Entry 351

"The top page's Key of the Twilight: a rigid model keys the sorted group on its own vertex box, not a Bbox chunk"

2026-10-02renderui

Filescrates/piney-desktop/src/assets.rscrates/piney-toppage/tests/toppage.rsdocs/engine/desktop.mddocs/formats/ccs-model.md

Issue #25 and the user's reference: on the top page (INF xdttopen0) the sword carries "Key of the Twilight" in yellow down its blade, lit by an orange glow. The port showed a bare blade.

What the port drew

The text (MDL_xdtworld1, alpha 0.6-0.9) and its glow (MDL_xdthotw, alpha 0.2-0.4, scale 1.0-1.2) are children of the sword (ExtObj parent OBJ_xdtsword). The text sits 2000 units in front of the blade, the glow 3000. Both are translucent (texture flag 0x08) and so go to the layer's sorted group. Both alpha-test GEQUAL trunc(aref * t) with AFAIL FB_ONLY, so each writes Z where its alpha passes. The port keyed both with the w-row formula, M[2][3] / M[3][3], because no Bbox chunk names them. That key is the model's z scale over its depth: the glow (scale over 1) came first, wrote Z over its bright core, and the text behind it failed the depth test. In the opening the text showed at frame 110 and went when the glow came in. Drawn alone, each landed in the right place.

The game's rule

Every rule on the way was checked and agreed:

• the w-row formula itself (ccModel::Draw 0x0013ebbc reads offsets 12, 28, 44 over 60);
• ccDLSort::Add and the in-order walk (StoreDLNode 0x00108100, ascending);
• AREF (ccSetMaterialPacket 0x0013e770, trunc(aref * t));
• ZBUF never masked (set once in SetScreenModeMain);
• the blend type (flag & 3; bit 3 dropped);
• vertex alpha 0x80.

So the game itself was run. tools/test_effect_draw_rs.py's SceneMachine decodes xdttopen0 with the game's own DecodeSetup and builds ANM_xdttop1a with the game's ccAnm. Its matrices match the port's at every step tried (1, 30, 60, 120). Letting ccModel::Draw run natively under the camera's view then stopped in _ccCheckBoundingBoxEx. The game had taken the bounding-box branch.

ccModel::Init (0x0013a550) sets ccModel +4, the box pointer, to the model chunk's own box (ccModelChunk +0x10) for every model without mtype & 6, and to null for a bone or skin model. Decode_Model fills that box for every model through ccBbox_SetBox (0x001388c0). Per axis it takes the integer min and max over all vertices, starting at 0x10000 and -0x10000, doubled about the centre only for mtype & 6. The centre at +0x30 is ((min + max) >> 1) * vertexScale / 4096. Decode_Bbox fills a list of its own and never reaches it.

So a rigid model always keys on the screen Z of its vertex box's centre. The w-row key is only for bone and skin models. The desktop page and the port had it the other way round: a key from a Bbox chunk, and the w row for everything without one. Read from the chunks the game decoded, the centre matches the port's new vertex_box_centre on all 41 models of xdttopen0 and all 138 of xddesk01.

With the real key the text (screen Z 5342) is drawn before the nearer glow (5368). The glow lays its orange over the lettering, as in the reference.

Tests

• the_key_of_the_twilight_is_drawn_before_its_glow (piney-toppage) at frames 150, 300 and 421. It fails with the old keying ("the glow drawn before the text").
• model_box_centres_are_the_games: the sword, text and glow centres as the game's decoder gives them.

The desktop, demo, stream, world and game suites pass unchanged.