Syplheed-Reborn/docs/HANDOFF-ship-placement-2026-07-26.md

Handoff — capital-ship placement: static assembly EXACT (2026-07-26)

Branch feature/ipfb-idxd-parser, commits c6ca5ce..ea32e77. Canary fork branch capture-ship-placement, tip 877163792.

Outcome

Capital ships assemble exactly from the ISO alone — no captures needed. Runtime captures served as the oracle that cracked the static encoding and now remain as verification tooling only.

The two format discoveries (docs/re/ship-placement-runtime-capture.md)

1. Node-TRS off-by-one (mesh.rs::read_trs9): an XBG7 scene-graph node's joint table holds NINE keyframe channels [TX TY TZ RY RX RZ SX SY SZ] behind EIGHT pointer slots shifted one track forward — channel k+1 = f64@(ptr[k]+8), channel 0 (TX) = f64@(ptr[0]−0x48). The old read never saw TX (why hulls stacked on the centreline) and misassigned TY/RY.
2. Euler order (mesh.rs::node_rotation): Ry(ch3)·Rx(ch4)·Rz(ch5), angles direct — pinned by decomposing the captured engine-nacelle rotation Rx(−15°)·Rz(30°) element-for-element.

assemble_ship (ship.rs) — fully static, capture-validated

• Every composite-node instance (alias duplicates collapsed), including cross-id turret mounts (rou_e303_wep_01_root ×2 on e106; e108 places 7).
• Engine cluster rig e_rou_<id>_eng mounts at GN_Engine_01 (two mirrored nacelles + centre engine).
• brg/sld at their exact GN frames (frames were always exact).
• Shared-geometry ±TX twins: the instance whose offset opposes the geometry's dominant side draws X-reflected (viewer reverses winding on det<0).
• exhaust_frames(): the GN_Jet/GN_SJet frames — the game renders its engine-exhaust FX there; the viewer draws marker cones at them (the engine GEOMETRY is recessed in the hull by design; without the FX ships read engine-less — that was the "engines inside the hull" report).

Validation chain (all green)

• ship::tests::static_assembly_matches_runtime_capture (SYLPHEED_ISO): static == the baked e106 capture, all 8 parts, T<1.0 R<0.02, nacelles ×2, turrets ×2, starboard mirror.
• examples/capture_verify.rs over the user's 5 multi-angle snapshots (43 e106 instances): dominant clusters match static to ~1 unit on every part incl. both nacelles. Run: cargo run --release --example capture_verify -- <Stage.xpr> e106 <logs…>
• Offline audit (examples/ship_audit.rs) across all stages: no outliers among composite ships; composite-less families (fighter morph sets, turret/prop parts) are filtered from the viewer's Ships catalog.
• 81 formats-lib + all disc tests; viewer builds.

Canary instrumentation (branch capture-ship-placement)

F10 → xenia_ship_capture_NN.log per press (multi-angle in one run), de-duped by (vertex-buffer, c0..c2 WVP hash) so every instance of every part is recorded. Launch:

cd xenia-canary/build/bin/Linux/Release
./xenia_canary "<…/Project Sylpheed - Arc of Deception (USA, Europe) (En,Ja).iso>"

Open / next

• Verify remaining capital ships against the existing 5 snapshots (they contain every on-screen ship): e105 (its engine block assembles dorsally — the one placement to confirm), f105/f106/f101, e102. Same capture_verify invocation with the other ship id.
• AA-gun models at GN_AAGun* frames need the Vessel recipe binding.
• SJet (maneuvering-vernier) marker orientation is approximate; main aft plumes verified.
• _mov node animation keys (hull-opening sequences) — rest pose rendered now.
• Viewer GUI verification by the user (exhaust cones, catalog cleanup).