let mut out = HashMap::new();
+    for w in toks.windows(2) {
+        let (name, num) = (w[0], w[1]);
+        if name.starts_with(b"VOICE_")
+            && !num.is_empty()
+            && num.iter().all(u8::is_ascii_digit)
+        {
+            if let (Ok(n), Ok(id)) = (
+                std::str::from_utf8(name),
+                std::str::from_utf8(num).unwrap().parse::<u32>(),
+            ) {
+                out.entry(n.to_string()).or_insert(id);
+            }
+        }
+    }
+    out
+}
+
+/// The constant marker following a cue's sound-id in its inline trailer.
+const DESC_MARK: u32 = 0x11;
+/// The trailer repeats its id this many bytes later; used to reject a coincidental
+/// `(id, 0x11)` pair inside XMA audio (probability of a false match is ~2⁻⁶⁴).
+const DESC_REPEAT: usize = 0x800;
+
+/// Byte offset (within `buf`, a slice of the continuous voice stream) of cue
+/// `id`'s trailer descriptor: the first big-endian `(id, 0x11)` pair whose id
+/// repeats at `+0x800`. `None` if absent in the slice.
+pub fn find_descriptor(buf: &[u8], id: u32) -> Option<usize> {
+    if buf.len() < DESC_REPEAT + 4 {
+        return None;
+    }
+    let be = |o: usize| u32::from_be_bytes([buf[o], buf[o + 1], buf[o + 2], buf[o + 3]]);
+    let end = buf.len() - (DESC_REPEAT + 4);
+    let mut o = 0;
+    while o <= end {
+        if be(o) == id && be(o + 4) == DESC_MARK && be(o + DESC_REPEAT) == id {
+            return Some(o);
+        }
+        o += 4;
+    }
+    None
+}
+
+/// Plausible sound-id range for a trailer (all real cue ids are well under this).
+const ID_MAX: u32 = 0x1_0000;
+
+/// Offset of the nearest cue trailer strictly **before** `before` — any plausible
+/// id. Used to bound a cue's START when its `id-1` predecessor doesn't exist
+/// (the sound-id sequence has gaps, e.g. `VOICE_D_453`=7142 → `VOICE_D_454`=7188).
+/// Scans downward and returns the first (largest-offset) valid `(id, 0x11)` with
+/// the `+0x800` id-repeat. `None` if there is no trailer below `before`.
+pub fn find_descriptor_before(buf: &[u8], before: usize) -> Option<usize> {
+    if before < 4 {
+        return None;
+    }
+    let cap = buf.len().checked_sub(DESC_REPEAT + 4)?;
+    let be = |o: usize| u32::from_be_bytes([buf[o], buf[o + 1], buf[o + 2], buf[o + 3]]);
+    // Start strictly below `before` (4-aligned) so the target's own trailer is
+    // never returned as its predecessor.
+    let mut o = (before - 1).min(cap) & !3;
+    loop {
+        let id = be(o);
+        if id >= 1 && id < ID_MAX && be(o + 4) == DESC_MARK && be(o + DESC_REPEAT) == id {
+            return Some(o);
+        }
+        if o < 4 {
+            return None;
+        }
+        o -= 4;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn registry_parses_name_id_pairs() {
+        let mut b = Vec::new();
+        for tok in [
+            b"VOICE_ADV".as_slice(),
+            b"1600",
+            b"VOICE_RT01A",
+            b"1601",
+            b"NOT_A_VOICE",
+            b"9",
+        ] {
+            b.extend_from_slice(tok);
+            b.push(0);
+        }
+        let ids = registry_voice_ids(&b);
+        assert_eq!(ids.get("VOICE_ADV"), Some(&1600));
+        assert_eq!(ids.get("VOICE_RT01A"), Some(&1601));
+        assert_eq!(ids.get("NOT_A_VOICE"), None);
+    }
+
+    #[test]
+    fn find_descriptor_matches_repeat() {
+        let mut b = vec![0u8; 0x900];
+        b[0..4].copy_from_slice(&1601u32.to_be_bytes());
+        b[4..8].copy_from_slice(&0x11u32.to_be_bytes());
+        b[DESC_REPEAT..DESC_REPEAT + 4].copy_from_slice(&1601u32.to_be_bytes());
+        assert_eq!(find_descriptor(&b, 1601), Some(0));
+        assert_eq!(find_descriptor(&b, 1600), None);
+    }
+}

crates/sylpheed-formats/src/pak.rs

@@ -178,6 +178,18 @@ impl PakArchive {
         })
     }
 
+    /// Parse **only** the TOC from a `.pak` index, without any segment data.
+    ///
+    /// For huge archives (`sound.pak` is ~1 GB across `.p00..p04`) this lets a
+    /// caller locate one entry's `(offset, size)` and read just that byte range
+    /// from the segments, instead of loading the whole thing into memory. The
+    /// returned entries are sorted by `name_hash` (binary-searchable).
+    pub fn parse_toc(index: &[u8]) -> Result<Vec<PakEntry>, PakError> {
+        // Reuse from_parts' validation with an empty data blob; the entries are
+        // independent of the data.
+        Ok(Self::from_parts(index, Vec::new())?.entries)
+    }
+
     /// All TOC entries, in stored order (ascending `name_hash`).
     pub fn entries(&self) -> &[PakEntry] {
         &self.entries

crates/sylpheed-formats/src/slb.rs

@@ -0,0 +1,370 @@
+//! `.slb` XACT sound banks → a decodable XMA1 `RIFF`.
+//!
+//! `dat/sound.pak` (9519 entries across `sound.p00..p04`) is the game's audio
+//! bank. Each entry is an XACT `.slb` wrapping **XMA1** (`fmt ` tag `0x0165`,
+//! 48 kHz). Files are named `<lang>\Voice\…`, `<lang>\Movie\VOICE_<movie>.slb`,
+//! `BGM_###.slb`, etc. (see `docs/re/structures/sound-slb.md`); look them up by
+//! [`crate::hash::name_hash`]. Movie cutscene voice = one continuous
+//! `<eng|jpn>\Movie\VOICE_<movie>.slb` track meant to play from the video start.
+//!
+//! Two on-disc layouts, both reversed statically:
+//! - **RIFF present** — a standard `RIFF/WAVE` sits inside the bank; its 32-byte
+//!   `fmt ` is the real `XMAWAVEFORMAT` and the XMA packets are everything after
+//!   that RIFF's `data` chunk header (the declared `data` size is unreliable, so
+//!   we take to end and let the caller clamp to the known media length).
+//! - **Headerless** — no RIFF at all; a fixed **1392-byte** header precedes raw
+//!   XMA1 packets (48 kHz, 2 channels).
+//!
+//! [`to_xma_riff`] rebuilds a standalone `RIFF/WAVE` (XMA1) for either layout,
+//! ready to hand to an XMA decoder (e.g. FFmpeg's `xma1`). The content is always
+//! mono (some clips put it in the left channel only, others duplicate L=R), so
+//! the decode step should downmix to mono (take the left channel).
+
+/// Fixed offset of the raw XMA1 stream in a headerless `.slb` (no `RIFF`).
+pub const HEADERLESS_DATA_OFFSET: usize = 1392;
+
+/// Voice language for cutscene audio. Only English and Japanese voice exist on
+/// the disc (subtitles cover more languages, voice does not).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum VoiceLang {
+    English,
+    Japanese,
+}
+
+impl VoiceLang {
+    fn code(self) -> &'static str {
+        match self {
+            VoiceLang::English => "eng",
+            VoiceLang::Japanese => "jpn",
+        }
+    }
+
+    /// `eng` / `jpn` — the `sound.pak` path prefix for this voice language.
+    pub fn code_pub(self) -> &'static str {
+        self.code()
+    }
+}
+
+/// The `sound.pak` entry name for a movie's continuous voice track, e.g.
+/// `eng\Movie\VOICE_RT07A.slb`. Hash it with [`crate::hash::name_hash`] to get
+/// the `sound.pak` TOC key.
+pub fn movie_voice_name(movie_basename: &str, lang: VoiceLang) -> String {
+    format!("{}\\Movie\\VOICE_{}.slb", lang.code(), movie_basename)
+}
+
+/// One playable voice clip discovered in `sounds.tbl`.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct VoiceClip {
+    /// Full `sound.pak` entry name, e.g. `eng\Voice\VOICE_ADAN_010.slb`.
+    pub name: String,
+    /// Speaker/category code, e.g. `ADAN`, `TCAF`, `A`, `RT07A`.
+    pub speaker: String,
+    /// Short UI label, e.g. `ADAN 010`.
+    pub display: String,
+}
+
+/// Enumerate the voice/dialog clips named in a decompressed `sounds.tbl` (the
+/// IDXD in `tables.pak`). Extracts every `<lang>\{Voice,etc,Movie,Briefing}\…`
+/// path ending in `.slb` for `lang`, parsed into `(name, speaker, display)`.
+pub fn list_voice_clips(sounds_tbl: &[u8], lang: VoiceLang) -> Vec<VoiceClip> {
+    let prefix = format!("{}\\", lang.code());
+    let mut seen = std::collections::BTreeSet::new();
+    let mut out = Vec::new();
+    // Scan for printable-ASCII runs; keep those that look like a voice path.
+    let mut i = 0;
+    while i < sounds_tbl.len() {
+        let start = i;
+        while i < sounds_tbl.len() && (0x20..=0x7e).contains(&sounds_tbl[i]) {
+            i += 1;
+        }
+        if i - start >= 6 {
+            if let Ok(s) = std::str::from_utf8(&sounds_tbl[start..i]) {
+                // Every spoken-line category, so the standalone player covers them
+                // all: in-mission radio (`\Voice\`, `\etc\`) and bound movie voices
+                // (`\Movie\`) all carry `VOICE_`; mission-briefing lines live in
+                // `\Briefing\` as `BR<NN>_<MM>.slb` (no `VOICE` in the name).
+                let is_voice = s.contains("VOICE") || s.contains("\\Briefing\\");
+                if s.starts_with(&prefix) && s.ends_with(".slb") && is_voice {
+                    if seen.insert(s.to_string()) {
+                        out.push(parse_voice_clip(s));
+                    }
+                }
+            }
+        }
+        i += 1;
+    }
+    out
+}
+
+fn parse_voice_clip(name: &str) -> VoiceClip {
+    // `<lang>\<cat>\VOICE_<SPK>_<NNN>.slb` or `..\VOICE_<movie>.slb`.
+    let stem = name
+        .rsplit('\\')
+        .next()
+        .unwrap_or(name)
+        .strip_suffix(".slb")
+        .unwrap_or(name);
+    let body = stem.strip_prefix("VOICE_").unwrap_or(stem);
+    let (speaker, display) = match body.rsplit_once('_') {
+        Some((spk, num)) if num.chars().all(|c| c.is_ascii_digit()) => {
+            (spk.to_string(), format!("{spk} {num}"))
+        }
+        _ => (body.to_string(), body.to_string()),
+    };
+    VoiceClip {
+        name: name.to_string(),
+        speaker,
+        display,
+    }
+}
+
+/// Build one standalone XMA1 `RIFF/WAVE` **per sub-wave** in a `.slb` bank, in
+/// on-disc order. A `.slb` is an XACT bank of one or more sub-waves; the sub-waves
+/// are EITHER alternate takes (the first is the whole track — e.g. `VOICE_S01A`,
+/// `VOICE_ADV`: sub0 ≈ the movie length) OR sequential segments that must be
+/// concatenated (e.g. `VOICE_RT07A`: 24s + 14s + 11s ≈ the 50s movie). The caller
+/// decodes each to PCM, concatenates in order, and clamps to the movie length —
+/// that yields the full track for segment banks while the clamp drops the
+/// duplicate takes for alternate-take banks. (Dynamic RE via Canary file-I/O
+/// tracing confirmed the movie→voice binding; this fixes the *decode* of `RT*`.)
+pub fn to_xma_riffs(slb: &[u8]) -> Vec<Vec<u8>> {
+    let mut out = Vec::new();
+    if find(slb, b"RIFF", 0).is_none() {
+        // Headerless single-stream bank.
+        if let Some(data) = slb.get(HEADERLESS_DATA_OFFSET..) {
+            if !data.is_empty() {
+                out.push(build_riff(&synth_xma1_fmt(2, 2, 48000), data));
+            }
+        }
+        return out;
+    }
+    let mut pos = 0usize;
+    while let Some(ri) = find(slb, b"RIFF", pos) {
+        // Parse this sub-wave's fmt + data (declared size is honest per sub-wave).
+        let Some(fi) = find(slb, b"fmt ", ri) else { break };
+        let Some(fsz) = le32(slb, fi + 4) else { break };
+        let Some(fmt_end) = fi.checked_add(8).and_then(|v| v.checked_add(fsz as usize)) else {
+            break;
+        };
+        if fmt_end > slb.len() {
+            break;
+        }
+        let Some(di) = find(slb, b"data", fi) else { break };
+        let Some(dsz) = le32(slb, di + 4) else { break };
+        let Some(ds) = di.checked_add(8) else { break };
+        let de = ds
+            .checked_add(dsz as usize)
+            .unwrap_or(slb.len())
+            .min(slb.len());
+        if let Some(data) = slb.get(ds..de) {
+            if !data.is_empty() {
+                out.push(build_riff(&slb[fi..fmt_end], data));
+            }
+        }
+        // Advance past this sub-wave's data to find the next RIFF.
+        pos = de.max(ri + 4);
+    }
+    out
+}
+
+/// Build a standalone XMA1 `RIFF/WAVE` from a `.slb` bank, decodable by FFmpeg's
+/// `xma1`. Returns `None` if the bank is too small / malformed. This is the
+/// FIRST sub-wave only; prefer [`to_xma_riffs`] for correct multi-segment banks.
+pub fn to_xma_riff(slb: &[u8]) -> Option<Vec<u8>> {
+    if let Some(ri) = find(slb, b"RIFF", 0) {
+        // RIFF layout: a `.slb` is an XACT bank of one or more sub-waves, each
+        // `[seek][RIFF: fmt + Dmmy pad + data][declared_size XMA bytes]`. Take the
+        // FIRST sub-wave, bounded by its **declared `data` size** (which is
+        // honest per sub-wave). Decoding to end-of-file instead would append the
+        // later sub-waves — for multi-take story movies those are ALTERNATE takes,
+        // which is what made S10–S16 play the wrong audio.
+        let fi = find(slb, b"fmt ", ri)?;
+        let fsz = le32(slb, fi + 4)? as usize;
+        let fmt_end = fi.checked_add(8)?.checked_add(fsz)?;
+        if fmt_end > slb.len() {
+            return None;
+        }
+        let fmt_chunk = &slb[fi..fmt_end];
+        let di = find(slb, b"data", fi)?;
+        let dsz = le32(slb, di + 4)? as usize;
+        let end = di.checked_add(8)?.checked_add(dsz)?.min(slb.len());
+        let data = slb.get(di + 8..end)?;
+        Some(build_riff(fmt_chunk, data))
+    } else {
+        // Headerless: 1392-byte header, then raw XMA1 (48 kHz, 2 channels).
+        let data = slb.get(HEADERLESS_DATA_OFFSET..)?;
+        if data.is_empty() {
+            return None;
+        }
+        Some(build_riff(&synth_xma1_fmt(2, 2, 48000), data))
+    }
+}
+
+/// Rebuild a standalone XMA1 `RIFF/WAVE` from a single-stream `.slb`, robust to
+/// the layout variants seen in `<lang>\etc\` radio clips. Unlike [`to_xma_riff`]
+/// (which assumes `RIFF → fmt → data` in order), this picks the **largest `data`
+/// chunk anywhere** in the bank — some radio banks store the audio *before* the
+/// trailing `RIFF`/`fmt` metadata (an empty post-`RIFF` `data` chunk), which the
+/// ordered scan misses. Pairs it with the first `fmt ` chunk; falls back to the
+/// headerless layout. Returns `None` only when no usable audio can be found.
+pub fn to_xma_riff_best(slb: &[u8]) -> Option<Vec<u8>> {
+    // Largest usable `data` chunk (bounded by its declared size and the buffer).
+    let mut best: Option<(usize, usize)> = None; // (data offset, usable payload len)
+    let mut i = 0;
+    while let Some(di) = find(slb, b"data", i) {
+        let declared = le32(slb, di + 4).unwrap_or(0) as usize;
+        let usable = declared.min(slb.len().saturating_sub(di + 8));
+        if best.map_or(true, |(_, b)| usable > b) {
+            best = Some((di, usable));
+        }
+        i = di + 4;
+    }
+    if let (Some(fi), Some((di, sz))) = (find(slb, b"fmt ", 0), best) {
+        if sz > 512 {
+            let fsz = le32(slb, fi + 4)? as usize;
+            let fmt_end = (fi + 8 + fsz).min(slb.len());
+            let data = slb.get(di + 8..di + 8 + sz)?;
+            return Some(build_riff(slb.get(fi..fmt_end)?, data));
+        }
+    }
+    // Headerless fallback: fixed data offset, synthesized XMA1 stereo/48k fmt.
+    let data = slb.get(HEADERLESS_DATA_OFFSET..)?;
+    (!data.is_empty()).then(|| build_riff(&synth_xma1_fmt(2, 2, 48000), data))
+}
+
+/// A minimal `fmt ` chunk carrying an XMA1 `XMAWAVEFORMAT` (one stream).
+fn synth_xma1_fmt(channels: u8, channel_mask: u16, rate: u32) -> Vec<u8> {
+    let mut fmt = Vec::with_capacity(40);
+    fmt.extend_from_slice(b"fmt ");
+    fmt.extend_from_slice(&32u32.to_le_bytes());
+    // XMAWAVEFORMAT header
+    fmt.extend_from_slice(&0x0165u16.to_le_bytes()); // wFormatTag = XMA1
+    fmt.extend_from_slice(&16u16.to_le_bytes()); // BitsPerSample
+    fmt.extend_from_slice(&0u16.to_le_bytes()); // EncodeOptions
+    fmt.extend_from_slice(&0u16.to_le_bytes()); // LargestSkip
+    fmt.extend_from_slice(&1u16.to_le_bytes()); // NumStreams
+    fmt.push(0); // LoopCount
+    fmt.push(3); // Version
+    // XMASTREAMFORMAT[0]
+    fmt.extend_from_slice(&(rate * channels as u32 * 2).to_le_bytes()); // PsuedoBytesPerSec
+    fmt.extend_from_slice(&rate.to_le_bytes()); // SampleRate
+    fmt.extend_from_slice(&0u32.to_le_bytes()); // LoopStart
+    fmt.extend_from_slice(&0u32.to_le_bytes()); // LoopEnd
+    fmt.push(4); // SubframeData
+    fmt.push(channels); // Channels
+    fmt.extend_from_slice(&channel_mask.to_le_bytes()); // ChannelMask
+    fmt
+}
+
+fn build_riff(fmt_chunk: &[u8], data: &[u8]) -> Vec<u8> {
+    let mut body = Vec::with_capacity(4 + fmt_chunk.len() + 8 + data.len());
+    body.extend_from_slice(b"WAVE");
+    body.extend_from_slice(fmt_chunk);
+    body.extend_from_slice(b"data");
+    body.extend_from_slice(&(data.len() as u32).to_le_bytes());
+    body.extend_from_slice(data);
+    let mut out = Vec::with_capacity(8 + body.len());
+    out.extend_from_slice(b"RIFF");
+    out.extend_from_slice(&(body.len() as u32).to_le_bytes());
+    out.extend_from_slice(&body);
+    out
+}
+
+fn find(hay: &[u8], needle: &[u8], from: usize) -> Option<usize> {
+    if from >= hay.len() {
+        return None;
+    }
+    hay[from..]
+        .windows(needle.len())
+        .position(|w| w == needle)
+        .map(|p| p + from)
+}
+
+fn le32(b: &[u8], o: usize) -> Option<u32> {
+    let s = b.get(o..o + 4)?;
+    Some(u32::from_le_bytes([s[0], s[1], s[2], s[3]]))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::hash::name_hash;
+
+    #[test]
+    fn movie_voice_names_and_hashes() {
+        assert_eq!(
+            movie_voice_name("RT07A", VoiceLang::English),
+            "eng\\Movie\\VOICE_RT07A.slb"
+        );
+        // Verified present in dat/sound.pak (VOICE_S00A == TOC entry 0x2A4F97D3).
+        assert_eq!(name_hash("eng\\Movie\\VOICE_RT07A.slb"), 0xA44A_EA1C);
+        assert_eq!(name_hash("eng\\Movie\\VOICE_S00A.slb"), 0x2A4F_97D3);
+    }
+
+    #[test]
+    fn takes_only_first_subwave() {
+        // Bank of TWO sub-waves; only the FIRST must be extracted (bounded by its
+        // declared `data` size), not everything to end-of-file.
+        let fmt = synth_xma1_fmt(2, 2, 48000);
+        let mut slb = vec![0xAB; 16];
+        slb.extend_from_slice(b"RIFF");
+        slb.extend_from_slice(&999u32.to_le_bytes()); // riff size (ignored)
+        slb.extend_from_slice(b"WAVE");
+        slb.extend_from_slice(&fmt);
+        slb.extend_from_slice(b"data");
+        slb.extend_from_slice(&4u32.to_le_bytes()); // declared: 4 bytes
+        slb.extend_from_slice(&[1, 2, 3, 4]); // sub-wave 1 XMA
+        slb.extend_from_slice(&[9, 9, 9, 9]); // a second sub-wave's bytes — excluded
+        let riff = to_xma_riff(&slb).unwrap();
+        assert_eq!(&riff[0..4], b"RIFF");
+        let dpos = find(&riff, b"data", 0).unwrap();
+        assert_eq!(le32(&riff, dpos + 4), Some(4));
+        assert_eq!(&riff[dpos + 8..], &[1, 2, 3, 4]);
+    }
+
+    #[test]
+    fn list_voice_clips_covers_movie_radio_and_briefing() {
+        // The standalone player must enumerate every spoken-line category: bound
+        // movie voices, in-mission radio (\etc\ + \Voice\), and briefing (\Briefing\,
+        // whose BR<NN>_<MM> names lack "VOICE"). Music (BGM_*) must stay excluded.
+        let mut tbl = Vec::new();
+        for s in [
+            "eng\\Movie\\VOICE_S13A.slb",
+            "eng\\etc\\VOICE_D_450.slb",
+            "eng\\Voice\\VOICE_ADAN_010.slb",
+            "eng\\Briefing\\BR01_01.slb",
+            "eng\\bgm\\BGM_001.slb",
+        ] {
+            tbl.extend_from_slice(s.as_bytes());
+            tbl.push(0);
+        }
+        let names: Vec<String> = list_voice_clips(&tbl, VoiceLang::English)
+            .into_iter()
+            .map(|c| c.name)
+            .collect();
+        for want in [
+            "eng\\Movie\\VOICE_S13A.slb",
+            "eng\\etc\\VOICE_D_450.slb",
+            "eng\\Voice\\VOICE_ADAN_010.slb",
+            "eng\\Briefing\\BR01_01.slb",
+        ] {
+            assert!(names.iter().any(|n| n == want), "missing {want}");
+        }
+        assert!(
+            !names.iter().any(|n| n.contains("BGM_")),
+            "music must not be listed as a voice clip"
+        );
+    }
+
+    #[test]
+    fn rebuilds_riff_from_headerless() {
+        let mut slb = vec![0u8; HEADERLESS_DATA_OFFSET];
+        slb.extend_from_slice(&[9, 8, 7, 6]);
+        let riff = to_xma_riff(&slb).unwrap();
+        let dpos = find(&riff, b"data", 0).unwrap();
+        assert_eq!(&riff[dpos + 8..], &[9, 8, 7, 6]);
+        // synthetic fmt advertises XMA1.
+        let fpos = find(&riff, b"fmt ", 0).unwrap();
+        assert_eq!(le32(&riff, fpos + 8).map(|v| v as u16), Some(0x0165));
+    }
+}

