Parse tfrag terrain blocks and decode all vertex positions

Adds gc-formats::tfrag and a gated real-disc test that decodes every
terrain patch on all 27 levels.

Two rules decide whether this works at all, and both were wrong on the
first attempt:

The vector format is encoded in the VIF command byte, not the immediate.
Bits 2-3 are vn (component count minus one) and bits 0-1 are vl, giving
element_size = ((32 >> vl) * (vn + 1)) / 8. Terrain positions are vn=2,
vl=1 (command 0x69) at 6 bytes. Reading the format from the immediate's
low nibble -- the more commonly quoted layout -- yields 12 or 16 bytes
and misaligns everything after it.

Unpack payloads pad to 4 bytes, and several non-unpack commands carry
inline data: STROW/STCOL 16, STMASK 4, MPG num*8, DIRECT/DIRECTHL
imm*16. Ignoring those desynchronises the walk, and a desynchronised
walk still finds code words that look like positions.

Positions are also not all in one list. vert_count is the total across
three tiers, the LOD-2 list precedes the common one in offset order, and
the LOD-0-only list starts at shared_ofs + lod_1_size*16 rather than at
an offset of its own. Scanning only the common list matches about a
quarter of patches, each exactly -- which is what made it convincing.

The invariant is that every patch yields exactly its own vert_count. It
caught both errors, so tfrag_positions returns None on a mismatch rather
than partial geometry. All 27 levels: 18,102 patches, 821,439 vertices,
zero mismatches.

Still open: strips, triangle indices and the RGBA vertex-info arrays.
Positions alone are not drawable.

crates/gc-extract/tests/disc.rs

@@ -385,3 +385,105 @@ fn level_geometry_decompresses() {
         core.len()
     );
 }
+
+/// The terrain itself. This is the first test to touch real level content: it
+/// parses the tfrags directory and decodes vertex positions from every patch.
+///
+/// The gate is deliberately about *coverage*, not correctness of a sample: a
+/// parser that works on one tfrag and silently returns `None` on the rest would
+/// otherwise pass.
+#[test]
+fn tfrag_vertices_decode_for_every_level() {
+    require_disc!();
+    // Opening the image also confirms the disc is readable, not merely present.
+    gc_formats::iso9660::IsoImage::open(std::path::Path::new(
+        &std::env::var("GC_DISC").expect("GC_DISC is set by require_disc!"),
+    ))
+    .expect("ISO should open");
+    // Where the unpacked data lives. GC_EXTRACTED overrides; otherwise the
+    // default location used during development.
+    let data = std::env::var_os("GC_EXTRACTED")
+        .map(PathBuf::from)
+        .unwrap_or_else(|| {
+            let home = std::env::var("HOME").unwrap_or_else(|_| ".".to_string());
+            PathBuf::from(home).join("Documents/gc-extracted")
+        });
+    if !data.exists() {
+        eprintln!(
+            "skipped: unpack the disc first, e.g.\
+             gc-extract unpack <disc.iso> ~/Documents/gc-extracted\
+             or point GC_EXTRACTED at an existing unpack"
+        );
+        return;
+    }
+
+    let mut levels = 0usize;
+    let mut total_verts = 0usize;
+    let mut total_tfrags = 0usize;
+
+    for n in 0..27 {
+        let dir = data.join(format!("level{n}"));
+        let core_path = dir.join("core_data.bin");
+        let index_path = dir.join("core_index.bin");
+        if !core_path.exists() || !index_path.exists() {
+            continue;
+        }
+        let core = std::fs::read(&core_path).expect("core_data");
+        let index = std::fs::read(&index_path).expect("core_index");
+
+        // The tfrags block starts at the core header's 0x08 field and runs to
+        // the next region that exists.
+        let region_start =
+            |off: usize| -> i32 { i32::from_le_bytes(index[off..off + 4].try_into().unwrap()) };
+        let start = region_start(0x08) as usize;
+        let end = [0x0c_usize, 0x10, 0x14]
+            .iter()
+            .map(|&o| region_start(o))
+            .filter(|&v| v > 0)
+            .min()
+            .unwrap_or(core.len() as i32) as usize;
+
+        let block = &core[start..end];
+        let header = gc_formats::tfrag::TfragsBlock::parse(block).expect("tfrags block header");
+        assert_eq!(header.table_offset, 0x40, "level{n}: table offset");
+
+        let table_offset = header.table_offset as usize;
+        let mut decoded = 0usize;
+        let mut verts = 0usize;
+
+        for i in 0..header.tfrag_count as usize {
+            let th = header.header(block, i).expect("tfrag header");
+            let data_base = table_offset.wrapping_add(th.data as usize);
+            if data_base >= block.len() {
+                continue;
+            }
+            let slice = &block[data_base..];
+            if let Some(p) = gc_formats::tfrag::tfrag_positions(slice, &th) {
+                decoded += 1;
+                verts += p.len();
+                assert_eq!(p.len(), th.vert_count as usize);
+            }
+        }
+
+        assert!(header.tfrag_count > 0, "level{n}: no tfrags");
+        assert_eq!(
+            decoded, header.tfrag_count as usize,
+            "level{n}: only {decoded} of {} tfrags decoded",
+            header.tfrag_count
+        );
+        levels += 1;
+        total_tfrags += header.tfrag_count as usize;
+        total_verts += verts;
+    }
+
+    assert!(levels > 0, "no unpacked levels were found");
+    assert!(
+        total_tfrags > 5000,
+        "expected thousands of tfrags, got {total_tfrags}"
+    );
+    assert!(
+        total_verts > 100_000,
+        "expected 100k+ vertices, got {total_verts}"
+    );
+    eprintln!("decoded {total_verts} vertices from {total_tfrags} tfrags across {levels} levels");
+}

