Player-character mesh packs Confirmed

meta[1] (0x2CBA0) is the sum of the three descriptors' decompressed sizes (0xB49C + 0x41E0 + 0x1D524) - the ordinary scene-bundle +0x04 word, which nothing reads. It was read as "the byte offset of a VDF data tail past the LZS payload inside this entry", and the entry's own size refutes that: the raw-TOC gap makes it exactly 0x19800 bytes, so 0x2CBA0 is 78 KB past the end.

The loop that produced the misreading is real, but walks another entry. FUN_800520F0 hands each base + offset of a flat [count][offsets] pack to FUN_8001FBCC, the append for the VDF pointer table at 0x80083E58. Its four load constants 0x368 / 0x369 / 0x36a / 0x36b are raw TOC indices 872..875 = extraction entries 870..873, the befect_data block's etim / etmd / vdf / efect. The pack walked is vdf (extraction 872), whose header is a genuine flat pack: count 0x20, then 32 ascending offsets. The character pack is extraction 874 = raw 876 - the same number in the other index space, which is how the two entries were fused.

The editing contract is unchanged: legaia_asset::party_swap::fieldize keeps the first four words byte-exact and pins §0's decoded size at 46 236 bytes. A rebuild that changed that size was observed to hang the next battle load on a wild read at 0x808425F8. With the VDF-pointer mechanism falsified, the shape of that address is now narrowed rather than unexplained: it is not a materialised constant, because no lui anywhere on the disc carries an immediate of 0x8084 or 0x8085 (all 84 images, 2.1 MB), so it is computed - a base plus a word read out of the container. The battle loader's two pack walks were the standing candidates and both are now out: a mid-battle save state shows one walking the vdf pack and the other the etmd pack, neither of which is this container. The walk that does read it is the one inside the boot-time player_data loader, over §0's own decoded pack, and the observed address sits a whole word offset above that buffer's measured base. The container's buffer is also sized from its own header at boot rather than from a constant, so a rebuild that keeps the header byte-exact while changing what the stream decodes to gets a truncated pack - which is how an offset word past the table becomes mesh payload read as an offset.

The upload writes (clut_x, clut_y, w*h, 1), which is why a character spans several CBA columns of one row instead of the declared CLUT rect. The row-479 NPC CLUTs are a separate, STP-set upload. Kept resident across transitions by FIELD_SHARED_BLOCKS.

Only face, eyes, skin and parts of the clothing sample the atlas. Hair, vest and boots are untextured flat / gouraud prims whose (cba, tsb) is (0, 0), so a texture-only renderer samples empty VRAM and leaves holes. The colour block sits at the start of the prim, before the vertex indices, with length vertex_offset: flat prims carry one RGB, gouraud one per corner at a 4-byte stride. legaia_tmd::mesh::tmd_to_vram_mesh_field_hybrid returns the per-vertex shading alongside the mesh; the characters page's field form draws it that way.

Each strip carries authored alternate-face frames (blink / mouth variants) below the live texel rows. Cutscene scripts stamp them over the live cell with the field-VM op 4C 60, a literal-operand MoveImage whose six u16 operands sit at arbitrary byte parity (read through FUN_8003CE9C; handler at 0x801E1B28..0x801E1B90 inside FUN_801DE840). A script that stamps once and never stamps back leaves the cell parked on the alternate frame until the band is re-uploaded. The engine queues the same stamps (World::queue_script_vram_move / apply_script_vram_moves). The actor wrap-scroll (FUN_80021DF4, installed by move-VM opcode 0x1E) scrolls water and ambient cells but never the character atlas.

Per-character animation is otherwise not in the pack: the actor tick consumes a record pointer at actor[+0x4C] installed by FUN_80024CFC, and battle actions feed a parallel struct at actor[+0x234] - see ANM.

On a full-party save, DAT_8007C018[0] = 0x80165E38 = the assembler's ctx+0x50 blob + 0x18 exactly; the assembled TMD reads nobj=17, bone-id bytes [0..14, 200, 201], attach array [5, 8]. With Vahn equipped Hunter Clothes / Survival Knife / Ra-Seru Meta, every one of the 17 object vertex pools byte-matches a PLAYER1 section, equipment-selectively: body only in the id 0x43 section, weapon objects only in id 0x22, the Ra-Seru extra in the Meta tiers, defaults elsewhere. The five equipped-variant objects appear nowhere in PROT 1204; the other 12 default-section objects are byte-shared with it. The full party installs 0x80165E38 / 0x8017A908 / 0x8018D550 into slots 0..2 (game_mode 0x15).

The install loop in FUN_800513F0 is gated per slot by DAT_8007bd10[i], the active-member ID (1 Vahn / 2 Noa / 3 Gala / 0 empty) - not a 0/1 flag. A Vahn-solo fight ([1,0,0,0]) installs only slot 0 there and FUN_800542C8 fills the rest (per-member loop bounded by *(rec+0x4a)). The loader FUN_800520F0 separately registers PROT 0x36a (etmd.dat) into the effect window [3..], never the party slots.

Posed at frame 0 (Y down, battle-world units): Vahn 202 × 425 × 339 with the head at −423 and feet at +3; Noa 217 × 390 × 296; Gala 208 × 503 × 340. Terra's file (0866) is a quadruped - long in Z by design - and has no runtime texture band, so it is not drawn as a party member.

1203 stores u8 << 4 angles where the player streams are full 12-bit, so the two idles agree only up to quantisation. The inline record[0] streams are raw (2 + parts*frames*9 bytes); the character's art clips carry the same decoded bytes as channel-delta-coded "ME" bodies in readef.DAT - two container encodings, one pose format.

Palette port accuracy against the reference capture: Vahn byte-exact, Gala 100 %, Noa ~98 % (late-game equipment patches in the capture). Texture placement reproduces live battle VRAM at 99.7-100 % per member. The 1204 atlases match the live band only at 73-98 % (default-equipment texels).

PROT 1204 is a flat streaming-format container (no LZS wrapper) of exactly 75 sectors (0x25800): five chunks of asset type 0x09 plus a terminator. PROT 1205 holds the eight atlases as type-0x00 TIM chunks at 0x8224 stride. The minigame overlay loads data\field\other5.lzs + PROT 1205/1206.

• “Battle reuses the field pack” and “battle renders PROT 1204 directly, posed from 1203” - the partial byte match was the default-section geometry both sources share; the equipped-variant objects are in no 1204 chunk, and no 1203 record is resident in battle RAM.
• “Seven atlases inside 1204, the last truncated” - an over-reading entry size ran 1204 into 1205; there are eight whole atlases in 1205.
• “Nominal CBA, no relocation, the palette is scene residue at rows 490..497” - those rows hold the scene environment palette; the party palette is uploaded to 481..483 by the battle loader.
• “The per-frame patch loop adds the equipment groups” - it is a pose toggle and never changes the object count. The “+2-row scroll phase” reading of a Noa strip variant was a cutscene blink stamp.

Full entries: do-not-re-walk § battle, § containers, § field.