crates/sylpheed-formats/src/t8ad.rs

@@ -1,27 +1,25 @@
 //! `T8aD` — the game's 2D UI/HUD texture format.
 //!
-//! A linear (untiled) 32bpp surface stored **A8R8G8B8** (Xbox byte order), with
+//! A 32bpp **A8R8G8B8** (Xbox byte order) surface stored as **256×256 raster
-//! a small fixed-per-variant header:
+//! tiles in row-major order** — each tile prefixed by a 16-byte tile header, edge
+//! tiles clipped to the image bounds. Fully reversed 2026-07-17 from the file
+//! header and verified against the running game (title screen).
 //!
 //! ```text
 //! 0x00  4         Magic "T8aD"
 //! 0x14  4         width  (BE u32)
 //! 0x18  4         height (BE u32)
-//! 0x1c  4   type field → header size: 1→64  2→84  3→104  4→124  15→344
+//! 0x1c  4         tile count (BE u32) = ceil(w/256) * ceil(h/256)
-//! <header> w*h*4 bytes of A8R8G8B8 pixel data, row-major
+//! 0x2c  tiles*4   offset table: absolute byte offset of each row-major tile
+//! <off>  16       per-tile header (flags + tile w/h), then:
+//! <off+16>        tile_w * tile_h * 4 bytes of A8R8G8B8 pixels, row-major
 //! ```
 //!
-//! Static forensics over the disc showed ~85% of entries decode exactly with
+//! Surfaces ≤256px wide are a single tile column, so the first tile's pixels sit
-//! this rule; the rest (unknown type field or a `w*h*4` that doesn't fit — most
+//! at `44 + tiles*4 + 16 = 64` — which is why the old "type→header size 64/84/…"
-//! likely DXT / palettized variants) return `None` rather than a wrong image.
+//! rule (header = 44 + tiles*20) happened to decode small textures correctly: for
-//!
+//! one tile it lands on the same pixel start. Wide textures were garbled because
-//! Keying the header size off the type field (not `len - w*h*4`) makes the
+//! the offset table + 16-byte per-tile headers weren't accounted for.
-//! decoder **container-safe**: RATC/LSTA hand us a child slice that may have
-//! trailing padding before the next child, so we must not infer the header from
-//! the slice length.
-//!
-//! ⚠️ Colour correctness (channel order / endianness / sRGB) is unverified
-//! against the running game — see the RE backlog.
 
 /// Magic at the start of every T8aD surface.
 pub const T8AD_MAGIC: [u8; 4] = *b"T8aD";
@@ -39,54 +37,74 @@ pub fn is_t8ad(bytes: &[u8]) -> bool {
     bytes.len() >= 4 && bytes[0..4] == T8AD_MAGIC
 }
 
-/// Header size in bytes for a given type field (`@0x1c`), or `None` for an
-/// unrecognized variant.
-fn header_size(type_field: u32) -> Option<usize> {
-    match type_field {
-        1 => Some(64),
-        2 => Some(84),
-        3 => Some(104),
-        4 => Some(124),
-        15 => Some(344),
-        _ => None,
-    }
-}
-
 #[inline]
 fn be32(b: &[u8], off: usize) -> u32 {
     u32::from_be_bytes([b[off], b[off + 1], b[off + 2], b[off + 3]])
 }
 
+/// Side of the square storage tile, in texels, and the per-tile header size.
+const TILE: usize = 256;
+const TILE_HDR: usize = 16;
+
 /// Decode a T8aD surface from a slice whose first bytes ARE the magic. Returns
 /// `None` for non-T8aD input or a variant we can't decode as RGBA (never guesses).
+///
+/// Layout (reversed from the header + verified against the running game):
+/// a 44-byte base header, then a `tiles`-entry big-endian u32 **offset table** at
+/// `0x2c`, where `tiles` = the field at `0x1c` = `ceil(w/256) * ceil(h/256)`.
+/// Each entry is the absolute byte offset of a **row-major** 256×256 tile; every
+/// tile is a 16-byte tile header followed by `tile_w*tile_h*4` A8R8G8B8 pixels,
+/// edge tiles clipped to the image bounds.
 pub fn parse(bytes: &[u8]) -> Option<T8adImage> {
     if !is_t8ad(bytes) || bytes.len() < 0x40 {
         return None;
     }
-    let width = be32(bytes, 0x14);
+    let width = be32(bytes, 0x14) as usize;
-    let height = be32(bytes, 0x18);
+    let height = be32(bytes, 0x18) as usize;
     if !(1..=4096).contains(&width) || !(1..=4096).contains(&height) {
         return None;
     }
-    let header = header_size(be32(bytes, 0x1c))?;
+    let tiles = be32(bytes, 0x1c) as usize;
-    let n = (width as usize) * (height as usize) * 4;
+    let cols = width.div_ceil(TILE);
-    if bytes.len() < header + n {
+    let rows = height.div_ceil(TILE);
-        return None; // DXT/palettized/short variant — defer, don't misdecode
+    // The field at 0x1c must be the tile count; otherwise it's a variant we don't
+    // decode (e.g. DXT / palettized) — defer rather than misdecode.
+    if tiles == 0 || tiles != cols * rows {
+        return None;
+    }
+    const TABLE: usize = 0x2c;
+    if bytes.len() < TABLE + tiles * 4 {
+        return None;
     }
 
+    let mut rgba = vec![0u8; width * height * 4];
+    for ty in 0..rows {
+        for tx in 0..cols {
+            let tile = ty * cols + tx;
+            let pixels = be32(bytes, TABLE + tile * 4) as usize + TILE_HDR;
+            let tw = TILE.min(width - tx * TILE);
+            let th = TILE.min(height - ty * TILE);
+            if pixels + tw * th * 4 > bytes.len() {
+                return None; // truncated / not the layout we expect
+            }
+            for row in 0..th {
+                let mut s = pixels + row * tw * 4;
+                let mut d = ((ty * TILE + row) * width + tx * TILE) * 4;
+                for _ in 0..tw {
                     // A8R8G8B8 → RGBA8.
-    let src = &bytes[header..header + n];
+                    rgba[d] = bytes[s + 1];
-    let mut rgba = vec![0u8; n];
+                    rgba[d + 1] = bytes[s + 2];
-    for (px, out) in src.chunks_exact(4).zip(rgba.chunks_exact_mut(4)) {
+                    rgba[d + 2] = bytes[s + 3];
-        let (a, r, g, b) = (px[0], px[1], px[2], px[3]);
+                    rgba[d + 3] = bytes[s];
-        out[0] = r;
+                    s += 4;
-        out[1] = g;
+                    d += 4;
-        out[2] = b;
+                }
-        out[3] = a;
+            }
+        }
     }
     Some(T8adImage {
-        width,
+        width: width as u32,
-        height,
+        height: height as u32,
         rgba,
     })
 }
@@ -95,13 +113,17 @@ pub fn parse(bytes: &[u8]) -> Option<T8adImage> {
 mod tests {
     use super::*;
 
-    /// Build a synthetic type-1 (header 64) T8aD with a known A8R8G8B8 pattern.
+    /// Build a synthetic single-tile T8aD (`w,h ≤ 256`) with a known A8R8G8B8
+    /// pattern: base header + 1-entry offset table + 16-byte tile header + pixels.
     fn synth(w: u32, h: u32) -> Vec<u8> {
-        let mut b = vec![0u8; 64];
+        assert!(w <= 256 && h <= 256);
+        let mut b = vec![0u8; 0x2c];
         b[0..4].copy_from_slice(&T8AD_MAGIC);
         b[0x14..0x18].copy_from_slice(&w.to_be_bytes());
         b[0x18..0x1c].copy_from_slice(&h.to_be_bytes());
-        b[0x1c..0x20].copy_from_slice(&1u32.to_be_bytes()); // type 1 → header 64
+        b[0x1c..0x20].copy_from_slice(&1u32.to_be_bytes()); // 1 tile
+        b.extend_from_slice(&0x30u32.to_be_bytes()); // offset table: tile 0 @ 0x30
+        b.extend_from_slice(&[0u8; 16]); // 16-byte tile header → pixels at 0x40
         for i in 0..(w * h) {
             b.extend_from_slice(&[(i & 0xff) as u8, 0x24, 0x63, 0xB2]); // A, R, G, B
         }
@@ -121,8 +143,31 @@ mod tests {
     }
 
     #[test]
-    fn rejects_unknown_variant_and_short() {
+    fn assembles_row_major_tiles_via_offset_table() {
-        // type field 7 (unknown) → None
+        // 300×1 → 2 tiles: (0,0)=256×1 red, (1,0)=44×1 blue, each +16-byte header.
+        let (w, h): (u32, u32) = (300, 1);
+        let mut b = vec![0u8; 0x2c];
+        b[0..4].copy_from_slice(&T8AD_MAGIC);
+        b[0x14..0x18].copy_from_slice(&w.to_be_bytes());
+        b[0x18..0x1c].copy_from_slice(&h.to_be_bytes());
+        b[0x1c..0x20].copy_from_slice(&2u32.to_be_bytes()); // 2 tiles
+        let off0 = 0x2c + 2 * 4; // after the 2-entry table
+        let off1 = off0 + 16 + 256 * 4; // tile-0 header + its 256 pixels
+        b.extend_from_slice(&(off0 as u32).to_be_bytes());
+        b.extend_from_slice(&(off1 as u32).to_be_bytes());
+        b.extend_from_slice(&[0u8; 16]);
+        b.extend_from_slice(&[0xFF, 0xFF, 0, 0].repeat(256)); // A,R,G,B red
+        b.extend_from_slice(&[0u8; 16]);
+        b.extend_from_slice(&[0xFF, 0, 0, 0xFF].repeat(44)); // A,R,G,B blue
+        let img = parse(&b).expect("decodes");
+        assert_eq!((img.width, img.height), (300, 1));
+        assert_eq!(&img.rgba[0..4], &[0xFF, 0, 0, 0xFF]); // tile 0 → red
+        assert_eq!(&img.rgba[256 * 4..256 * 4 + 4], &[0, 0, 0xFF, 0xFF]); // tile 1 → blue
+    }
+
+    #[test]
+    fn rejects_wrong_tilecount_and_short() {
+        // tile-count field that isn't ceil(w/256)*ceil(h/256) → None
         let mut b = synth(2, 2);
         b[0x1c..0x20].copy_from_slice(&7u32.to_be_bytes());
         assert!(parse(&b).is_none());