crates/gc-formats/src/lib.rs

@@ -3,4 +3,5 @@ pub mod global_wad;
 pub mod iso9660;
 pub mod level_data;
 pub mod level_wad;
+pub mod tfrag;
 pub mod wad;

crates/gc-formats/src/tfrag.rs

@@ -0,0 +1,588 @@
+//! Terrain fragments: the `tfrags` block of a level's geometry.
+//!
+//! Spec: `docs/formats/tfrag.md`.
+//!
+//! A level's terrain is a set of **tfrags** — small patches of ground, each
+//! with its own bounding sphere, its own vertex set and its own index strips,
+//! organised into three levels of detail. The block is a directory followed by
+//! one data block per tfrag.
+//!
+//! # Block layout
+//!
+//! ```text
+//! +0x00  4     byte offset of the header table, from the start of the block
+//! +0x04  4     number of tfrags
+//! +0x08  4     float, LOD distance parameter
+//! +0x10  ..    padding, zero, up to the table
+//! +table_offset   TfragHeader[tfrag_count], 0x40 bytes each
+//!                 tfrag 0's data block
+//!                 tfrag 1's data block
+//!                 ...
+//! ```
+//!
+//! A tfrag header's `data` field is a byte offset **relative to the start of the
+//! header table**, not to the start of the block. Every other offset in the
+//! record is relative to that tfrag's own data block. Two different bases in one
+//! record is the sort of detail that produces silently wrong geometry.
+//!
+//! # What is recovered here
+//!
+//! Vertex **positions**: a V3_16 signed unpack inside the tfrag's VIF command
+//! list, giving 6 bytes per vertex as three `s16` world coordinates.
+//!
+//! Not yet recovered: the vertex-info array (texture coordinates), the strip and
+//! index arrays that define the triangles, and the ad-gif texture bindings. See
+//! `docs/plan/tfrag.md`.
+
+use crate::buf::{Buf, Error, ErrorKind, Result};
+
+/// Size of one tfrag header record.
+pub const TFRAG_HEADER_SIZE: usize = 0x40;
+/// Size of the block header before the padding.
+pub const TFRAGS_BLOCK_HEADER_SIZE: usize = 0x10;
+
+/// The header at the start of a tfrags block.
+#[derive(Debug, Clone, PartialEq)]
+pub struct TfragsBlock {
+    /// Byte offset of the header table, relative to the start of the block.
+    pub table_offset: u32,
+    pub tfrag_count: u32,
+    /// LOD distance parameter. Preserved; semantics not confirmed.
+    pub lod_distance: f32,
+    /// The second unnamed word. Preserved.
+    pub reserved: u32,
+}
+
+impl TfragsBlock {
+    /// Parse the block header from the start of the tfrags region.
+    pub fn parse(bytes: &[u8]) -> Result<TfragsBlock> {
+        let mut b = Buf::new(bytes);
+        let table_offset = b.u32("tfrags table offset")?;
+        let tfrag_count = b.u32("tfrag count")?;
+        let lod_distance = b.f32("tfrags LOD distance")?;
+        let reserved = b.u32("tfrags reserved word")?;
+        Ok(TfragsBlock {
+            table_offset,
+            tfrag_count,
+            lod_distance,
+            reserved,
+        })
+    }
+
+    /// The tfrag header at index `i`, given the whole block.
+    pub fn header(&self, bytes: &[u8], i: usize) -> Result<TfragHeader> {
+        let off = self.table_offset as usize + i * TFRAG_HEADER_SIZE;
+        if off + TFRAG_HEADER_SIZE > bytes.len() {
+            return Err(Error {
+                offset: off,
+                context: "tfrag header table",
+                kind: ErrorKind::OutOfBounds {
+                    needed: TFRAG_HEADER_SIZE,
+                    available: bytes.len().saturating_sub(off),
+                },
+            });
+        }
+        TfragHeader::parse(&bytes[off..off + TFRAG_HEADER_SIZE])
+    }
+
+    /// Every tfrag header in the block.
+    pub fn headers(&self, bytes: &[u8]) -> Result<Vec<TfragHeader>> {
+        (0..self.tfrag_count as usize)
+            .map(|i| self.header(bytes, i))
+            .collect()
+    }
+}
+
+/// One tfrag's header record.
+#[derive(Debug, Clone, PartialEq)]
+pub struct TfragHeader {
+    /// Bounding sphere: centre xyz and radius w. The centre is a VU1 memory
+    /// address on this game rather than a world position, so it is not usable
+    /// as a world-space bound without further work.
+    pub bsphere: [f32; 4],
+    /// Byte offset of this tfrag's data block, **relative to the table**.
+    pub data: i32,
+    pub lod_2_ofs: u16,
+    pub shared_ofs: u16,
+    pub lod_1_ofs: u16,
+    pub lod_0_ofs: u16,
+    /// Offset of the ad-gif texture payload inside the common VIF list.
+    pub tex_ofs: u16,
+    /// Offset of the per-vertex RGBA array.
+    pub rgba_ofs: u16,
+    pub common_size: u8,
+    pub lod_2_size: u8,
+    pub lod_1_size: u8,
+    pub lod_0_size: u8,
+    pub texture_count: u8,
+    pub rgba_size: u8,
+    pub msphere_ofs: u16,
+    pub light_ofs: u16,
+    pub light_end_ofs: u16,
+    pub cube_ofs: u16,
+    /// Occlusion / visibility index for this tfrag.
+    pub occl_index: u16,
+    /// Total vertex count. `u8`, so at most 255 positions per tfrag.
+    pub vert_count: u8,
+    /// Triangle count at LOD 0.
+    pub tri_count: u8,
+}
+
+impl TfragHeader {
+    /// Parse one 0x40-byte header record.
+    pub fn parse(bytes: &[u8]) -> Result<TfragHeader> {
+        let mut b = Buf::new(bytes);
+        let mut bsphere = [0f32; 4];
+        for slot in bsphere.iter_mut() {
+            *slot = b.f32("tfrag bounding sphere")?;
+        }
+        let data = b.i32("tfrag data offset")?;
+        let lod_2_ofs = b.u16("tfrag lod_2_ofs")?;
+        let shared_ofs = b.u16("tfrag shared_ofs")?;
+        let lod_1_ofs = b.u16("tfrag lod_1_ofs")?;
+        let lod_0_ofs = b.u16("tfrag lod_0_ofs")?;
+        let tex_ofs = b.u16("tfrag tex_ofs")?;
+        let rgba_ofs = b.u16("tfrag rgba_ofs")?;
+        // 0x20..0x2d are single bytes; 0x2d..0x2e then a u16, so read in order.
+        let common_size = b.u8("tfrag common_size")?;
+        let lod_2_size = b.u8("tfrag lod_2_size")?;
+        let lod_1_size = b.u8("tfrag lod_1_size")?;
+        let lod_0_size = b.u8("tfrag lod_0_size")?;
+        let _lod_2_rgba = b.u8("tfrag lod_2_rgba_count")?;
+        let _lod_1_rgba = b.u8("tfrag lod_1_rgba_count")?;
+        let _lod_0_rgba = b.u8("tfrag lod_0_rgba_count")?;
+        let _base_only = b.u8("tfrag base_only")?;
+        let texture_count = b.u8("tfrag texture_count")?;
+        let rgba_size = b.u8("tfrag rgba_size")?;
+        let _rgba_verts_loc = b.u8("tfrag rgba_verts_loc")?;
+        let _occl_index_stash = b.u8("tfrag occl_index_stash")?;
+        let _msphere_count = b.u8("tfrag msphere_count")?;
+        let _flags = b.u8("tfrag flags")?;
+        let msphere_ofs = b.u16("tfrag msphere_ofs")?;
+        let light_ofs = b.u16("tfrag light_ofs")?;
+        let light_end_ofs = b.u16("tfrag light_end_ofs")?;
+        let _dir_lights_one = b.u8("tfrag dir_lights_one")?;
+        let _dir_lights_upd = b.u8("tfrag dir_lights_upd")?;
+        let _point_lights = b.u16("tfrag point_lights")?;
+        let cube_ofs = b.u16("tfrag cube_ofs")?;
+        let occl_index = b.u16("tfrag occl_index")?;
+        let vert_count = b.u8("tfrag vert_count")?;
+        let tri_count = b.u8("tfrag tri_count")?;
+        let _mip_dist = b.u16("tfrag mip_dist")?;
+
+        Ok(TfragHeader {
+            bsphere,
+            data,
+            lod_2_ofs,
+            shared_ofs,
+            lod_1_ofs,
+            lod_0_ofs,
+            tex_ofs,
+            rgba_ofs,
+            common_size,
+            lod_2_size,
+            lod_1_size,
+            lod_0_size,
+            texture_count,
+            rgba_size,
+            msphere_ofs,
+            light_ofs,
+            light_end_ofs,
+            cube_ofs,
+            occl_index,
+            vert_count,
+            tri_count,
+        })
+    }
+}
+
+/// A VIF code word decoded into its fields.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct VifCode {
+    pub irq: bool,
+    /// Command field, bits 30..24.
+    pub cmd: u8,
+    /// Payload word count, bits 23..16. Zero means 256.
+    pub num: u16,
+    /// Immediate, bits 15..0.
+    pub imm: u16,
+}
+
+impl VifCode {
+    /// Re-encode into a code word. Used by tests to build packets.
+    pub fn to_word(&self) -> u32 {
+        ((self.irq as u32) << 31)
+            | ((self.cmd as u32) << 24)
+            | ((self.num as u32) << 16)
+            | self.imm as u32
+    }
+
+    pub fn decode(word: u32) -> VifCode {
+        VifCode {
+            irq: (word >> 31) & 1 == 1,
+            cmd: ((word >> 24) & 0x7F) as u8,
+            num: ((word >> 16) & 0xFF) as u16,
+            imm: (word & 0xFFFF) as u16,
+        }
+    }
+
+    /// Is this an UNPACK? Recognised by the top two bits of the command.
+    pub fn is_unpack(&self) -> bool {
+        (self.cmd & 0b110_0000) == 0b110_0000
+    }
+
+    /// Number of components minus one: 0..3 = 1..4 components.