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 +1,426 @@
+//! 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::{material_groups, node_transforms, submesh_albedos, 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 hero_ship_submesh_material_graph() {
+    let Some(dir) = res3d_dir() else {
+        eprintln!("SKIP: resource3d dir not found (set SYLPHEED_RES3D)");
+        return;
+    };
+    // The XBG7 node/material graph tags each sub-mesh record with its own albedo
+    // (`rou_…_col` → TX2D name). DeltaSaber_T `f001`: the 7 detail parts use
+    // bdy_04/06/07 — NOT the body's bdy_01a — which is the per-part texturing the
+    // viewer needs so fins aren't painted with the hull map.
+    let bytes = std::fs::read(dir.join("DeltaSaber_T.xpr")).unwrap();
+    let mut map = std::collections::HashMap::new();
+    for (v, i, name) in submesh_albedos(&bytes, "f001") {
+        map.insert((v, i), name);
+    }
+    // (vtx, idx) → expected albedo (rou_ stripped = TX2D name).
+    assert_eq!(map.get(&(314, 645)).map(String::as_str), Some("f001_bdy_04_col"));
+    assert_eq!(map.get(&(48, 84)).map(String::as_str), Some("f001_bdy_04_col"));
+    assert_eq!(map.get(&(96, 180)).map(String::as_str), Some("f001_bdy_06_col"));
+    assert_eq!(map.get(&(60, 108)).map(String::as_str), Some("f001_bdy_07_col"));
+    // Names strip cleanly to real TX2D resources.
+    let tex = sylpheed_formats::texture::X360Texture::texture_names(&bytes);
+    for name in map.values() {
+        assert!(tex.iter().any(|t| t == name), "albedo {name} not a TX2D resource");
+    }
+}
+
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn hero_ship_body_splits_by_material() {
+    let Some(dir) = res3d_dir() else {
+        eprintln!("SKIP: resource3d dir not found (set SYLPHEED_RES3D)");
+        return;
+    };
+    // The merged body (sub0 = 24561 indices / 8187 tris) is drawn as 9 groups
+    // sharing one index buffer, each with its own albedo (bdy_01a hull, bdy_01b,
+    // bdy_02/03, and the `daiza` hangar stand). `material_groups` recovers the
+    // cumulative-offset chain so the viewer can texture each slice.
+    let bytes = std::fs::read(dir.join("DeltaSaber_T.xpr")).unwrap();
+    let groups = material_groups(&bytes, "f001", 24561);
+    assert_eq!(groups.len(), 9, "body must split into 9 material groups");
+    // Offsets chain from 0 and cover the whole index buffer exactly.
+    let mut cursor = 0usize;
+    for g in &groups {
+        assert_eq!(g.idx_offset, cursor, "group offsets must be contiguous");
+        cursor += g.idx_count;
+    }
+    assert_eq!(cursor, 24561, "groups must cover all body indices");
+    // Expected per-group albedo (verified against the GPU draw log).
+    let albedos: Vec<&str> = groups.iter().map(|g| g.albedo.as_str()).collect();
+    assert_eq!(
+        albedos,
+        [
+            "f001_bdy_01a_col",
+            "f001_bdy_01a_col",
+            "f001_bdy_01a_col",
+            "f001_bdy_01b_col",
+            "f001_bdy_01a_col",
+            "f001_bdy_01b_col",
+            "f001_bdy_02_col",
+            "f001_bdy_03_col",
+            "f001_bdy_daiza_col",
+        ]
+    );
+    // A single-material detail part (sub1 = 645 indices) → one group, bdy_04.
+    let fin = material_groups(&bytes, "f001", 645);
+    assert_eq!(fin.len(), 1);
+    assert_eq!(fin[0].albedo, "f001_bdy_04_col");
+    assert_eq!(fin[0].idx_offset, 0);
+}
+
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn hero_ship_node_transforms_move_fins_to_tail() {
+    let Some(dir) = res3d_dir() else {
+        eprintln!("SKIP: resource3d dir not found (set SYLPHEED_RES3D)");
+        return;
+    };
+    // For the DeltaSaber, node_transforms returns the MEASURED runtime placements
+    // (Canary draw-log WVP; see mesh.rs::DELTASABER_MEASURED) — the fin assemblies
+    // are mounted OUT on the nacelles (world X ≈ −4.4..−6.6), an offset the static
+    // graph omits. Body identity; each part a left/right mirrored pair. Vertex
+    // space: X right, Y up, Z fore/aft (engines at −Z).
+    let bytes = std::fs::read(dir.join("DeltaSaber_T.xpr")).unwrap();
+    let places = node_transforms(&bytes, "f001");
+    assert!(!places.is_empty(), "expected node placements");
+    // Body (sub 0) identity, drawn once.
+    let body: Vec<_> = places.iter().filter(|p| p.sub_index == 0).collect();
+    assert_eq!(body.len(), 1, "body drawn once");
+    assert!(body[0].t.iter().all(|c| c.abs() < 0.1), "body must not translate");
+    assert!(!body[0].reflect, "body is not mirrored");
+    // Fin bdy_04 (sub 1): mirrored pair near the tail, mounted OUT on the nacelle
+    // (|X| ≈ 5.2, not the graph's inboard ≈1.1).
+    let fins: Vec<_> = places.iter().filter(|p| p.sub_index == 1).collect();
+    assert_eq!(fins.len(), 2, "V-tail is a mirrored pair");
+    assert!(fins.iter().all(|f| (f.t[2] + 9.72).abs() < 0.3), "fin fore/aft ≈ −9.7");
+    assert!(fins.iter().all(|f| f.t[0].abs() > 4.0), "fin mounted out on the nacelle");
+    assert!(fins[0].t[0] * fins[1].t[0] < 0.0, "the pair mirrors across X");
+    assert!(fins.iter().any(|f| f.reflect), "one of the pair is reflected");
+    // Winglet bdy_06 (sub 4): on the nacelle (|X| ≈ 6.6, Z ≈ −15.5).
+    let wings: Vec<_> = places.iter().filter(|p| p.sub_index == 4).collect();
+    assert_eq!(wings.len(), 2, "L/R winglet pair");
+    assert!(wings.iter().all(|p| p.t[0].abs() > 6.0 && (p.t[2] + 15.5).abs() < 0.3),
+        "winglets out on the nacelle");
+    assert!(wings[0].t[0] * wings[1].t[0] < 0.0, "winglets mirror across X");
+    // Small fin bdy_10 (sub 6): inboard nacelle stub (|X| ≈ 4.45, Z ≈ −17).
+    let sfins: Vec<_> = places.iter().filter(|p| p.sub_index == 6).collect();
+    assert_eq!(sfins.len(), 2, "L/R small-fin pair");
+    assert!(sfins.iter().all(|p| (p.t[0].abs() - 4.45).abs() < 0.3), "small fins on the stub");
+    assert!(sfins[0].t[0] * sfins[1].t[0] < 0.0, "small fins mirror across X");
+}
+
+#[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()));
+            }
+        }
+    }
+}
+
+/// The hero ship's **grouped-pool** layout decodes fully: `DeltaSaber_T.xpr`'s
+/// `f001` resource is one vertex+index pool shared by 8 sub-meshes (body + 7
+/// detail parts). Reversed statically and cross-checked against a Canary GPU
+/// draw-log capture — every sub-mesh's stored normals agree with its triangle
+/// winding (0 degenerate, full vertex coverage). This is the layout the old
+/// per-block adjacency anchor rendered as a spiky phantom.
+#[test]
+#[ignore = "requires extracted disc models — set SYLPHEED_RES3D"]
+fn hero_ship_grouped_pool_decodes() {
+    let Some(dir) = res3d_dir() else {
+        eprintln!("SKIP: resource3d dir not found (set SYLPHEED_RES3D)");
+        return;
+    };
+    let bytes = std::fs::read(dir.join("DeltaSaber_T.xpr")).unwrap();
+    let models = Xbg7Model::stage_models(&bytes);
+
+    // The neutral pose `f001` (the mnv*/turn180 resources are animation poses).
+    let f001 = models.iter().find(|m| m.name == "f001").expect("f001 decoded");
+    assert_eq!(f001.meshes.len(), 8, "body + 7 detail sub-meshes");
+    let (v, t) = f001.totals();
+    assert_eq!(v, 11607, "vertex total across sub-meshes");
+    assert_eq!(t, 8650, "triangle total (body 8187 + 7 parts)");
+
+    // Sub-mesh 0 is the body: exactly the draw-log-verified geometry.
+    let body = &f001.meshes[0];
+    assert_eq!(body.positions.len(), 10891);
+    assert_eq!(body.indices.len(), 24561);
+
+    for (i, m) in f001.meshes.iter().enumerate() {
+        // Every index in range.
+        assert!(
+            m.indices.iter().all(|&ix| (ix as usize) < m.positions.len()),
+            "sub{i}: index out of range"
+        );
+        // Correct alignment ⇒ unit normals, and the winding agrees with them
+        // (the decisive correctness signal, ~1.0, not the ~0.5 of a mis-carve).
+        let n = m.normals.len();
+        assert_eq!(n, m.positions.len(), "sub{i}: a normal per vertex");
+        let mean_nlen: f32 = m
+            .normals
+            .iter()
+            .map(|nv| (nv[0] * nv[0] + nv[1] * nv[1] + nv[2] * nv[2]).sqrt())
+            .sum::<f32>()
+            / n as f32;
+        assert!((mean_nlen - 1.0).abs() < 0.05, "sub{i}: mean |normal| {mean_nlen} ≠ 1");
+
+        let mut agree = 0usize;
+        let mut counted = 0usize;
+        for tri in m.indices.chunks_exact(3) {
+            let (a, b, c) = (tri[0] as usize, tri[1] as usize, tri[2] as usize);
+            let (pa, pb, pc) = (m.positions[a], m.positions[b], m.positions[c]);
+            let u = [pb[0] - pa[0], pb[1] - pa[1], pb[2] - pa[2]];
+            let w = [pc[0] - pa[0], pc[1] - pa[1], pc[2] - pa[2]];
+            let f = [
+                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 f[0] * f[0] + f[1] * f[1] + f[2] * f[2] < 1e-12 {
+                continue;
+            }
+            let sn = [
+                m.normals[a][0] + m.normals[b][0] + m.normals[c][0],
+                m.normals[a][1] + m.normals[b][1] + m.normals[c][1],
+                m.normals[a][2] + m.normals[b][2] + m.normals[c][2],
+            ];
+            if f[0] * sn[0] + f[1] * sn[1] + f[2] * sn[2] > 0.0 {
+                agree += 1;
+            }
+            counted += 1;
+        }
+        let na = agree as f32 / counted.max(1) as f32;
+        assert!(
+            na.max(1.0 - na) > 0.95,
+            "sub{i}: winding-vs-normal agreement {na} — mis-carved (should be ~1.0 or ~0.0)"
+        );
+    }
+}
+
+/// 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-formats/tests/movie_manifest_disc.rs

@@ -0,0 +1,102 @@
+//! Real-disc test for the movie manifest → voice binding. Skipped without
+//! `SYLPHEED_DISC`.
+
+use std::path::PathBuf;
+
+use sylpheed_formats::movie_manifest;
+use sylpheed_formats::slb::VoiceLang;
+use sylpheed_formats::PakArchive;
+
+fn disc_root() -> Option<PathBuf> {
+    let p = PathBuf::from(std::env::var("SYLPHEED_DISC").ok()?);
+    p.join("dat").is_dir().then_some(p)
+}
+
+/// Read the manifest + `eng\sounds.tbl` out of `tables.pak`.
+fn load_manifest_and_sounds(root: &PathBuf) -> (Vec<u8>, Vec<u8>) {
+    let pak = PakArchive::open(root.join("dat/tables.pak")).unwrap();
+    let manifest = pak
+        .entries()
+        .iter()
+        .find_map(|e| pak.read(e).ok().filter(|b| movie_manifest::is_manifest(b)))
+        .expect("movie manifest present in tables.pak");
+    let sounds = pak
+        .read_by_name("eng\\sounds.tbl")
+        .expect("sounds.tbl present")
+        .unwrap();
+    (manifest, sounds)
+}
+
+#[test]
+fn binds_and_resolves_movie_voice() {
+    let Some(root) = disc_root() else {
+        eprintln!("SKIP: set SYLPHEED_DISC");
+        return;
+    };
+    let (manifest, sounds) = load_manifest_and_sounds(&root);
+    let entries = movie_manifest::parse(&manifest);
+    assert!(entries.len() > 90, "expected ~101 movies, got {}", entries.len());
+
+    // Standard story movie → VOICE_<movie> in <lang>\Movie\.
+    assert_eq!(
+        movie_manifest::voice_token(&manifest, "S13A").as_deref(),
+        Some("VOICE_S13A")
+    );
+    assert_eq!(
+        movie_manifest::resolve_voice_entry(&manifest, &sounds, "S13A", VoiceLang::English)
+            .as_deref(),
+        Some("eng\\Movie\\VOICE_S13A.slb")
+    );
+
+    // A resupply movie bound to an in-mission radio clip → VOICE_D_* in \etc\.
+    // (This is what a `VOICE_<movie>` guess would miss entirely.)
+    assert_eq!(
+        movie_manifest::voice_token(&manifest, "hokyu_LS_s02A").as_deref(),
+        Some("VOICE_D_450")
+    );
+    assert_eq!(
+        movie_manifest::resolve_voice_entry(
+            &manifest,
+            &sounds,
+            "hokyu_LS_s02A",
+            VoiceLang::English
+        )
+        .as_deref(),
+        Some("eng\\etc\\VOICE_D_450.slb")
+    );
+
+    // A resupply movie the game leaves unbound in the manifest → no DIRECT
+    // binding. (Extending to unbound movies by shared demo line was verified
+    // WRONG against the running game, so we do NOT resolve these.)
+    let e = entries
+        .iter()
+        .find(|e| e.movie == "hokyu_DS_s13A")
+        .expect("hokyu_DS_s13A present in manifest");
+    assert_eq!(e.voice_token, None, "hokyu_DS_s13A has no direct voice binding");
+    assert_eq!(
+        movie_manifest::resolve_voice_entry(&manifest, &sounds, "hokyu_DS_s13A", VoiceLang::English),
+        None
+    );
+
+    // Every resolved voice entry must actually exist in sound.pak.
+    let snd = std::fs::read(root.join("dat/sound.pak")).unwrap();
+    let keys: std::collections::HashSet<u32> = PakArchive::parse_toc(&snd)
+        .unwrap()
+        .iter()
+        .map(|e| e.name_hash)
+        .collect();
+    let mut resolved = 0;
+    let mut missing = Vec::new();
+    for e in &entries {
+        if let Some(full) =
+            movie_manifest::resolve_voice_entry(&manifest, &sounds, &e.movie, VoiceLang::English)
+        {
+            resolved += 1;
+            if !keys.contains(&sylpheed_formats::hash::name_hash(&full)) {
+                missing.push(full);
+            }
+        }
+    }
+    assert!(resolved > 80, "expected 80+ voiced movies, got {resolved}");
+    assert!(missing.is_empty(), "resolved but absent in sound.pak: {missing:?}");
+}