+    pub fn vn(&self) -> u8 {
+        (self.cmd >> 2) & 3
+    }
+
+    /// Component width: 0 = 32-bit, 1 = 16-bit, 2 = 8-bit, 3 = 5-bit.
+    pub fn vl(&self) -> u8 {
+        self.cmd & 3
+    }
+
+    /// Bytes per unpacked element on disk.
+    ///
+    /// The whole vector format is encoded in the command byte: `vn` gives the
+    /// component count and `vl` the component width. Nothing about the format
+    /// lives in the immediate.
+    pub fn element_size(&self) -> usize {
+        ((32usize >> self.vl()) * (self.vn() as usize + 1)) / 8
+    }
+
+    /// The PS2 vector format name, for diagnostics.
+    pub fn format_name(&self) -> String {
+        const NAMES: [&str; 4] = ["V1", "V2", "V3", "V4"];
+        const WIDTHS: [&str; 4] = ["32", "16", "8", "5"];
+        if self.is_unpack() {
+            format!(
+                "{}_{}",
+                NAMES[self.vn() as usize],
+                WIDTHS[self.vl() as usize]
+            )
+        } else {
+            "not an unpack".to_string()
+        }
+    }
+
+    /// Effective element count: a `num` of zero means 256.
+    pub fn count(&self) -> usize {
+        if self.num == 0 {
+            256
+        } else {
+            self.num as usize
+        }
+    }
+
+    /// Inline payload size in bytes, not counting the 4-byte code word.
+    ///
+    /// An unpack's payload is padded up to 4 bytes. Only a few non-unpack
+    /// commands carry inline data; everything else is a bare code word. Getting
+    /// this wrong desynchronises the walk and yields plausible garbage, so the
+    /// set is enumerated rather than assumed empty.
+    pub fn payload_size(&self) -> usize {
+        if self.is_unpack() {
+            (self.count() * self.element_size() + 3) & !3
+        } else {
+            match self.cmd {
+                STMASK => 4,
+                STROW | STCOL => 16,
+                MPG => self.num as usize * 8,
+                DIRECT | DIRECTHL => {
+                    if self.imm == 0 {
+                        65536 * 16
+                    } else {
+                        self.imm as usize * 16
+                    }
+                }
+                _ => 0,
+            }
+        }
+    }
+
+    /// Total packet size in bytes, including the 4-byte code word.
+    pub fn packet_size(&self) -> usize {
+        4 + self.payload_size()
+    }
+}
+
+/// VIF command bytes for non-unpack commands that carry inline data.
+pub const STMASK: u8 = 0x20;
+pub const STROW: u8 = 0x30;
+pub const STCOL: u8 = 0x31;
+pub const MPG: u8 = 0x4a;
+pub const DIRECT: u8 = 0x50;
+pub const DIRECTHL: u8 = 0x51;
+
+/// The VIF lists that make up a tfrag's data block, in the order the game
+/// transfers them.
+///
+/// The four offsets do not run in ascending order: the LOD-2 list sits *before*
+/// the common list. Each is a byte range from the start of the tfrag's data
+/// block. Positions appear in the common list and the two LOD-0 lists only.
+fn position_lists(h: &TfragHeader) -> [(usize, usize); 3] {
+    // The LOD-0-only list starts here rather than at an offset of its own.
+    let lod0_start = h.shared_ofs as usize + h.lod_1_size as usize * 0x10;
+    [
+        (h.shared_ofs as usize, h.lod_1_ofs as usize),
+        (h.lod_0_ofs as usize, lod0_start),
+        (lod0_start, h.rgba_ofs as usize),
+    ]
+}
+
+/// Collect the V3_16 position payloads from one byte range.
+fn positions_in_list(block: &[u8], start: usize, end: usize, out: &mut Vec<[i32; 3]>) -> bool {
+    if start >= end || end > block.len() {
+        // Empty lists and inverted ranges (a level with no LOD 2) are normal.
+        return start >= end || end > block.len();
+    }
+    let mut pos = start;
+    while pos + 4 <= end {
+        let code = u32::from_le_bytes([block[pos], block[pos + 1], block[pos + 2], block[pos + 3]]);
+        let packet = VifCode::decode(code);
+        if pos + packet.packet_size() > end {
+            return false;
+        }
+        if packet.is_unpack() && packet.vl() == 1 && packet.vn() == 2 {
+            let base = pos + 4;
+            for v in 0..packet.count() {
+                let o = base + v * 6;
+                if o + 6 > end {
+                    return false;
+                }
+                out.push([
+                    i16::from_le_bytes([block[o], block[o + 1]]) as i32,
+                    i16::from_le_bytes([block[o + 2], block[o + 3]]) as i32,
+                    i16::from_le_bytes([block[o + 4], block[o + 5]]) as i32,
+                ]);
+            }
+        }
+        pos += packet.packet_size();
+    }
+    true
+}
+
+/// Vertex positions of a tfrag, in game order.
+///
+/// `block` is the tfrag's data block, starting at its own `data` offset;
+/// the offsets in `header` are relative to that. `vert_count` covers all three
+/// position tiers (common, LOD 0+1 and LOD 0), so positions are gathered from
+/// each list in transfer order.
+///
+/// Returns `None` if the lists cannot be walked cleanly or do not yield
+/// exactly `vert_count` positions: a wrong count means the walk desynchronised,
+/// and returning partial geometry would hide that.
+pub fn tfrag_positions(block: &[u8], header: &TfragHeader) -> Option<Vec<[i32; 3]>> {
+    if header.vert_count == 0 {
+        return Some(Vec::new());
+    }
+    let mut out = Vec::with_capacity(header.vert_count as usize);
+    for (start, end) in position_lists(header) {
+        if !positions_in_list(block, start, end, &mut out) {
+            return None;
+        }
+    }
+    if out.len() != header.vert_count as usize {
+        return None;
+    }
+    Some(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// A block header as measured on level 0.
+    fn block_header() -> Vec<u8> {
+        let mut v = vec![0u8; TFRAGS_BLOCK_HEADER_SIZE];
+        v[0x00..0x04].copy_from_slice(&0x40u32.to_le_bytes());
+        v[0x04..0x08].copy_from_slice(&833u32.to_le_bytes());
+        v[0x08..0x0c].copy_from_slice(&32.0f32.to_le_bytes());
+        v
+    }
+
+    #[test]
+    fn parses_the_measured_block_header() {
+        let h = TfragsBlock::parse(&block_header()).unwrap();
+        assert_eq!(h.table_offset, 0x40);
+        assert_eq!(h.tfrag_count, 833);
+        assert_eq!(h.lod_distance, 32.0);
+    }
+
+    #[test]
+    fn parses_a_tfrag_header() {
+        let mut v = vec![0u8; TFRAG_HEADER_SIZE];
+        v[0x00..0x10].copy_from_slice(&[0u8; 16]);
+        v[0x10..0x14].copy_from_slice(&53_312i32.to_le_bytes());
+        v[0x14..0x16].copy_from_slice(&0u16.to_le_bytes());
+        v[0x16..0x18].copy_from_slice(&208u16.to_le_bytes());
+        v[0x1c..0x1e].copy_from_slice(&256u16.to_le_bytes());
+        v[0x3c] = 75;
+        v[0x3d] = 62;
+
+        let h = TfragHeader::parse(&v).unwrap();
+        assert_eq!(h.data, 53_312);
+        assert_eq!(h.shared_ofs, 208);
+        assert_eq!(h.tex_ofs, 256);
+        assert_eq!(h.vert_count, 75);
+        assert_eq!(h.tri_count, 62);
+    }
+
+    #[test]
+    fn header_table_access_is_bounds_checked() {
+        let mut bytes = block_header();
+        bytes.resize(0x40 + 4 * TFRAG_HEADER_SIZE, 0);
+        let block = TfragsBlock::parse(&bytes).unwrap();
+        block.header(&bytes, 3).unwrap();
+        assert!(block.header(&bytes, 99).is_err());
+    }
+
+    #[test]
+    fn vif_code_decoding() {
+        // 0x694b_8022, measured in level 0's first tfrag: a V3_16 unpack of 75
+        // positions to VU address 2.
+        let c = VifCode::decode(0x694b_8022);
+        assert!(c.is_unpack());
+        assert_eq!((c.cmd, c.num, c.imm), (0x69, 0x4b, 0x8022));
+        assert_eq!((c.vn(), c.vl()), (2, 1));
+        assert_eq!(c.element_size(), 6);
+        assert_eq!(c.format_name(), "V3_16");
+        // 75 * 6 = 450, padded up to 452, plus the code word.
+        assert_eq!(c.payload_size(), 452);
+        assert_eq!(c.packet_size(), 456);
+    }
+
+    #[test]
+    fn vector_formats_come_from_the_command_byte() {
+        // (code word, vn, vl, bytes per element)
+        for (word, vn, vl, size) in [
+            (0x6900_8009u32, 2u8, 1u8, 6usize), // V3_16
+            (0x6c00_0000, 3, 0, 16),            // V4_32
+            (0x6d00_c012, 3, 1, 8),             // V4_16
+            (0x6e00_0000, 3, 2, 4),             // V4_8
+            (0x6a00_0000, 2, 2, 3),             // V3_8
+            (0x6800_0000, 2, 0, 12),            // V3_32
+        ] {
+            let c = VifCode::decode(word);
+            assert_eq!(
+                (c.vn(), c.vl(), c.element_size()),
+                (vn, vl, size),
+                "{word:#010x}"
+            );
+        }
+    }
+
+    #[test]
+    fn vif_num_zero_means_256() {
+        assert_eq!(VifCode::decode(0x6e00_0000).count(), 256);
+        assert_eq!(VifCode::decode(0x6e00_0000).payload_size(), 256 * 4);