crates/sylpheed-formats/tests/movie_subtitle_disc.rs

@@ -0,0 +1,63 @@
+//! Real-disc test for the movie subtitle chain. Skipped when the extracted disc
+//! is absent (set `SYLPHEED_DISC` to the extract root to enable).
+
+use std::path::PathBuf;
+
+use sylpheed_formats::movie_subtitle::{self, SubLang};
+use sylpheed_formats::PakArchive;
+
+fn disc_root() -> Option<PathBuf> {
+    let p = PathBuf::from(std::env::var("SYLPHEED_DISC").ok()?);
+    p.join("dat").is_dir().then_some(p)
+}
+
+#[test]
+fn resolves_english_radio_subtitles() {
+    let Some(root) = disc_root() else {
+        eprintln!("SKIP: set SYLPHEED_DISC");
+        return;
+    };
+    let lang = SubLang::English;
+    let lang_pak = PakArchive::open(root.join(format!("dat/movie/{}.pak", lang.pak_code()))).unwrap();
+    let text_pak =
+        PakArchive::open(root.join(format!("dat/GP_MAIN_GAME_{}.pak", lang.game_code()))).unwrap();
+
+    // A radio movie: real timed English dialogue.
+    let cues = movie_subtitle::load("RT01C_1", &lang_pak, &text_pak);
+    assert!(!cues.is_empty(), "RT01C_1 should have timed cues");
+    assert!(cues[0].start >= 0.0 && cues[0].start < 5.0);
+    assert!(
+        cues.iter().any(|c| c.text.contains("Rhino Leader")),
+        "expected recognizable dialogue, got: {:?}",
+        cues.iter().map(|c| &c.text).collect::<Vec<_>>()
+    );
+
+    // No MSG_DEMO keys should leak into resolved caption text.
+    assert!(
+        cues.iter().all(|c| !c.text.contains("MSG_DEMO")),
+        "unresolved key leaked: {:?}",
+        cues.iter().map(|c| &c.text).collect::<Vec<_>>()
+    );
+
+    // A resupply movie: the canonical resupply line.
+    let hokyu = movie_subtitle::load("hokyu_DS_s02A", &lang_pak, &text_pak);
+    assert!(hokyu.iter().any(|c| c.text.contains("Resupply complete")));
+
+    // S00A carries inline narration text (not MSG_DEMO refs) with timecodes.
+    let story = movie_subtitle::load("S00A", &lang_pak, &text_pak);
+    assert!(!story.is_empty() && story.iter().any(|c| c.text.contains("27th century")));
+
+    // S13A stores a two-line caption as two consecutive text tokens sharing one
+    // timing; both lines must survive (the "Look at it father" regression — the
+    // first line used to be dropped).
+    let s13 = movie_subtitle::load("S13A", &lang_pak, &text_pak);
+    let multiline = s13
+        .iter()
+        .find(|c| c.text.contains("Look at it father"))
+        .expect("S13A should contain the 'Look at it father' caption");
+    assert_eq!(
+        multiline.text, "Look at it father\n& beautiful isn't it",
+        "both lines of the caption must be present"
+    );
+    assert!((multiline.start - 74.8).abs() < 0.1, "start {}", multiline.start);
+}

crates/sylpheed-formats/tests/slb_disc.rs

@@ -0,0 +1,104 @@
+//! Real-disc test: read a movie voice `.slb` out of the segmented `sound.pak`
+//! and rebuild a decodable XMA1 RIFF. Skipped without `SYLPHEED_DISC`.
+
+use std::fs::File;
+use std::io::{Read, Seek, SeekFrom};
+use std::path::PathBuf;
+
+use sylpheed_formats::hash::name_hash;
+use sylpheed_formats::slb::{self, VoiceLang};
+use sylpheed_formats::PakArchive;
+
+fn disc_root() -> Option<PathBuf> {
+    let p = PathBuf::from(std::env::var("SYLPHEED_DISC").ok()?);
+    p.join("dat").is_dir().then_some(p)
+}
+
+/// Read `[off, off+size)` from `dat/sound.p00..` (segments concatenated).
+fn read_range(root: &PathBuf, mut off: u64, size: usize) -> Vec<u8> {
+    let mut out = Vec::with_capacity(size);
+    let mut need = size;
+    for i in 0..100u32 {
+        if need == 0 {
+            break;
+        }
+        let seg = root.join(format!("dat/sound.p{i:02}"));
+        let Ok(meta) = std::fs::metadata(&seg) else { break };
+        let len = meta.len();
+        if off >= len {
+            off -= len;
+            continue;
+        }
+        let mut f = File::open(&seg).unwrap();
+        f.seek(SeekFrom::Start(off)).unwrap();
+        let take = need.min((len - off) as usize);
+        let start = out.len();
+        out.resize(start + take, 0);
+        f.read_exact(&mut out[start..]).unwrap();
+        need -= take;
+        off = 0;
+    }
+    out
+}
+
+#[test]
+fn reads_and_rebuilds_movie_voice_riff() {
+    let Some(root) = disc_root() else {
+        eprintln!("SKIP: set SYLPHEED_DISC");
+        return;
+    };
+    let toc = std::fs::read(root.join("dat/sound.pak")).unwrap();
+    let entries = PakArchive::parse_toc(&toc).unwrap();
+    assert_eq!(entries.len(), 9519);
+
+    for movie in ["S00A", "RT07A", "RT02D_1"] {
+        // RT02D_1 is a headerless variant; S00A/RT07A embed a RIFF.
+        let key = name_hash(&slb::movie_voice_name(movie, VoiceLang::English));
+        let idx = entries
+            .binary_search_by_key(&key, |e| e.name_hash)
+            .unwrap_or_else(|_| panic!("{movie} voice missing"));
+        let e = &entries[idx];
+        let bank = read_range(&root, e.offset as u64, e.comp_size as usize);
+
+        let riff = slb::to_xma_riff(&bank).expect("rebuild RIFF");
+        assert_eq!(&riff[0..4], b"RIFF", "{movie}");
+        assert_eq!(&riff[8..12], b"WAVE", "{movie}");
+        // fmt chunk advertises XMA1 (tag 0x0165).
+        let fpos = riff.windows(4).position(|w| w == b"fmt ").unwrap();
+        let tag = u16::from_le_bytes([riff[fpos + 8], riff[fpos + 9]]);
+        assert_eq!(tag, 0x0165, "{movie} should be XMA1");
+        // data chunk carries the XMA payload.
+        let dpos = riff.windows(4).position(|w| w == b"data").unwrap();
+        let dsz = u32::from_le_bytes([
+            riff[dpos + 4],
+            riff[dpos + 5],
+            riff[dpos + 6],
+            riff[dpos + 7],
+        ]) as usize;
+        assert!(dsz > 10_000, "{movie} data too small: {dsz}");
+        assert_eq!(riff.len(), dpos + 8 + dsz, "{movie} data length consistent");
+    }
+}
+
+#[test]
+fn enumerates_voice_clips_from_sounds_tbl() {
+    let Some(root) = disc_root() else {
+        eprintln!("SKIP: set SYLPHEED_DISC");
+        return;
+    };
+    let pak = PakArchive::open(root.join("dat/tables.pak")).unwrap();
+    let tbl = pak.read_by_name("eng\\sounds.tbl").unwrap().unwrap();
+    let clips = slb::list_voice_clips(&tbl, VoiceLang::English);
+    assert!(clips.len() > 1000, "expected many voice clips, got {}", clips.len());
+    // Every clip is an eng voice path present in sound.pak.
+    let sound = std::fs::read(root.join("dat/sound.pak")).unwrap();
+    let keys: std::collections::HashSet<u32> =
+        PakArchive::parse_toc(&sound).unwrap().iter().map(|e| e.name_hash).collect();
+    let present = clips.iter().filter(|c| keys.contains(&name_hash(&c.name))).count();
+    assert!(
+        present as f32 / clips.len() as f32 > 0.95,
+        "{present}/{} clips resolve in sound.pak",
+        clips.len()
+    );
+    assert!(clips.iter().any(|c| c.speaker == "ADAN"));
+}

crates/sylpheed-viewer/src/camera.rs

@@ -8,6 +8,10 @@
 
 use bevy::prelude::*;
 use bevy::input::mouse::{MouseMotion, MouseWheel};
+use bevy::render::render_asset::RenderAssetUsages;
+use bevy::render::render_resource::{
+    Extent3d, TextureDimension, TextureFormat, TextureViewDescriptor, TextureViewDimension,
+};
 
 pub struct OrbitCameraPlugin;
 
@@ -32,6 +36,10 @@ 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 {
@@ -44,29 +52,114 @@ impl Default for OrbitCamera {
             orbit_sensitivity: 0.005,
             zoom_sensitivity: 0.3,
             pan_sensitivity: 0.003,
+            min_radius: 0.05,
+            max_radius: 500.0,
         }
     }
 }
 
-fn spawn_camera(mut commands: Commands) {
+fn spawn_camera(mut commands: Commands, mut images: ResMut<Assets<Image>>) {
     let orbit = OrbitCamera::default();
     let transform = orbit_transform(&orbit);
 
+    // The game's ship shader reflects a shared HDR environment cubemap off the
+    // hull (three cube samples — see the shader RE). Reproduce that "look" with a
+    // procedural sky cube feeding Bevy's image-based lighting, so metallic
+    // surfaces reflect an environment instead of reading flat. Procedural (not a
+    // bundled KTX2) so it also works on WASM with no extra assets.
+    let env = make_env_cubemap(&mut images);
+
     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,
+        EnvironmentMapLight {
+            diffuse_map: env.clone(),
+            specular_map: env,
+            intensity: 900.0,
+            ..default()
+        },
     ));
 }
 
+/// Build a small procedural sky cubemap (6 faces) for image-based lighting.
+///
+/// A vertical gradient (bright zenith → warm horizon → dark ground) plus a warm
+/// "sun" highlight roughly where the game's key light sits — enough for a
+/// metallic hull to read as reflective rather than flat. Stored as an sRGB cube
+/// (filterable, no half-float encoding needed); a single mip means reflections
+/// stay sharp (fine for a shiny ship).
+fn make_env_cubemap(images: &mut Assets<Image>) -> Handle<Image> {
+    let size: i32 = 64;
+    // Approx key-light direction from the RE (PS c32 ≈ (0,-0.76,-0.65), i.e. light
+    // arriving from top-front); place the bright spot there.
+    let sun = Vec3::new(0.0, 0.85, 0.5).normalize();
+    let zenith = Vec3::new(0.70, 0.80, 0.95);
+    let horizon = Vec3::new(0.42, 0.45, 0.50);
+    let ground = Vec3::new(0.09, 0.09, 0.11);
+    let srgb = |c: f32| (c.clamp(0.0, 1.0).powf(1.0 / 2.2) * 255.0).round() as u8;
+
+    let mut data: Vec<u8> = Vec::with_capacity((size * size * 6 * 4) as usize);
+    for face in 0..6 {
+        for y in 0..size {
+            for x in 0..size {
+                let u = (x as f32 + 0.5) / size as f32 * 2.0 - 1.0;
+                let v = (y as f32 + 0.5) / size as f32 * 2.0 - 1.0;
+                // Standard cube-face → direction mapping.
+                let d = match face {
+                    0 => Vec3::new(1.0, -v, -u),
+                    1 => Vec3::new(-1.0, -v, u),
+                    2 => Vec3::new(u, 1.0, v),
+                    3 => Vec3::new(u, -1.0, -v),
+                    4 => Vec3::new(u, -v, 1.0),
+                    _ => Vec3::new(-u, -v, -1.0),
+                }
+                .normalize();
+                let mut col = if d.y >= 0.0 {
+                    horizon.lerp(zenith, (d.y).powf(0.6))
+                } else {
+                    horizon.lerp(ground, (-d.y).powf(0.5))
+                };
+                let s = d.dot(sun).max(0.0).powf(60.0);
+                col += Vec3::new(1.0, 0.85, 0.6) * s * 1.5;
+                data.extend_from_slice(&[srgb(col.x), srgb(col.y), srgb(col.z), 255]);
+            }
+        }
+    }
+
+    let mut image = Image::new(
+        Extent3d {
+            width: size as u32,
+            height: size as u32,
+            depth_or_array_layers: 6,
+        },
+        TextureDimension::D2,
+        data,
+        TextureFormat::Rgba8UnormSrgb,
+        RenderAssetUsages::RENDER_WORLD,
+    );
+    image.texture_view_descriptor = Some(TextureViewDescriptor {
+        dimension: Some(TextureViewDimension::Cube),
+        ..default()
+    });
+    images.add(image)
+}
+
 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 +192,20 @@ 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/lib.rs

@@ -72,9 +72,28 @@ pub fn run() {
 }
 
 fn setup_scene(mut commands: Commands) {
-    commands.spawn(DirectionalLight {
+    // Bright ambient so surfaces facing away from every light aren't pure black.
-        illuminance: 10_000.0,
+    commands.insert_resource(AmbientLight {
-        shadows_enabled: true,
+        color: Color::srgb(0.9, 0.93, 1.0),
-        ..default()
+        brightness: 900.0,
     });
+    // A multi-directional rig (key + fills + rim + underside) so an orbiting
+    // camera always has some light on the visible side — the single front light
+    // left the backside unreadable.
+    let lights = [
+        (Vec3::new(1.0, 2.0, 1.5), 10_000.0, true),  // key: front-top-right
+        (Vec3::new(-2.0, 1.0, 0.5), 4_500.0, false), // fill: left
+        (Vec3::new(0.5, 0.8, -2.0), 6_000.0, false), // rim: behind
+        (Vec3::new(0.0, -1.5, 0.5), 2_500.0, false), // underside fill
+    ];
+    for (from, lux, shadows) in lights {
+        commands.spawn((
+            DirectionalLight {
+                illuminance: lux,
+                shadows_enabled: shadows,
+                ..default()
+            },
+            Transform::from_translation(from).looking_at(Vec3::ZERO, Vec3::Y),
+        ));
+    }
 }

crates/sylpheed-viewer/src/ui.rs

@@ -10,10 +10,13 @@ use bevy::prelude::*;
 use bevy_egui::{egui, EguiContexts};
 
 use crate::iso_loader::{
-    FileInfo, FileSelected, ImageRgba, IsoState, PakContent, PakView, RequestOpenDir,
+    AudioPreview, FileInfo, FileSelected, ImageRgba, IsoState, ModelPreview, MovieSubtitles,
-    RequestOpenIso, SkyboxPreview, TextPreview, TexturePreview, VideoPreview, IsoLoaderSystemSet,
+    MovieVoice, PakContent, PakView, RequestAudio, RequestOpenDir, RequestOpenIso, RequestSubtitles,
+    RequestVoiceLibrary, SkyboxPreview, TextPreview, TexturePreview, VideoPreview, VoiceLibrary,
+    IsoLoaderSystemSet,
 };
 use crate::ViewerState;
+use sylpheed_formats::SubLang;
 
 pub struct ViewerUiPlugin;
 
@@ -123,6 +126,18 @@ fn render_dir(
     }
 }
 
+/// Bundled event writers for the UI, to keep `draw_viewer_ui` under Bevy's
+/// 16-parameter system limit.
+#[derive(bevy::ecs::system::SystemParam)]
+struct UiEvents<'w> {
+    open_iso: EventWriter<'w, RequestOpenIso>,
+    open_dir: EventWriter<'w, RequestOpenDir>,
+    file_selected: EventWriter<'w, FileSelected>,
+    subtitles: EventWriter<'w, RequestSubtitles>,
+    audio: EventWriter<'w, RequestAudio>,
+    voice_lib: EventWriter<'w, RequestVoiceLibrary>,
+}
+
 fn draw_viewer_ui(
     mut contexts: EguiContexts,
     mut viewer: ResMut<ViewerState>,
@@ -132,11 +147,14 @@ 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>,
+    mut subtitles: ResMut<MovieSubtitles>,
-    mut open_dir_events: EventWriter<RequestOpenDir>,
+    mut movie_voice: ResMut<MovieVoice>,
-    mut file_selected_events: EventWriter<FileSelected>,
+    mut audio: ResMut<AudioPreview>,
+    mut voice_lib: ResMut<VoiceLibrary>,
+    mut events: UiEvents,
 ) {
     let ctx = contexts.ctx_mut();
 
@@ -147,11 +165,11 @@ fn draw_viewer_ui(
                 #[cfg(not(target_arch = "wasm32"))]
                 {
                     if ui.button("Open ISO disc image…").clicked() {
-                        open_iso_events.send_default();
+                        events.open_iso.send_default();
                         ui.close_menu();
                     }
                     if ui.button("Open extracted folder…").clicked() {
-                        open_dir_events.send_default();
+                        events.open_dir.send_default();
                         ui.close_menu();
                     }
                     ui.separator();
@@ -166,6 +184,17 @@ fn draw_viewer_ui(
                 );
             });
 
+            ui.menu_button("View", |ui| {
+                if ui.button("🎙 Voice Lines…").clicked() {
+                    voice_lib.open = true;
+                    if !voice_lib.loaded && !voice_lib.loading {
+                        voice_lib.loading = true;
+                        events.voice_lib.send_default();
+                    }
+                    ui.close_menu();
+                }
+            });
+
             ui.menu_button("Help", |ui| {
                 if ui.button("Controls…").clicked() {
                     // TODO: controls popup
@@ -177,6 +206,101 @@ fn draw_viewer_ui(
         });
     });
 