+    }
+
+    #[test]
+    fn only_some_non_unpack_commands_carry_payload() {
+        // STROW and STCOL move a quadword each; a bare code word moves none.
+        assert_eq!(VifCode::decode(0x3000_0000).payload_size(), 16);
+        assert_eq!(VifCode::decode(0x3100_0000).payload_size(), 16);
+        assert_eq!(VifCode::decode(0x2000_0000).payload_size(), 4);
+        assert_eq!(VifCode::decode(0x4a02_0000).payload_size(), 16); // MPG num=2
+        assert_eq!(VifCode::decode(0x5000_0001).payload_size(), 16); // DIRECT imm=1
+        assert_eq!(VifCode::decode(0x0500_0001).payload_size(), 0); // STMOD
+        assert_eq!(VifCode::decode(0x1400_0000).payload_size(), 0); // MSCAL
+    }
+
+    /// Build a data block holding a single V3_16 unpack in the common list.
+    fn synthetic_tfrag(verts: &[[i16; 3]]) -> (Vec<u8>, TfragHeader) {
+        let mut block = Vec::new();
+        let n = verts.len() as u16;
+        // NUM is 8 bits, so a 256-element unpack has to encode zero.
+        let num = if n == 256 { 0u16 } else { n };
+        block.extend_from_slice(
+            &VifCode {
+                irq: false,
+                cmd: 0x69,
+                num,
+                imm: 0x8009,
+            }
+            .to_word()
+            .to_le_bytes(),
+        );
+        for v in verts {
+            for c in v {
+                block.extend_from_slice(&c.to_le_bytes());
+            }
+        }
+        // The next list starts after the padded packet, not after the raw
+        // vertices, so an odd element count advances the boundary.
+        let common_end = block.len().div_ceil(4) * 4;
+        block.resize(128, 0);
+
+        let mut h = TfragHeader::parse(&[0u8; TFRAG_HEADER_SIZE]).unwrap();
+        h.vert_count = verts.len() as u8;
+        h.shared_ofs = 0;
+        h.lod_1_ofs = common_end as u16;
+        h.lod_0_ofs = common_end as u16;
+        // The LOD-0-only list starts on a quadword boundary at or after the
+        // common list, so round the span up rather than truncating it.
+        h.lod_1_size = common_end.div_ceil(16) as u8;
+        h.rgba_ofs = 128;
+        (block, h)
+    }
+
+    #[test]
+    fn finds_positions_in_a_synthetic_block() {
+        let verts = [
+            [99, -3795, -5203],
+            [1123, -3795, -5203],
+            [-715, -3072, -5203],
+            [1123, -3072, -5203],
+        ];
+        let (block, header) = synthetic_tfrag(&verts);
+        let pos = tfrag_positions(&block, &header).expect("positions should be found");
+        assert_eq!(pos.len(), 4);
+        for (got, want) in pos.iter().zip(verts.iter()) {
+            assert_eq!(*got, [want[0] as i32, want[1] as i32, want[2] as i32]);
+        }
+    }
+
+    #[test]
+    fn an_odd_count_is_padded_to_four_bytes() {
+        // 3 elements = 18 bytes, padded to 20, so the next code word sits at
+        // +24 rather than +22.
+        let verts = [[1, 2, 3], [4, 5, 6], [7, 8, 9]];
+        let (block, header) = synthetic_tfrag(&verts);
+        let c = VifCode::decode(u32::from_le_bytes(block[0..4].try_into().unwrap()));
+        assert_eq!(c.payload_size(), 20);
+        let pos = tfrag_positions(&block, &header).expect("positions should be found");
+        assert_eq!(pos.len(), 3);
+    }
+
+    #[test]
+    fn a_desynchronised_walk_is_reported_rather_than_trusted() {
+        let verts = [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]];
+        let (block, header) = synthetic_tfrag(&verts);
+        // Claim more vertices than the block actually holds.
+        let mut bad = header;
+        bad.vert_count = 40;
+        assert!(tfrag_positions(&block, &bad).is_none());
+    }
+
+    #[test]
+    fn a_tfrag_with_no_unpack_returns_none() {
+        let block = vec![0u8; 128];
+        let mut header = TfragHeader::parse(&[0u8; TFRAG_HEADER_SIZE]).unwrap();
+        header.vert_count = 4;
+        header.shared_ofs = 0;
+        header.lod_1_ofs = 64;
+        header.lod_0_ofs = 64;
+        header.rgba_ofs = 128;
+        assert!(tfrag_positions(&block, &header).is_none());
+    }
+
+    #[test]
+    fn zero_vertex_count_gives_an_empty_result() {
+        let block = vec![0u8; 16];
+        let mut header = TfragHeader::parse(&[0u8; TFRAG_HEADER_SIZE]).unwrap();
+        header.vert_count = 0;
+        assert_eq!(tfrag_positions(&block, &header), Some(Vec::new()));
+    }
+}