+    // ── Standalone voice-line browser (floating window) ───────────────────
+    if voice_lib.open {
+        let mut open = true;
+        egui::Window::new("🎙 Voice Lines")
+            .default_width(360.0)
+            .default_height(480.0)
+            .open(&mut open)
+            .show(ctx, |ui| {
+                if voice_lib.loading {
+                    ui.horizontal(|ui| {
+                        ui.spinner();
+                        ui.label("Reading sounds.tbl…");
+                    });
+                    ctx.request_repaint();
+                } else if !voice_lib.loaded {
+                    ui.label("Open a game source first.");
+                } else {
+                    ui.horizontal(|ui| {
+                        ui.label("Filter:");
+                        ui.text_edit_singleline(&mut voice_lib.filter);
+                    });
+                    let f = voice_lib.filter.to_lowercase();
+                    // Group the thousands of entries as directory → speaker so the
+                    // list is navigable (e.g. browse `Voice` by character to find a
+                    // cutscene's radio line). `name` is `<lang>\<dir>\<file>.slb`.
+                    use std::collections::BTreeMap;
+                    let mut groups: BTreeMap<&str, BTreeMap<&str, Vec<&sylpheed_formats::slb::VoiceClip>>> =
+                        BTreeMap::new();
+                    for c in &voice_lib.clips {
+                        if !f.is_empty() && !c.name.to_lowercase().contains(&f) {
+                            continue;
+                        }
+                        let dir = c.name.rsplit('\\').nth(1).unwrap_or("?");
+                        groups
+                            .entry(dir)
+                            .or_default()
+                            .entry(c.speaker.as_str())
+                            .or_default()
+                            .push(c);
+                    }
+                    let shown: usize = groups.values().flat_map(|s| s.values()).map(Vec::len).sum();
+                    ui.label(
+                        egui::RichText::new(format!("{shown} / {} clips", voice_lib.clips.len()))
+                            .weak()
+                            .small(),
+                    );
+                    ui.separator();
+                    let filtering = !f.is_empty();
+                    egui::ScrollArea::vertical().show(ui, |ui| {
+                        for (dir, speakers) in &groups {
+                            let dtotal: usize = speakers.values().map(Vec::len).sum();
+                            egui::CollapsingHeader::new(format!("📁 {dir}  ({dtotal})"))
+                                .id_salt(("vdir", *dir))
+                                .default_open(filtering)
+                                .show(ui, |ui| {
+                                    for (speaker, clips) in speakers {
+                                        egui::CollapsingHeader::new(format!(
+                                            "{speaker}  ({})",
+                                            clips.len()
+                                        ))
+                                        .id_salt(("vspk", *dir, *speaker))
+                                        .default_open(filtering || clips.len() <= 6)
+                                        .show(ui, |ui| {
+                                            for c in clips {
+                                                ui.horizontal(|ui| {
+                                                    if ui
+                                                        .button("▶")
+                                                        .on_hover_text(&c.name)
+                                                        .clicked()
+                                                    {
+                                                        audio.generation =
+                                                            audio.generation.wrapping_add(1);
+                                                        audio.loading = true;
+                                                        audio.active = true;
+                                                        audio.name = c.display.clone();
+                                                        events.audio.send(RequestAudio {
+                                                            clip: c.name.clone(),
+                                                            display: c.display.clone(),
+                                                            movie: None,
+                                                            generation: audio.generation,
+                                                        });
+                                                    }
+                                                    ui.label(&c.display);
+                                                });
+                                            }
+                                        });
+                                    }
+                                });
+                        }
+                    });
+                }
+            });
+        voice_lib.open = open;
+    }
+
     // ── Left panel: file browser ──────────────────────────────────────────
     egui::SidePanel::left("file_browser")
         .resizable(true)
@@ -225,7 +349,7 @@ fn draw_viewer_ui(
                         force_open,
                         selected,
                         loading,
-                        &mut file_selected_events,
+                        &mut events.file_selected,
                         &files,
                     );
                 }
@@ -251,11 +375,43 @@ fn draw_viewer_ui(
     });
 
     // ── Central area ──────────────────────────────────────────────────────
-    if browser.selected.is_some() {
+    if audio.active {
+        // Standalone audio player takes over the central panel.
+        egui::CentralPanel::default().show(ctx, |ui| {
+            ctx.request_repaint();
+            draw_audio_player(ui, &mut audio);
+        });
+    } else 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
                 // background read + decode runs, instead of the previous item.
+                // Keep repainting so the spinner animates even if the app is in a
+                // reactive (idle) winit mode while the worker thread churns.
+                ctx.request_repaint();
                 let name = browser
                     .selected
                     .and_then(|i| browser.files.get(i))
@@ -268,7 +424,43 @@ fn draw_viewer_ui(
                     });
                 });
             } else if video.active {
-                draw_video_player(ui, &mut video);
+                let mut lang_changed = false;
+                let mut play_voice_solo = false;
+                draw_video_player(
+                    ui,
+                    &mut video,
+                    &mut subtitles,
+                    &mut movie_voice,
+                    &mut lang_changed,
+                    &mut play_voice_solo,
+                );
+                if play_voice_solo {
+                    if let Some(movie) = movie_voice.movie.clone() {
+                        video.playing = false; // pause the video; audio takes over
+                        audio.generation = audio.generation.wrapping_add(1);
+                        audio.loading = true;
+                        audio.active = true; // show the panel immediately (spinner)
+                        audio.name = format!("VOICE_{movie}");
+                        events.audio.send(RequestAudio {
+                            clip: String::new(),
+                            display: format!("VOICE_{movie}"),
+                            movie: Some((movie.clone(), movie_voice.lang)),
+                            generation: audio.generation,
+                        });
+                    }
+                }
+                if lang_changed {
+                    if let Some(movie) = subtitles.movie.clone() {
+                        subtitles.generation = subtitles.generation.wrapping_add(1);
+                        subtitles.cues.clear();
+                        subtitles.loading = true;
+                        events.subtitles.send(RequestSubtitles {
+                            movie,
+                            lang: subtitles.lang,
+                            generation: subtitles.generation,
+                        });
+                    }
+                }
             } else if !skybox.faces.is_empty() {
                 draw_skybox_grid(ui, &skybox);
             } else if pak_view.loaded {
@@ -806,7 +998,14 @@ fn fmt_time(secs: f32) -> String {
 /// keyboard shortcuts (space = play/pause, ←/→ = skip 10 s). Playback state
 /// lives in `VideoPreview`; the decode/audio engine reacts to it in
 /// `advance_video_playback`.
-fn draw_video_player(ui: &mut egui::Ui, video: &mut VideoPreview) {
+fn draw_video_player(
+    ui: &mut egui::Ui,
+    video: &mut VideoPreview,
+    subs: &mut MovieSubtitles,
+    voice: &mut MovieVoice,
+    lang_changed: &mut bool,
+    play_voice_solo: &mut bool,
+) {
     // ── Keyboard shortcuts (consumed so focused widgets don't also act). ──
     let (mut toggle_play, mut skip) = (false, 0.0_f32);
     ui.input_mut(|i| {
@@ -838,6 +1037,47 @@ fn draw_video_player(ui: &mut egui::Ui, video: &mut VideoPreview) {
             ui.separator();
             ui.colored_label(egui::Color32::YELLOW, "no audio");
         }
+
+        // Subtitle + voice controls (right-aligned): language, CC, and Voice.
+        ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+            let before = subs.lang;
+            egui::ComboBox::from_id_source("subtitle_lang")
+                .selected_text(subs.lang.label())
+                .show_ui(ui, |ui| {
+                    for lang in SubLang::ALL {
+                        ui.selectable_value(&mut subs.lang, lang, lang.label());
+                    }
+                });
+            if subs.lang != before {
+                *lang_changed = true;
+            }
+            ui.toggle_value(&mut subs.enabled, "CC")
+                .on_hover_text("Show subtitles");
+            if subs.loading {
+                ui.spinner();
+            } else if subs.enabled && subs.cues.is_empty() {
+                ui.weak("(no subtitles)");
+            }
+
+            ui.separator();
+            // Voice-over track: the movie's own stream carries only music+SFX, so
+            // this layers in the localized cutscene voice.
+            ui.toggle_value(&mut voice.enabled, "🗣 Voice")
+                .on_hover_text("Play the cutscene voice track over the movie");
+            if voice.loading {
+                ui.spinner();
+            } else if !voice.available {
+                ui.weak("(no voice)");
+            }
+            if voice.available
+                && ui
+                    .button("🎧")
+                    .on_hover_text("Listen to the voice track on its own")
+                    .clicked()
+            {
+                *play_voice_solo = true;
+            }
+        });
     });
     ui.separator();
 
@@ -849,6 +1089,7 @@ fn draw_video_player(ui: &mut egui::Ui, video: &mut VideoPreview) {
     let scale = (avail.x / w).min(frame_h / h);
     let (disp_w, disp_h) = (w * scale, h * scale);
 
+    let mut frame_rect = None;
     if let Some(egui_id) = video.egui_id {
         ui.allocate_ui(egui::vec2(avail.x, frame_h), |ui| {
             ui.vertical_centered(|ui| {
@@ -860,10 +1101,26 @@ fn draw_video_player(ui: &mut egui::Ui, video: &mut VideoPreview) {
                 if resp.clicked() {
                     video.playing = !video.playing;
                 }
+                frame_rect = Some(resp.rect);
             });
         });
     }
 
+    // Caption overlay: every cue active at the current position, stacked over
+    // the lower third of the frame (overlapping spans show together, newest —
+    // last in start order — lowest).
+    if let Some(rect) = frame_rect {
+        let active = subs.active_cues(video.position);
+        if !active.is_empty() {
+            let joined = active
+                .iter()
+                .map(|c| c.text.as_str())
+                .collect::<Vec<_>>()
+                .join("\n");
+            paint_caption(ui, rect, &joined);
+        }
+    }
+
     // ── Transport bar ──
     ui.horizontal(|ui| {
         if ui.button(if video.playing { "⏸" } else { "▶" }).clicked() {
@@ -907,6 +1164,86 @@ fn draw_video_player(ui: &mut egui::Ui, video: &mut VideoPreview) {
     });
 }
 
+/// Draw a subtitle caption centered along the bottom of `rect`, with a
+/// semi-transparent backing box so it stays legible over any frame.
+fn paint_caption(ui: &egui::Ui, rect: egui::Rect, text: &str) {
+    let painter = ui.painter_at(rect);
+    // Font scales with the frame; clamped so it's readable but not huge.
+    let size = (rect.height() * 0.045).clamp(13.0, 30.0);
+    let font = egui::FontId::proportional(size);
+    let wrap = rect.width() * 0.9;
+    let galley = painter.layout(
+        text.to_string(),
+        font,
+        egui::Color32::WHITE,
+        wrap,
+    );
+    let margin = egui::vec2(10.0, 6.0);
+    let box_size = galley.size() + margin * 2.0;
+    let top_left = egui::pos2(
+        rect.center().x - box_size.x / 2.0,
+        rect.bottom() - box_size.y - rect.height() * 0.04,
+    );
+    let bg = egui::Rect::from_min_size(top_left, box_size);
+    painter.rect_filled(bg, 4.0, egui::Color32::from_black_alpha(160));
+    painter.galley(top_left + margin, galley, egui::Color32::WHITE);
+}
+
+/// Standalone audio player: a transport bar for a decoded voice/sound track with
+/// no video. Playback state lives in `AudioPreview`; `advance_audio_playback`
+/// reacts to it.
+fn draw_audio_player(ui: &mut egui::Ui, audio: &mut AudioPreview) {
+    ui.horizontal(|ui| {
+        ui.heading(&audio.name);
+        ui.separator();
+        if audio.loading {
+            ui.spinner();
+            ui.label("decoding…");
+        } else {
+            ui.label("voice track");
+        }
+        ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
+            if ui.button("✖ Close").clicked() {
+                audio.active = false;
+            }
+        });
+    });
+    ui.separator();
+    ui.add_space(ui.available_height() * 0.4);
+
+    // Big centered play/pause + a speaker glyph, since there's nothing to show.
+    ui.vertical_centered(|ui| {
+        ui.label(egui::RichText::new("🔊").size(64.0));
+    });
+    ui.add_space(12.0);
+
+    // Transport bar.
+    ui.horizontal(|ui| {
+        if ui
+            .button(if audio.playing { "⏸" } else { "▶" })
+            .clicked()
+        {
+            audio.playing = !audio.playing;
+        }
+        ui.label(fmt_time(audio.position));
+        let tl_width = (ui.available_width() - 170.0).max(80.0);
+        ui.spacing_mut().slider_width = tl_width;
+        let mut pos = audio.position;
+        let resp = ui.add(
+            egui::Slider::new(&mut pos, 0.0..=audio.duration.max(0.1)).show_value(false),
+        );
+        if resp.changed() {
+            audio.position = pos;
+            audio.seek_request = Some(pos);
+        }
+        ui.label(fmt_time(audio.duration));
+        ui.separator();
+        ui.label("🔊");
+        ui.spacing_mut().slider_width = 90.0;
+        ui.add(egui::Slider::new(&mut audio.volume, 0.0..=1.0).show_value(false));
+    });
+}
+
 // ── World cubemap (skybox) viewer ─────────────────────────────────────────────
 
 /// Show a world cubemap's 6 faces in a labelled grid (D3D9 order). A true

docs/HANDOFF-movie-voice-subtitles-2026-07-19.md

@@ -0,0 +1,95 @@
+# Handoff — movie subtitles, voice, and the movie manifest (2026-07-19)
+
+Branch `feature/ipfb-idxd-parser`. This commit flushes several sessions of local
+WIP; the **new, finished** work is the movie subtitle + voice + audio pipeline
+and the movie manifest. One item is deliberately **left open** (see §4).
+
+## 1. What's DONE and verified (static RE + tests)
+
+### Movie subtitles (`crates/sylpheed-formats/src/movie_subtitle.rs`)
+- Full movie→track→text chain (see `docs/re/structures/movie-subtitles.md`).
+- **Multi-line caption fix**: a caption stored as several consecutive text tokens
+  sharing one timing (S13A: `"Look at it father"` + `"& beautiful isn't it"` @
+  `01:14.80`) — `parse_track` now joins them with `\n`. Previously all but the
+  last line were dropped. Disc test asserts both lines survive.
+- **Overlap rendering**: `MovieSubtitles::active_cues(t)` returns every cue active
+  at `t`; the viewer stacks them (was: only the first cue shown).
+- Umlauts / Latin-1 accents preserved (`utf16le_tokens`); Japanese label romanized
+  (CJK font still a known gap — see §5).
+
+### Voice decode (`crates/sylpheed-formats/src/slb.rs`)
+- `sound.pak` entries are XACT `.slb` banks wrapping **XMA1** (fmt tag `0x0165`,
+  48 kHz, mono content). `to_xma_riff` rebuilds a decodable RIFF; FFmpeg `xma1`
+  decodes it. Downmix mono via `-af pan=mono|c0=c0`.
+- **Multi-subwave fix**: take the FIRST sub-wave bounded by its declared `data`
+  size — NOT `data..EOF` (which appended later takes = the S10–S16 garble).
+  Verified: S13A→83.78s == video.
+- `list_voice_clips` enumerates the voice library from `sounds.tbl`.
+
+### Movie manifest (`crates/sylpheed-formats/src/movie_manifest.rs`) — the index
+- `tables.pak` entry `ADVERTISE_MOVIE` (IDXD, hash `0x5B983A08`) is the
+  authoritative **movie → subtitle → voice** map. Located by shape (`is_manifest`),
+  parsed by grouping the string pool on `.wmv`.
+- **Authoritative voice binding** replaces the old `VOICE_<movie>` guess:
+  101 movies, 83 with a `VOICE_` token, 18 without. The token is NOT always
+  `VOICE_<movie>` — 5 `hokyu_*` movies bind to `VOICE_D_450..454` in `\etc\`,
+  which a guess would miss. Token's subdir varies (Movie/etc/Voice) → resolve via
+  `sounds.tbl`. Disc test: all 83 resolved entries exist in `sound.pak`.
+
+### Viewer wiring (`iso_loader.rs`, `ui.rs`)
+- `resolve_movie_voice_clip` reads the manifest + `sounds.tbl` (DIRECT bindings
+  only) → `handle_voice_request` (movie player 🗣 toggle) and the 🎧 solo button
+  (`RequestAudio.movie`) use it. Unbound movies correctly post no audio.
+- Standalone **View → 🎙 Voice Lines** browser (filter + ▶ play any `sound.pak`
+  clip) via `VoiceLibrary` + `AudioPreview`.
+
+Tests: 46 formats-lib + 11 viewer-lib + disc (`movie_manifest_disc`,
+`movie_subtitle_disc`, `slb_disc`) all green. Full `cargo test --workspace` green.
+
+## 2. Also bundled in this commit (prior local WIP, not this session's focus)
+- Viewer **async stage/model loading** (off-thread `prepare_xpr`, cancellable) —
+  `iso_loader.rs`/`ui.rs`/`camera.rs`. See memory `reborn-viewer-async-loading`.
+- **Grouped-pool XBG7 / hero-ship** decode refinements in `mesh.rs`, `cli/main.rs`,
+  `mesh_disc.rs`. See memory `reborn-ship-drawlog` / `reborn-saber-texture-investigation`.
+- `tools/analyze_drawlog_wvp.py`, `tools/extract_movie_subtitles.py`.
+
+These build and test green but were not re-verified for behaviour this session.
+
+## 3. USER TO VERIFY IN GUI (couldn't be tested from CLI)
+1. S13A cutscene shows BOTH lines of the "Look at it father" caption.
+2. Story-movie voice (S13A etc.) lines up with the video after the subwave fix.
+3. The 5 directly-bound `hokyu_*` movies play voice; other hokyu are silent (see §4).
+4. View → 🎙 Voice Lines browser lists and plays clips.
+
+## 4. OPEN PROBLEM — unbound-hokyu resupply voice (do NOT re-guess)
+The resupply cutscenes clearly SHARE voice recordings (video-only varies per
+mission), but the correct join key for the **13 unbound `hokyu_*`** movies is
+**unknown**:
+- Manifest DIRECTLY binds only 5: `LS_s02A→450, LS_s09A→451, DS_s02A→452,
+  LS_s02H→453, DS_s07H→454` (grouping verified against the raw layout).
+- Keying unbound movies by subtitle **demo id** (600→450…604→454, so
+  `hokyu_DS_s13A` demo602→`VOICE_D_452`) was implemented, tested against the
+  running game by the user, and is **WRONG** (wrong line). It has been **reverted**
+  — unbound hokyu now stay unvoiced rather than play wrong audio.
+- Red flag: only `VOICE_D_450..454` exist; decoded durations `450=2.8s 451=1.6s
+  452=2.2s` but `453=0.14s 454=0.43s` — far too short for the spoken line, so
+  these `.slb` are likely **multi-subwave / not cleanly sliced** (same class as
+  the deferred B/C banks below).
+
+**Next-box options:** (a) get the real join key from mission-event data (mission
+scripts/IDXD tables that trigger resupply), or (b) build a proper multi-subwave
+`.slb` decoder and verify audio by ear. Needs a concrete calibration clue from the
+user (e.g. "s13A should sound like <movie X>" or the actual spoken words).
+
+## 5. Other deferred items (unchanged)
+- ~49 "layout-B/C" movie/voice banks (most RT + S07B/S11A/S12B/S13B) + some
+  individual clips need an XMA1 packet/seek-table reassembler.
+- CJK/Japanese text rendering in the viewer (needs a CJK TTF via egui FontDefinitions).
+- Texture colours (channel order / sRGB) still on the dynamic-RE backlog.
+
+## 6. Resume checklist (next box)
+1. `git pull` on `feature/ipfb-idxd-parser`; set `SYLPHEED_DISC=<extract root>`
+   (had `.../sylph_extract`).
+2. Build guardrail: `CARGO_BUILD_JOBS=4` always (15 GB box; full `-j` OOM'd before).
+3. `cargo test --workspace` (with `SYLPHEED_DISC`) should be green.
+4. Pick up §4 (unbound-hokyu voice) — the only open thread from this session.

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/movie-subtitles.md

@@ -0,0 +1,174 @@
+# Movie subtitles & the movie ↔ mission ↔ text chain
+
+Reverse-engineered 2026-07-19 (static, from the extracted disc). The full chain
+that links a cutscene movie to its on-screen subtitle text is now closed.
+
+## Files involved
+
+- `dat/movie/*.wmv` — the cutscene videos. **Named by mission** (see below).
+- `dat/movie/<lang>.pak` + `<lang>.p00` — per-language **subtitle timing tracks**
+  (`eng`, `jpn`, `deu`, `fra`, `esp`, `ita`) plus the caption font.
+- `dat/GP_MAIN_GAME_<L>.pak` + `.p00` — per-language **caption TEXT**
+  (`E`=Eng, `J`=Jpn, `D`=Deu, `F`=Fra, `I`=Ita, `S`=Esp).
+- `dat/tables.pak` entry `ADVERTISE_MOVIE` (hash `0x5B983A08`) — the master
+  **movie manifest**: all 101 `.wmv` names in mission-progression order, each
+  bound to its subtitle track and (optionally) its **voice bank** (see below).
+
+## Movie filename → mission
+
+Purely from the filename:
+
+| Pattern | Meaning |
+|---------|---------|
+| `S<NN><P>.wmv` | Stage `NN` **story** cutscene, part `P` (A/B/C…) — e.g. `S02C` = stage 2, 3rd story scene |
+| `RT<NN><P>.wmv` | Stage `NN` **radio / briefing** transmission, part `P` (`_1`/`_2` = split clips) |
+| `hokyu_<LS\|DS>_s<NN><P>.wmv` | Stage `NN` **resupply** scene (`hokyu` = 補給). `LS`/`DS` = the two resupply-ship variants |
+| `ADV.wmv` | Intro / title movie |
+
+97 movies total: 27 story, 50 radio, 19 resupply, 1 intro. (Some referenced
+stages — s24, s27, RT16 — exist as keys but the .wmv isn't in this extract.)
+
+## Subtitle timing track — `<lang>.pak`
+
+IPFB archive (`IPFB`, BE-u32 count, 16-byte header; TOC of
+`[name_hash u32][offset u32][size u32]` triples, sorted by hash, into `.p00`).
+
+- **Track key = `name_hash("subtitle_<movie_basename>.tbl")`** (the hash
+  lowercases internally, so basename case is irrelevant). This is the movie→track
+  link. Verified: `subtitle_S00A.tbl` → `0x6F2D9663`, `subtitle_hokyu_DS_s02A.tbl`
+  → `0x3662B1F8`, `subtitle_RT01C_1.tbl` → `0x756F69FB`.
+- The archive also holds **RATC** pre-rendered title-card / number textures
+  (`pwterop_s01a1.t32`, `pwrt_rt01_str.t32`, `pwnum0-9.t32`) + one TrueType font.
+- **Each track data block is `Z1`+zlib**: bytes `5A 31` ("Z1"), a small header,
+  then a raw zlib stream (`78 DA`/`78 9C`). `zlib.decompress(blob[blob.find(b"\x78\xda"):])`.
+- Decompressed = an **IXUD** container. Payload (UTF-16LE) is the timing sheet:
+  `SUBTITLE  MSG_DEMO_<demo>  <mm:ss.cc>  MSG_DEMO_<demo>  <mm:ss.cc> …`.
+  So the track says *which* demo-message shows *when*, not the text itself.
+
+## Caption text — `GP_MAIN_GAME_<L>.pak`
+
+Same IPFB+`.p00`. Among its ~1119 entries, **32 blocks are `Z1`+zlib → IXUD**
+string containers holding the movie caption text. Layout: `IXUD`, u32 version,
+hash@0x08, count@0x14, then `(recordhash,offset,len)` triples, then a UTF-16LE
+string region where **each line is stored as `text` immediately followed by its
+key** `MSG_DEMO_<demo>_<page>_<line>` (captions wrap across `_00`,`_01`, …).
+537 English lines recovered. Entries 1 & 19 are the IDXD *schema* records
+(`ID`, `PageCount`, `Character`=speaker e.g. `TCAFSUPPLY`, `Face`, line refs) —
+no text, just structure.
+
+## Movie → voice track: the manifest binding (`ADVERTISE_MOVIE`)
+
+The `ADVERTISE_MOVIE` manifest is **also the authoritative movie→voice index**.
+Its string pool emits, per movie, a run led by `<movie>.wmv` optionally followed
+by `<pak>+….prt` (overlay art), `<pak>+SUBTITLE_<movie>.tbl`, and a bare
+`VOICE_<token>`. Grouping the pool on `.wmv` (records are emitted in order)
+recovers `movie → Option<voice_token>` without decoding the IDXD record binary
+(`crate::movie_manifest`).
+
+The voice token is **not** always `VOICE_<movie>`, so the manifest is required —
+guessing both misses real bindings and invents tracks for silent movies:
+
+- **83 / 101 movies have a voice token.** Story/radio movies use `VOICE_<movie>`
+  in `<lang>\Movie\`.
+- **5 `hokyu_*` resupply movies bind to in-mission radio clips** — e.g.
+  `hokyu_LS_s02A → VOICE_D_450`, which lives in `<lang>\etc\`, *not* `Movie`.
+  A `VOICE_<movie>` guess would never find these.
+- **18 movies have no direct voice token** = 4 boot logos + 1 HD test pattern +
+  **13 `hokyu_*` movies** (incl. `hokyu_DS_s13A`). Only the manifest's **direct**
+  bindings are trusted for playback.
+
+### Shared resupply voice — UNRESOLVED for unbound movies
+
+The manifest directly binds only **5** resupply movies, each to a shared
+`VOICE_D_45x` clip in `<lang>\etc\`:
+
+| bound movie | subtitle demo | clip |
+|-------------|---------------|------|
+| `hokyu_LS_s02A` | 600 | `VOICE_D_450` |
+| `hokyu_LS_s09A` | 601 | `VOICE_D_451` |
+| `hokyu_DS_s02A` | 602 | `VOICE_D_452` |
+| `hokyu_LS_s02H` | 603 | `VOICE_D_453` |
+| `hokyu_DS_s07H` | 604 | `VOICE_D_454` |
+
+The resupply cutscenes clearly **share** voice recordings (only the video varies
+per mission), so the 13 unbound `hokyu_*` movies must reuse one of these — but
+the **correct join key is not yet known**:
+
+- Keying by subtitle **demo id** (so `hokyu_DS_s13A`, demo 602 → `VOICE_D_452`)
+  was tried and is **WRONG** — it plays the wrong recording in-game. Do not use.
+- Only `VOICE_D_450..454` exist (no 44x/45x neighbours). Decoded durations are
+  suspicious — `450`=2.8s, `451`=1.6s, `452`=2.2s, but `453`=**0.14s**,
+  `454`=**0.43s** — far too short for the spoken line, so these `.slb` banks are
+  likely **multi-subwave / not cleanly sliced** by the current extractor (same
+  class as the deferred B/C banks).
+
+⇒ The unbound-hokyu voice mapping is **open** (needs either the real join key
+from mission data, or a proper multi-subwave `.slb` decode + audio verification).
+`hokyu_LS_s24A`/`s27A` have no subtitle track at all (stages absent from this
+extract).
+
+The token's `sound.pak` **subdirectory is not fixed** (`Movie` / `etc` / `Voice`),
+so resolve it via `sounds.tbl` (which lists the full `<lang>\…\<token>.slb` path)
+rather than assuming a directory. `movie_manifest::resolve_voice_entry` does the
+full chain. Verified: all 83 resolved entries exist in `sound.pak`.
+
+## Caption packing quirks (parser must handle)
+
+- **Multi-line captions are split into consecutive text tokens** that share one
+  trailing timing, e.g. S13A stores `"Look at it father"` + `"& beautiful isn't
+  it"` before `01:14.80-01:17.60`. Accumulate every text token since the last
+  timing and join with `\n`; pairing strictly 1:1 silently drops all but the last
+  line.
+- **Overlapping spans**: some tracks show two captions at once (an open-ended
+  radio line still up when the next range line starts). The viewer stacks every
+  cue active at `t` (`MovieSubtitles::active_cues`) instead of showing only the
+  first.
+
+## The full join
+
+```
+tables.pak / ADVERTISE_MOVIE        → list of movies (mission order)
+<movie>            ─ nh("subtitle_<movie>.tbl") ─→  <lang>.pak track
+track (IXUD)       → [ (MSG_DEMO_<d>, timecode), … ]
+MSG_DEMO_<d>       → GP_MAIN_GAME_<L>  → "the localized caption line(s)"
+```
+
+## Coverage
+
+- 92 / 97 movies have a subtitle track.
+- **66 movies carry timed `MSG_DEMO` captions** — the **radio (`RT*`)** and
+  **resupply (`hokyu_*`)** movies. These fully decode to timed text.
+- The **27 story (`S*`) movies have a track but 0 timed captions** — their text
+  is delivered as the **pre-rendered title-card textures** (`pwterop_*`, burned
+  styling), not MSG_DEMO lines.
+
+## Worked examples (English)
+
+```
+RT01C_1.wmv  (Stage 1 radio, part C):
+  00:00.50  [14] We did it! Okay, all pilots follow my lead!
+  00:06.80  [15] Rhino Leader to ACROPOLIS. We made it through and we're coming
+                 home. Roger. It's good to see you're all safe.
+  00:19.30  [17] Yeah, but Brandon ...  Damn. There's only seven of us. …
+
+hokyu_DS_s02A.wmv  (Stage 2 resupply):
+  00:00.00  [602] Resupply complete. You are cleared for take-off!
+```
+
+## Reusable extractor
+
+`tools/extract_movie_subtitles.py` — parses `<lang>.pak`, resolves each movie's
+track, cross-references `GP_MAIN_GAME_<L>` text, and prints per-movie timed
+transcripts + the movie→mission table.
+
+## In-mission dialogue (future work)
+
+`GP_MAIN_GAME_<L>.pak` is the **global** message store, not just movie captions:
+its `MSG_DEMO_*` table also holds the in-mission radio/dialogue lines (same demo
+id space). So the *text* of gameplay dialogue is already decodable with
+[`crate::movie_subtitle::build_demo_text`]. What's missing is the **trigger** —
+which demo id fires at which mission event — and that lives in the **mission
+data** (mission scripts / `GP_MAIN_GAME` IDXD tables), not in the text pack. When
+reversing mission data, look for demo-id references there to bind dialogue to
+events; the IDXD "Message" schema records also carry `Character` (speaker) and
+`Face` (portrait) per line.

docs/re/structures/sound-slb.md

@@ -0,0 +1,62 @@
+# Sound bank audio — `sound.pak` / `.slb` / XMA1
+
+Reverse-engineered 2026-07-19 (static). `dat/sound.pak` (+ `sound.p00..p04`,
+~1.08 GB) is the game's entire audio bank: **9519** IPFB entries, each an XACT
+`.slb` sound bank wrapping **XMA1** audio.
+
+## Names
+
+Entry keys are `name_hash(<string>)` of filenames stored verbatim in the IDXD
+string pools of `tables.pak` → `eng\sounds.tbl` (entry 39) and `jpn\sounds.tbl`
+(entry 45). All 9519 recovered. Families:
+
+- `<lang>\Voice\VOICE_<SPK>_<NNN>.slb` — per-line character voice (SPK = ADAN,
+  TCAF, RHIN, ADPL, BIRD, ACRO, ZZZZ…). Only `eng`/`jpn` voice exists.
+- `<lang>\etc\VOICE_{A,B,C,D}_<NNN>.slb`, `<lang>\Briefing\BR<n>_<m>.slb`.
+- **`<lang>\Movie\VOICE_<movie>.slb`** — the continuous voice track for cutscene
+  `<movie>.wmv` (e.g. `eng\Movie\VOICE_RT07A.slb`). Played from the video start.
+- `BGM_###.slb` (32) — music. `Static.slb`, `JNGL_###` — SFX/ambience.
+
+Only English and Japanese have voice; subtitles cover more languages, voice does
+not. See [`crate::slb::movie_voice_name`].
+
+## Codec
+
+XMA1: RIFF `fmt ` tag **0x0165**, a 32-byte `XMAWAVEFORMAT`, **48000 Hz**. The
+content is **mono** — some clips carry it in the **left channel only** (right is
+silence), others are **dual-mono** (L = R). Downmix to mono by taking the left
+channel (it always holds the full signal).
+
+`XMAWAVEFORMAT` (one stream, 32-byte `fmt `): SampleRate @ fmt+16 (u32 LE),
+Channels @ fmt+29 (u8), ChannelMask @ fmt+30 (u16). Movie voices report 2
+channels / mask 0x2 despite the mono content.
+
+## `.slb` layouts → decodable RIFF
+
+Two on-disc layouts (`sound.pak` movie voices split ≈ 63 RIFF / 15 headerless):
+
+- **RIFF present** — a `RIFF/WAVE` sits inside the bank; its 32-byte `fmt ` is
+  the real header. The XMA packets are **everything after that RIFF's `data`
+  chunk header** (`slb[data_pos+8..]`); the declared `data` size is unreliable
+  (sometimes half the real length), so take to end and clamp to the media length
+  downstream. Some banks have trailing bank data past the real audio — always
+  beyond the video length, so clamping to the video duration drops it.
+- **Headerless** — no `RIFF` anywhere; a fixed **1392-byte** header precedes raw
+  XMA1 packets (48 kHz, 2 channels). Data = `slb[1392..]`.
+
+[`crate::slb::to_xma_riff`] rebuilds a standalone XMA1 `RIFF/WAVE` for either
+layout, ready for FFmpeg's `xma1` decoder.
+
+## Decode (FFmpeg)
+
+`ffmpeg -i rebuilt.riff -af "pan=mono|c0=c0" -t <media_len> out.wav` — the RIFF
+is fed to the `xma1` decoder, downmixed to the left channel, clamped to length.
+Verified: `VOICE_S00A` → 93.9 s == `S00A.wmv` 93.87 s; `VOICE_ADV` → 137.3 s ==
+137.7 s; `VOICE_RT07A` → 49 s ≈ 50 s.
+
+## Reading one entry cheaply
+
+`sound.pak` is ~1 GB. Use [`crate::PakArchive::parse_toc`] on the small `.pak`
+index, binary-search the name-hash, then read just `[offset, offset+size)` from
+the `.pNN` segments (the viewer's `read_segment_range` seeks into the right
+segment) instead of loading the whole archive.