docs/plan/level-assets.md

@@ -205,6 +205,9 @@ Next: rather than guess the header further, identify `tfrag` by its own
 structure. Terrain is the largest subsystem and the one that decides whether
 anything draws, so it is the right next target even though it is the harder one.
 
+*(Done — see "tfrag terrain" below. Identifying tfrag by its own structure
+worked where the class-table guess did not.)*
+
 ## `data`: the real target, untouched
 
 6,484 sectors on level 0, down to 2,102 on level 24 — 4.2 to 21.9 MB. This is
@@ -238,3 +241,79 @@ that holds everywhere. The sector/byte ambiguity in Phase 1 survived because it
 happened to be correct on the one container checked by hand. The level header
 was fixed in a single pass because the tiling check failed immediately on
 levels 1, 2, 4, 7, 8, 11, 19 and 20.
+## tfrag terrain: block structure and vertex positions
+
+Settled. `gc-formats::tfrag` parses the block, and a gated test decodes every
+patch on all 27 levels.
+
+### Block and patch framing
+
+The tfrags block opens with a 12-byte header: `table_offset` (always `0x40`),
+`tfrag_count`, and an `f32` LOD distance (`32.0`). The patch table follows at
+`table_offset`, `tfrag_count` records of `0x40` bytes each. On level 0 that is
+833 patches.
+
+A patch header's `data` offset is **relative to `table_offset`**, not to the
+start of the block. Getting this wrong happens to work for the first patch and
+silently produces garbage afterwards.
+
+### Positions live in three lists, not one
+
+`vert_count` is the *total* across all three position tiers. Within a patch's
+data block the VIF lists do not run in ascending offset order — the LOD-2 list
+sits *before* the common list — and the LOD-0-only list has no offset of its
+own, starting instead at `shared_ofs + lod_1_size * 16`. The ranges, in transfer
+order, are:
+
+| List | Range (from the patch data block) | Holds |
+|---|---|---|
+| common | `shared_ofs .. lod_1_ofs` | VU header, ad-gifs, vertex info, origin, positions |
+| LOD 2 | `lod_2_ofs .. shared_ofs` | strips, indices |
+| LOD 1 | `lod_1_ofs .. lod_0_ofs` | strips, indices |
+| LOD 0+1 | `lod_0_ofs .. shared_ofs + lod_1_size*16` | parents, vertex info, positions |
+| LOD 0 | `shared_ofs + lod_1_size*16 .. rgba_ofs` | positions, strips, indices, parents |
+
+Scanning only the common list and expecting `num == vert_count` finds positions
+in roughly a quarter of patches. The other three quarters keep most of their
+positions in the LOD lists. That was the first wrong turn and it looked
+plausible: every patch it did match matched *exactly*.
+
+### The VIF1 code word
+
+Two rules had to be right before any of this worked, and both were wrong on the
+first attempt.
+
+**The vector format lives in the command byte, not the immediate.** Bits 2-3 of
+the command are `vn` (component count minus one) and bits 0-1 are `vl` (0 =
+32-bit, 1 = 16-bit, 2 = 8-bit, 3 = 5-bit), so
+
+    element_size = ((32 >> vl) * (vn + 1)) / 8
+
+Terrain positions are `vn=2, vl=1` — command `0x69` — at 6 bytes each. Reading
+the format out of the immediate's low nibble, which is the more commonly quoted
+layout, yields 12 or 16 bytes and quietly misaligns everything after it.
+
+**Unpack payloads pad to 4 bytes, and several non-unpack commands carry inline