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)
+### ✅ CONFIRMED — on-disk order is **ARGB** (alpha first); `[a,r,g,b]→[r,g,b,a]` is correct
-Our decoders currently assume on-disk **A8R8G8B8** and reorder `[a,r,g,b] → [r,g,b,a]`. Canary
+Settled by two independent lines that agree, without an emulator run:
-proves the *host* target is R8G8B8A8 identity-swizzled, but the *guest→host* byte order also depends
+1. **Measured on the disc data** — across 789 real T8aD textures (GP_HANGAR_ARSENAL, GP_MAIN_GAME_E,
-on the texture's runtime **endian** field (e.g. `8in32`), which is set by the game at fetch time and
+   DefTables), the byte position that is most-often exactly `0xFF` (the opaque-alpha signature of UI
-**cannot be read statically** from the format table. So whether our reorder matches is unverified.
+   art) is **byte 0 in 767/789 (97%)** — tf1 385/0, tf2 382/12. So texel byte 0 = alpha ⇒ on-disk
-→ *What would confirm/refute it:* one **measured** decoded texel — either capture Canary's
+   layout is **A,R,G,B**. (An earlier bimodality test tied byte0/byte3 only because colour channels
-framebuffer over a texture with distinct R≠G≠B pixels of known colour, or patch Canary's already-
+   are also `0x00`-heavy from black backgrounds; the `==0xFF` discriminator isolates alpha cleanly.)
-present-but-unwired `TextureDump` (`src/xenia/gpu/texture_dump.cc`) to dump the decoded texture and
+2. **Cross-check vs our Canary-validated renderer** — `xenia-rs`'s `decode_k8888_tiled` (the path
-diff bytes against our decode. **Measure, don't guess** — a symmetric colour (white logo) can't
+   behind the user-confirmed M1 splash / M2 video colours) nets memory-`ARGB` → endian-swap →
-disambiguate channel order.
+   `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
+- `2026-07-11` — static read of Canary `vulkan_texture_cache.cc` host-format table; no `R8G8B8A8_SRGB`
-  table + confirmed no `R8G8B8A8_SRGB` entry exists. → sRGB/UNORM claim `CONFIRMED`; channel-order
+  entry exists → sRGB/UNORM claim `CONFIRMED`.
-  claim remains `HYPOTHESIS` pending a measured texel.
+- `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 +1,254 @@
+# XBG7 — mesh geometry (inside XPR2 model containers)
+
+- **Confidence:** 🟡 `PROBABLE` for the single-stream layout (below); ❔ `HYPOTHESIS` / undecoded
+  for the complex multi-stream body layout.
+- **Parser in:** `sylpheed-formats/src/mesh.rs` (`Xbg7Model::from_xpr2`), tests
+  `tests/mesh_disc.rs`. Container parsing reused from `src/texture.rs` (`Xpr2Header` /
+  `Xpr2ResourceEntry`).
+- **Applies to:** ship / weapon / prop models in `hidden/resource3d/*.xpr` (166 files).
+- **Method:** clean-room — static hex inspection of the retail disc + geometric validation of the
+  recovered triangles (non-degenerate area, indices in range, bbox matches the descriptor's stored
+  size). No game code decompiled or copied.
+
+## Where XBG7 lives
+
+Models are ordinary **`XPR2`** containers (see the XPR2 texture doc / `texture.rs`). The 16-byte
+resource-directory entries (from file offset `0x10`) carry `TX2D` texture resources **and** one or
+more `XBG7` geometry resources:
+
+```
+entry = [ tag:4 ][ data_offset:u32 ][ descriptor_size:u32 ][ name_offset:u32 ]   (big-endian)
+```
+
+Offsets are relative to the directory base `0x10`. The `XBG7` *descriptor* (at `data_offset+0x10`,
+`descriptor_size` bytes) is a **scene / material / node graph** — it holds node names
+(`rou_f001_mnt1_root`, `Light`, …), a bounding value (`0x41F00000` = 30.0 ≈ the ship's ~30-unit
+length), material names matching the `TX2D` channels (`_col` albedo, `_spc` specular, `_gls` gloss,
+`_lum` luminance), and per-sub-mesh records. The **vertex / index buffers** live in the container's
+shared **data section** (from `header_size`).
+
+## Sub-mesh records (in the descriptor)
+
+Read in file order by a sliding 4-byte scan; each is a big-endian tuple:
+
+```
+[ vtx_count:u32 ][ 0:u32 ][ idx_count:u32 ][ tail:u32 ]
+   3..=65535        ==0     mult. of 3        1..=64
+```
+
+(For `rou_f001_wep_00`: `vtx_count=215`, `idx_count=1092` — matches the recovered geometry exactly.)
+
+## The single-stream data layout — 🟡 PROBABLE (decoded, GPU-cross-checked)
+
+For 36 of the 166 models (weapons, simple props) the data section is a straight sequence of
+sub-meshes, carved from `header_size` in record order:
+
+```
+per sub-mesh block:
+  [ 12-byte header  (contents undecoded) ]
+  [ index buffer  : idx_count × u16 BE ]        triangle list (prim=4, GPU-confirmed)
+  [ vertex buffer : vtx_count × stride bytes ]  ← declaration-driven
+  (pad to 16 bytes → next sub-mesh)
+```
+
+The 12-byte header precedes the **index** buffer (the same block shape as stage
+resources — see below); the vertex buffer follows the indices with no further
+gap. (Earlier this was mis-modelled as `[index][12-byte gap][vertex]`, which put
+the vertex buffer at the identical offset but read the index buffer 12 bytes too
+early — turning the 12 header bytes into 6 junk indices = **2 leading degenerate
+triangles** (a stray-triangle artifact) and dropping the last 6 real indices.
+Skipping the header fixes the triangle list with no change to vertex coverage.)
+
+**Vertex declaration.** The layout is **not fixed-stride**. The descriptor holds a declaration table
+(right after the `(index_bytes, index_count)` marker) of `{offset:u32, format-code:u32, usage<<16:u32}`
+big-endian triples, terminated by `offset == 0x00FF0000` / `code == 0xFFFFFFFF`:
+
+| usage | element  | format code | format | size |
+|-------|----------|-------------|--------|------|
+| 0x00  | POSITION | `0x2A23B9`   | f32×3  | 12 B |
+| 0x03  | NORMAL   | `0x1A2360`   | f16×4 (use xyz) | 8 B |
+| 0x05  | TEXCOORD | `0x2C235F`   | f16×2 (u,v) | 4 B |
+
+Stride = max element extent. Models **omit elements** → variable stride (20 = pos+normal, no UV;
+24 = pos+normal+uv). Each element is read in **naive big-endian component order** (the raw file
+bytes; see the endianness note below). Assuming a fixed stride-24 was why the old decoder mis-aligned
+and declined the pos+normal-only models.
+
+**Alignment pinned by the normals.** The vertex buffer starts at `index_end + 12` (a fixed 12-byte
+header — NOT `align16`, which lands 4 bytes early on most models). The correct offset is the unique
+one where recovered normals are exactly unit-length.
+
+**Safety gate:** every index is validated `< vtx_count`, and when the declaration has a normal
+element the mean recovered `|normal|` must be ≈1 (`[0.5, 2.0]`). A model failing either is rejected
+(`MeshError::UnsupportedLayout`) rather than emitting garbage.
+
+### Endianness — file bytes are naive-BE, `k8in32` is a red herring ✅
+
+The Canary GPU capture reports every vertex stream with fetch `endian = 2` (`k8in32`, each 32-bit
+word byte-reversed). This describes the **guest-memory** copy the GPU fetches — **not** the `.xpr`
+file bytes. Reading the file with a `k8in32` transform breaks the normals (mean `|normal|` → 1.33);
+the plain per-element big-endian read yields exactly unit normals. So the game **rearranges** the
+vertex data between the on-disc `.xpr` and the uploaded buffer; the decoder reads the file directly
+and must use naive BE.
+
+## Stage containers — multi-resource, grouped pools ✅ (content-anchored)
+
+`hidden/resource3d/Stage_S*.xpr` are not single models but **collections of
+enemy / prop sub-models** — up to ~400 `XBG7` resources each (e.g. `Stage_S07` =
+378). Their data layout differs from the weapon files:
+
+- Each resource is a block `[12-byte header][index buffer][vertex buffer]`.
+  Unlike weapons there is **no 12-byte gap between index and vertex** — the
+  vertex buffer directly follows the indices.
+- The **index count** is the descriptor's `(index_bytes, index_count)` marker
+  (the *total* for the resource — a resource may have several sub-meshes summing
+  to it), **not** the first sub-mesh record. The **vertex count** is a `u32`
+  stored **32 bytes before** the marker.
+- The blocks are **scattered through the data section, interleaved with the
+  container's texture data**, in an allocation order that is **not** directory
+  order and is **not stored** in any descriptor field we could find (the
+  descriptor holds sizes — `rel 160 ≈ index_bytes+3`, `rel 164 =
+  0x1000_0000 | (vertex_bytes+2)` — but no data offset). Reconstructing that
+  allocation order is unsolved.
+
+Because the offset is not stored, each resource's block is located by
+**content**: parser `Xbg7Model::stage_models` does one `O(file)` pass per
+distinct stride to find every **vertex-buffer start** (an offset whose NORMAL —
+`f16×4` at vertex `+12` — is unit length while the *previous* stride slot's is
+not, i.e. a run boundary; ~one candidate per block, not millions), then pins each
+resource to the unique candidate where the `index_count` indices ending just
+before it are all `< vertex_count`, reference nearly all vertices, and yield
+non-degenerate triangles with real extent. This is fast (≤ ~2 s on a 70 MB
+stage) and unambiguous (no two resources collide). Blocks that fail validation —
+the few quantized hero bodies — are **skipped**, never emitted as garbage.
+
+**Coverage:** 5662 sub-models decode across the 22 stage files (e.g. `Stage_S07`
+366/378, `Stage_S10` 7/9 — including the main enemy bodies `e003`/`e005`, their
+LODs, weapons, and props). The viewer (`spawn_stage_models`) lays the decoded
+sub-models out as a side-by-side "cast sheet", skipping the few huge skybox-plane
+resources (300 k-unit quads). Cross-checked: `e003` = 2383 v / 1436 t bbox
+23.5×9.5×32.5; `e005` = 2566 v / 1507 t; both 0 degenerate.
+
+## Not yet decoded — ❔ the complex body layout
+
+The hero-ship *body* meshes (`DeltaSaber_A/_T/_W.xpr` resource `f004`, and ~100 other models) still
+decline. Two open sub-problems: (1) **multi-sub-mesh models** whose *first* sub-mesh decodes but a
+later one's inter-mesh offset isn't yet handled (the `align16` advance is a guess) — these are
+declined whole; (2) the big body meshes, where the data section does not start with an index buffer
+and the geometry sits at descriptor-addressed offsets. NB the GPU capture showed **every** rendered
+mesh is *single-stream* (just wider strides, e.g. 44 bytes = pos + f32×3 + colour + 2×f16×4), so the
+body is likely single-stream-with-a-richer-declaration rather than the "separate streams" first
+guessed — it was simply not rendered in the captured session (menu only). A capture taken *in a
+mission* (where `DeltaSaber` renders) would hand over its exact declaration directly. `DeltaSaber_A`'s
+5 XBG7 blocks are `f004` (body) + `_rou_f004_mnv01_L/_R`, `_mnv02`, `_turn180` (maneuver / pose).
+
+## Evidence log
+
+- 2026-07-12 — `rou_f001_wep_00.xpr`: XPR2 dir = 1×XBG7 (`f001_wep_00`) + 3×TX2D
+  (`_col/_gls/_spc`). Data section starts with a u16-BE index run (max 214), then stride-24
+  vertices. Descriptor record `[215,0,1092,4]` at desc `+0x2B0`; index-buffer byte size `0x888`
+  (=2184=1092×2) at desc `+0x1A0`. Recovered mesh = 215 v / 364 t, 362 non-degenerate, median tri
+  area 0.015 — coherent. → single-stream layout **PROBABLE**.
+- 2026-07-12 — descriptor-driven sequential carving over all 166 models: **42 carve** under the
+  index-only check. Real 3D extent confirmed on `rou_f001_wep_04` (bbox 3.7×1.2×3.5).
+- 2026-07-12 (refine) — attribute ranges differed per model (`wep_00` UV≈attr0/2, `wep_04`
+  attr4/5, `wep_03` attr1≈±60000 = garbage) → **refuted the fixed "6 half attrs, UV=attr0/2"
+  reading.** Found the descriptor **vertex declaration** (offsets 0x0C/0x14, usages 0x03 NORMAL /
+  0x05 TEXCOORD). Solved the vertex-base offset with a **unit-normal validator**: `index_end + 12`
+  gives median `|normal| = 1.000` on every weapon model (`wep_00/02/03/04`), vs `align16` landing 4
+  bytes early. UVs then land in `[0,1]`. Adding the normal gate: **25 models decode clean +
+  normal-valid** (the rest — incl. `Stage_S*` degenerate blobs — correctly declined).
+- 2026-07-12 (DYNAMIC) — added a cvar-gated draw logger to Canary
+  (`command_processor.cc::LogDrawForRE`, cvar `log_draws`), captured the Ready Room / Briefings.
+  **GPU ground truth confirmed the static layout exactly**: primitive `prim=4` = **triangle LIST**
+  (settles the list-vs-strip question), and a stream with `f32x3 @offset0` + `f16x4 @3dw` +
+  `f16x2 @5dw`, stride 6 dwords = 24 bytes — matching `POSITION@0, NORMAL@0x0C, TEXCOORD@0x14`.
+  Revealed **stride varies** (24, 20, 28, 44 …) and that all streams are **single-stream** →
+  parsed the declaration for variable stride: coverage **25 → 36** (e.g. `wep_05` is pos+normal,
+  stride 20, no UV — previously mis-aligned). Also confirmed the endianness note above: fetch
+  `endian=2` (k8in32) is the *guest* copy; file stays naive-BE. `Stage_S*` now decode (stride 20).
+- 2026-07-12 (STAGE) — `Stage_S*.xpr` decoded as multi-resource containers. Found each geometry
+  block is `[12B hdr][index buffer][vertex buffer]`, index count = descriptor marker (total, e.g.
+  `e003` = 4308 spanning 2 sub-meshes), vertex count = `u32` at marker−32 (`e003` = 2383 → verified
+  by max-index 2382 and 0 degenerate tris, bbox 23.5×9.5×32.5). Blocks are scattered among texture
+  data with **no stored offset** (descriptor rel 160 = idx_bytes+3, rel 164 = `0x1000_0000 |
+  (vtx_bytes+2)` are sizes, not offsets; block starts e.g. `e003`@0x10230, `e003_l`@0x5d000,
+  `e005_l`@0x9b000 are not directory-ordered). Solved by **content anchoring**: a single per-stride
+  pass finds vertex-run starts (unit NORMAL at +12 whose previous slot isn't), then match each
+  resource by strict index+triangle validation. **5662 sub-models decode across 22 stages** (S07
+  366/378), incl. the previously-declined main bodies `e005` (2566 v) and weapons — ≤2 s on 70 MB.
+  Parser `Xbg7Model::stage_models`, tests `stage_models_{decode,sweep,quality_audit}`.
+- 2026-07-17 — **triangle-LIST re-confirmed; a strip interlude refuted; winding-consistency gate
+  added.** A 2026-07 change had briefly re-read the index buffers as triangle *strips* (to "fill
+  holes"). Refuted objectively with a new `XVERIFY` diagnostic that compares both readings by
+  **stored-normal agreement** (each triangle's cross-product face normal vs the sum of its vertices'
+  stored normals): the LIST reading gives agreement **1.000** on every clean weapon (`wep_00/03/04/19`
+  = only possible with correct topology + winding), the STRIP reading **~0.49** (random). The strip
+  reading also over-generated ~2.5× the triangles (wep_00: 938 vs 364) — a hole-filling garbage soup.
+  Reverted to LIST in both paths (`from_xpr2`, `read_pool_mesh`), matching the `prim=4` GPU capture.
+  Added an objective **winding-consistency gate** `max(na, 1-na)`: a correct carve is internally
+  consistent (agreement ≈1.0, or ≈0.0 for inverted-but-consistent winding — a real single-sided
+  mesh), a mis-carve scatters to the ≈0.5 middle. `from_xpr2` declines sub-meshes below 0.90 (e.g.
+  `wep_23` na=0.398 → declined instead of a spike-mess); the single-model **content-anchor fallback**
+  gates at 0.85; the large multi-resource **stage** path stays ungated (its enemy meshes span a
+  continuous 0.5–1.0 consistency range — a hard gate there dropped ~48/314 legit S07 blocks).
+  **Routing fixed:** `decode_models` (CLI) and the viewer now route by `count_xbg7` (1 → validated
+  records-based list decode, fallback to strict-gated anchor; >1 → stage anchor) instead of the old
+  "whichever decoder yields more verts" rule — that rule let stage content-anchoring win on
+  single-model weapon files and fabricate **phantom** blocks (a `wep_00` clone appearing inside
+  `wep_19`), duplicates, and spike-mess anchors. Weapons now: 33 clean-decode / 26 declined (declined
+  = genuinely multi-stream or un-carvable, shown as nothing rather than garbage); stage coverage
+  unchanged (S07 314). `expand_triangle_strip` retained as an `XVERIFY`-only diagnostic.
+- 2026-07-12 — `DeltaSaber_A.xpr` body: data does **not** begin with indices; plain-`f32×3` runs
+  with ship-scale extent (span ≈27–34, matching bbox 30.0) found only at high offsets
+  (`data+0x28634C`, …) → multi-stream, **undecoded**.
+- 2026-07-18 — **GROUPED-POOL layout cracked → the hero ship (Delta Saber) fully decodes.** The
+  detailed models (`DeltaSaber_*.xpr` + ~100 others) were declined for **location**, not format —
+  their vertex format is the standard stride-24 triangle list. A resource's *several* sub-meshes
+  don't interleave `[idx][vtx]` per block; they share **two grouped pools**: an **index pool**
+  (buffers concatenated in descriptor-marker order, each **4-byte aligned**) followed by a **vertex
+  pool** (each sub-pool `vtx_count × stride`, same order), with the index pool ending **exactly**
+  where the vertex pool begins. So the whole resource pivots on one unknown, the first vertex-pool
+  start `vb0` (= index-pool end, found by the unit-normal vertex-run scan); everything else is
+  derived: `ib0 = vb0 − span`, `ib[i] = align4(ib[i-1] + idx_count[i-1]·2)`,
+  `vb[i] = vb[i-1] + vtx_count[i-1]·stride`. Reversed statically from `DeltaSaber_T.xpr` and
+  **cross-checked against a Canary GPU draw-log capture** (mission ship = `DeltaSaber_T.xpr`, found
+  via the `--log_file_io` kernel hook): `f001` = body (idx@`data+0xC` = 0x5500C, vtx@0x61ACC, 10891 v
+  / 8187 t) + **7 detail parts** (fins/cockpit/wingtips, markers at descriptor 0x3BEC…0x58FC) =
+  **8650 tris**, and **every sub-mesh decodes at 0 degenerate / full coverage / winding-agreement
+  1.000**. This is the layout the per-block adjacency anchor (`ib = vb − idx_bytes`) rendered as a
+  **spiky phantom** (it read 24561 indices starting 2782 B too late, agree 0.64, 1277 degenerate).
+  Insight: a single index marker reduces the grouped model to `index_end = vb0`, i.e. the existing
+  adjacency `ib = vb − idx_bytes` — so grouped **generalises** the single-block anchor (n=1 is
+  identical). Implemented as `anchor_grouped_meshes` (mesh.rs): `anchor_models` routes resources with
+  >1 index marker to it (validated per sub-mesh; on failure falls back to the old first-marker
+  adjacency anchor so stage coverage never regresses); single-marker stages/props keep the exact
+  prior path. The shared acceptance test is factored into `validate_block` (the connectivity
+  heuristic is relaxed for *derived* grouped parts, which are pinned by in-range + consistency, so
+  small flat fins aren't mis-rejected). Render self-check: `sylpheed-cli mesh render DeltaSaber_T.xpr
+  --only f001` (exact-name match excludes the `_rou_f001_mnv*` animation poses) → clean complete
+  fighter. Test `hero_ship_grouped_pool_decodes`. Colours/UVs still pending the running-game oracle.
+- 2026-07-18 (refinement) — **4-byte vertex-pool alignment + weapon recovery.** The grouped-pool
+  rule "index pool ends exactly where the vertex pool begins" is really "the vertex pool is **4-byte
+  aligned** after the index pool": `vb0 = align4(ib0 + span)`, so 0..=3 bytes of padding can sit
+  between them. DeltaSaber's index pool ended already-aligned (pad 0), which hid this; **19
+  weapon/`*_hangar` models** (single- and multi-marker: `wep_08/11/34/58/62/69/81/83…`) have pad 2
+  and so decoded to *nothing* — the viewer then showed them as a flat 2D texture instead of a model.
+  Fix: both anchors try `pad ∈ 0..=3` (`ib = vb − idx_bytes − pad` for the single-block adjacency
+  anchor; `ib0 = vb0 − span − pad` for the grouped pivot), validated — a wrong pad reads shifted
+  indices → agreement collapses < 0.85, so only the true pad passes. pad>0 in the ungated stage path
+  is gated at a strict 0.85 to avoid a false anchor; pad 0 keeps its exact prior behaviour (stages
+  unchanged). Result: all 19 now decode as clean models (e.g. `wep_34` 1243 v / 1233 t, a
+  long-barrelled gun-pod; `wep_08` 3 sub-meshes / 478 t). Viewer routing already falls through
+  `from_xpr2` → `anchor_models(0.85)` for single-XBG7 files, so the recovered grouped/padded weapons
+  now preview as meshes.
+- 2026-07-18 (refinement 2) — **pivot on the largest sub-mesh; all 19 recovered.** Three weapons
+  (`wep_81`, `wep_81_hangar`, `wep_30_hangar`) still declined because the grouped pivot validated
+  `markers[0]`, which for these is a tiny *elongated* lead bracket that fails the connectivity gate
+  even when perfectly placed. Fixed by pivoting the alignment check on the **largest** marker (max
+  index count) — the sub-mesh whose triangle-quality/connectivity signature most reliably confirms
+  `(ib0, vb0)`. Once the pivot validates, markers up to it are read unconditionally (a legitimately
+  tiny/flat lead part may fail the quality gates yet still be real), and markers after it stay
+  validated so a stray trailing marker ends the chain. Result: **all 19 previously-declined weapons
+  decode** (`wep_81` 460 t missile w/ tail fins; `wep_30_hangar` 334 t). DeltaSaber unchanged (its
+  body IS the largest marker → same pivot). 7/7 disc tests green, stage quality audit unchanged.

tools/analyze_drawlog_wvp.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""Recover each ship part's runtime WORLD transform from a Canary draw log.
+
+The draw logger (branch capture-drawlog) now dumps, per distinct draw:
+  - the first ~8 object-space vertex positions (pre-transform), and
+  - the vertex-shader float constants c0..c31 (which hold the transform matrices).
+
+The game bakes each part's WORLD matrix into the constants, so the shader's
+World*View*Proj (WVP) for a part is  VP * part_World.  The body (bdy_01) is drawn
+with an identity world, so body_WVP = VP.  Hence for any part drawn in the SAME
+frame (the ship's parts always are):
+
+    part_World = inv(body_WVP) * part_WVP          (View/Proj cancel)
+
+We don't know a-priori which 4 consecutive constants form the WVP, so we try every
+block c[i..i+3] and keep the one for which inv(body_WVP)*part_WVP is a RIGID
+transform (orthonormal 3x3 + translation) for the parts — that uniquely identifies
+the matrix block.  The translation column of each part's world matrix is the
+placement we want (esp. the ±X nacelle offset the static XBG7 graph lacks).
+
+Usage:  analyze_drawlog_wvp.py xenia_re_draws.log
+Identify the body draw via --body-indices (the draw with the most indices) or by
+matching object-space positions to our decoded sub-meshes.
+"""
+import sys, re
+import numpy as np
+
+def parse(path):
+    draws = []
+    cur = None
+    consts = {}
+    mode = None
+    for line in open(path):
+        if line.startswith("DRAW "):
+            if cur is not None:
+                cur["consts"] = consts
+                draws.append(cur)
+            cur = {"header": line.strip(), "positions": [], "textures": [],
+                   "size_words": 0}
+            consts = {}
+            mode = None
+            m = re.search(r"index\[base=0x([0-9A-Fa-f]+) count=(\d+)", line)
+            cur["idx_base"] = int(m.group(1), 16) if m else None
+            cur["idx_count"] = int(m.group(2)) if m else None
+            m = re.search(r"indices=(\d+)", line)
+            cur["indices"] = int(m.group(1)) if m else None
+        elif cur is not None and line.strip().startswith("stream "):
+            # Largest vertex buffer (size_words) = the hull (body, identity world).
+            m = re.search(r"size_words=(\d+)", line)
+            if m:
+                cur["size_words"] = max(cur["size_words"], int(m.group(1)))
+        elif cur is None:
+            continue
+        elif line.strip().startswith("positions:"):
+            cur["positions"] = [tuple(map(float, t.split(",")))
+                                for t in re.findall(r"\(([^)]+)\)", line)]
+        elif line.strip().startswith("textures:"):
+            cur["textures"] = re.findall(r"base=0x([0-9A-Fa-f]+)", line)
+        elif line.strip().startswith("vsconst"):
+            mode = "vsconst"
+        elif mode == "vsconst":
+            m = re.match(r"\s*c(\d+)\s+(-?[\d.]+)\s+(-?[\d.]+)\s+(-?[\d.]+)\s+(-?[\d.]+)", line)
+            if m:
+                consts[int(m.group(1))] = np.array([float(m.group(i)) for i in range(2, 6)])
+            else:
+                mode = None
+    if cur is not None:
+        cur["consts"] = consts
+        draws.append(cur)
+    return draws
+
+def wvp_from(consts, i):
+    """4x4 from constants c[i..i+3] as rows."""
+    if not all(k in consts for k in (i, i+1, i+2, i+3)):
+        return None
+    return np.array([consts[i], consts[i+1], consts[i+2], consts[i+3]])
+
+def is_rigid(M):
+    """3x3 upper-left orthonormal (rotation/reflection), within tolerance."""
+    R = M[:3, :3]
+    if not np.all(np.isfinite(R)):
+        return False
+    G = R @ R.T
+    return np.allclose(G, np.eye(3), atol=1e-2) or np.allclose(G / (np.trace(G)/3), np.eye(3), atol=1e-2)
+
+def main():
+    if len(sys.argv) < 2:
+        print(__doc__); return
+    draws = parse(sys.argv[1])
+    print(f"parsed {len(draws)} draws")
+    # Body = the draw with the largest vertex buffer (the ~10891-vert hull, drawn
+    # with an identity world). idx_count is unreliable (the hull's 9 material
+    # groups dedup to the first group's partial count).
+    body = max(draws, key=lambda d: d.get("size_words") or 0)
+    print(f"body draw: {body['header'][:90]} size_words={body['size_words']}")
+    # Try every constant block; report the one giving rigid part transforms.
+    for i in range(0, 29):
+        bwvp = wvp_from(body["consts"], i)
+        if bwvp is None:
+            continue
+        try:
+            binv = np.linalg.inv(bwvp)
+        except np.linalg.LinAlgError:
+            continue
+        worlds = []
+        rigid = 0
+        for d in draws:
+            pw = wvp_from(d["consts"], i)
+            if pw is None:
+                continue
+            W = binv @ pw           # column-vector convention; may need transpose
+            worlds.append((d, W))
+            if is_rigid(W) or is_rigid(W.T):
+                rigid += 1
+        if rigid >= max(2, len(worlds)//2):
+            print(f"\n=== WVP block c{i}..c{i+3}: {rigid}/{len(worlds)} rigid ===")
+            for d, W in worlds:
+                # translation is the last column (or last row, convention-dependent)
+                tcol = W[:3, 3]
+                trow = W[3, :3]
+                print(f"  idx={d['idx_count']:>6} base={d['idx_base'] and hex(d['idx_base'])}"
+                      f"  t_col=({tcol[0]:+.2f},{tcol[1]:+.2f},{tcol[2]:+.2f})"
+                      f"  t_row=({trow[0]:+.2f},{trow[1]:+.2f},{trow[2]:+.2f})")
+
+if __name__ == "__main__":
+    main()

tools/extract_movie_subtitles.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+import struct, zlib, re, os
+MOD=0x00FFF9D7; RECIP=0x80031493
+def rol(v,n,b=32):
+    v&=(1<<b)-1; return ((v<<n)|(v>>(b-n)))&((1<<b)-1)
+def nh(name):
+    bs=bytearray(name.encode('latin-1'))
+    for i,b in enumerate(bs):
+        if 0x41<=b<=0x5A: bs[i]=b+0x20
+    a=0;bb=0
+    for byte in bs:
+        c=(byte if byte<0x80 else byte-0x100)&0xFFFFFFFF
+        a=((rol(a,8)&0xFFFFFF00)+c)&0xFFFFFFFF
+        bb=(bb+c)&0xFFFFFFFF
+        hi=((a*RECIP)>>32)&0xFFFFFFFF; q=rol(hi,9)&0x1FF
+        a=(a-(q*MOD))&0xFFFFFFFF
+    return (((bb<<24)&0xFF000000)|(a&0xFFFFFF))&0xFFFFFFFF
+
+EX="/home/fabi/RE Project Sylpheed/sylph_extract"
+def load_pak(base):
+    """base like dat/movie/eng ; returns list of (hash,off,size), data bytes"""
+    pak=open(f"{EX}/{base}.pak","rb").read()
+    n=struct.unpack(">I",pak[4:8])[0]
+    es=[struct.unpack(">III",pak[16+12*i:28+12*i]) for i in range(n)]
+    p00=f"{EX}/{base}.p00"
+    data=open(p00,"rb").read() if os.path.exists(p00) else pak
+    return es, data
+def unz(blob):
+    if blob[:2]==b'Z1':
+        zi=blob.find(b'\x78\xda');
+        if zi<0: zi=blob.find(b'\x78\x9c')
+        return zlib.decompress(blob[zi:])
+    return blob
+
+# --- verify the track key scheme ---
+es,data=load_pak("dat/movie/eng")
+keys={h for h,_,_ in es}
+for mv,exp in [("S00A",0x6F2D9663),("hokyu_DS_s02A",0x3662B1F8),("RT01C_1",0x756F69FB),("S16A",0xEEB85377)]:
+    k=nh(f"subtitle_{mv}.tbl")
+    print(f"subtitle_{mv}.tbl -> {k:08X}  expect {exp:08X}  inpak={k in keys}  {'OK' if k==exp else 'MISMATCH'}")
+
+# --- build MSG_DEMO text table from GP_MAIN_GAME_E ---
+es2,data2=load_pak("dat/GP_MAIN_GAME_E")
+text={}  # 'MSG_DEMO_ddd_ppp_ll' -> string
+keyrx=re.compile(r'^MSG_DEMO_\d+(_\d+)*$')
+for h,o,s in es2:
+    dec=unz(data2[o:o+s])
+    if dec[:4]!=b'IXUD': continue
+    # walk null-terminated utf-16le strings in the whole block
+    toks=re.findall(rb'(?:[\x20-\x7e]\x00){2,}', dec)
+    toks=[t.decode('utf-16-le') for t in toks]
+    # pair: text followed by its key
+    for i in range(len(toks)-1):
+        if keyrx.match(toks[i+1]) and not keyrx.match(toks[i]):
+            text[toks[i+1]]=toks[i]
+print(f"\nMSG_DEMO text lines loaded: {len(text)}")
+
+# --- render a movie's subtitle transcript ---
+def track_for(mv):
+    k=nh(f"subtitle_{mv}.tbl")
+    for h,o,s in es:
+        if h==k: return unz(data[o:o+s])
+    return None
+# index text by integer demo id -> ordered list of lines
+bydemo={}
+for k,v in text.items():
+    m=re.match(r'MSG_DEMO_(\d+)_',k)
+    if m: bydemo.setdefault(int(m.group(1)),[]).append((k,v))
+for d in bydemo: bydemo[d].sort()
+
+def transcript(mv):
+    dec=track_for(mv)
+    if not dec: return f"[no track for {mv}]"
+    toks=re.findall(rb'(?:[\x20-\x7e]\x00){2,}', dec)
+    toks=[t.decode('utf-16-le') for t in toks]
+    out=[f"=== {mv}.wmv ==="]
+    # collect all MSG_DEMO ids and timecodes in order; pair positionally
+    seq=[('id',int(re.match(r'MSG_DEMO_(\d+)$',t).group(1))) if re.match(r'MSG_DEMO_(\d+)$',t)
+         else ('tc',t) if re.match(r'\d\d:\d\d\.\d\d$',t) else ('x',t) for t in toks]
+    ids=[v for k,v in seq if k=='id']; tcs=[v for k,v in seq if k=='tc']
+    for j,demo in enumerate(ids):
+        tc=tcs[j] if j<len(tcs) else "--:--.--"
+        lines=[v for _,v in bydemo.get(demo,[])]
+        out.append(f"  {tc}  [{demo:3d}] "+(" ".join(lines) if lines else "(no text on disc)"))
+    return "\n".join(out)
+
+for mv in ["S00A","S16A","hokyu_DS_s02A"]:
+    print(); print(transcript(mv))