+data.** An unpack advances `(count * element_size + 3) & !3`. `STROW`/`STCOL`
+carry 16 bytes, `STMASK` 4, `MPG` `num * 8`, `DIRECT`/`DIRECTHL` `imm * 16`;
+every other non-unpack is a bare 4-byte code word. Ignoring the first four
+desynchronises the walk, and a desynchronised walk still finds code words that
+look like positions — it just finds the wrong number of them.
+
+The 4-byte padding is easy to misread. With 75 elements the payload is 450
+bytes and the following word lands at +452, where `0x01000404` (STCYCL) reads
+perfectly plausibly. The word at +450 is `0x04040000` (ITOP), which is the real
+one. Taking the plausible-looking reading is how this was briefly wrong.
+
+### Verification
+
+Requiring every patch to yield exactly its own `vert_count` is the invariant
+that caught both errors, and it is the check the parser is built around: a
+mismatch means the walk desynchronised, so `tfrag_positions` returns `None`
+rather than partial geometry. Across all 27 levels: 18,102 patches,
+821,439 vertices, zero mismatches.
+
+### Still open
+
+Strips and triangle indices, the RGBA vertex-info arrays, texture binding, and
+the LOD/morph/cull rules in `TfragProc`. Positions alone are not drawable.

docs/roadmap.md

@@ -20,7 +20,7 @@ early, while they are still cheap to retire.
       header, and the `data` extent is parsed as a container of byte ranges. The
       geometry decompresses: level 0 yields 16,969,296 bytes of level content
       from a 9.7 MB stream.
-- [ ] Next: the geometry directory's class tables, then `tfrag` terrain.
+- [ ] Next: tfrag triangle topology, then the geometry directory's class tables.
 
 The disc being the exact build the reference projects pin is the single
 biggest favourable fact so far: no disc conversion, no regional timing work,
@@ -114,9 +114,16 @@ Phase 1 hands over the extents; this phase works out what each one contains.
 - [x] Region layout of `core_data` established: tfrag, occlusion, sky, collision
       and `textures_base` are offsets, strictly increasing on all 27 levels.
       Terrain is `[0, occlusion)` and carries inline texels.
-- [ ] **tfrag vertex format: not recovered.** The 64-byte record framing and the
+- [x] **tfrag vertex format: recovered.** The tfrags block header gives a table
-      scale-8 fixed-point reading were both tested and disproven; see
+      offset (`0x40`) and a patch count; each patch is a 0x40-byte header whose
-      `plan/level-assets.md` for what was ruled out.
+      `data` offset is relative to the table, not to the block. Positions are
+      V3_16 payloads inside a VIF1 command list, split across three lists per
+      patch (common, LOD 0+1, LOD 0). The whole vector format is encoded in the
+      command byte and every element size follows from it. Verified on all 27
+      levels: **18,102 patches, 821,439 vertices**, each patch yielding exactly
+      its own `vert_count`. See `plan/level-assets.md` and `gc-formats::tfrag`.
+- [ ] tfrag topology: strips, triangle indices and the RGBA vertex-info arrays,
+      which is what turns positions into drawable geometry.
 - [ ] `gameplay` — begins `WAD` + a per-level byte, and does *not* decode with
       the WAD LZ77 decoder. All 27 levels fail identically, so the input is
       wrong rather than the decoder. Ruled out: bit order, header length, the