a68405216f
Static external-part placement (ship::assemble_ship) is only approximate — a GN_* frame is the mount pivot, and the real offset lives in a runtime detail rig. So capture ground truth from Canary instead. Finding: the guest vertex buffer holds LOCAL coords (byte-identical to the .xpr), so capital-ship parts are placed entirely in the vertex shader. The per-part transform is in the VS float constants — c0..c2 are the WorldViewProjection rows; normalising each by its norm (= projection x/y/z scales, sy/sx = 16:9) yields the rigid WorldView (rows orthonormal, det +1). Expressing every part relative to a reference part cancels the camera: rel_p = WV_ref⁻¹·WV_p, a pure ship-space rigid transform. - examples/correlate_capture.rs: parse xenia_ship_capture.log, decode the ship's parts, match by vertex count, recover each part's ship-relative placement. - docs/re/ship-placement-runtime-capture.md: full method + the capture protocol (Canary branch capture-ship-placement, F10 hotkey) + validated e106 result. Validated on e106 (ADAN Destroyer, Stage_S01): all 7 drawn parts matched by vertex count; bdy_01/bdy_02 recovered as a PORT/STARBOARD pair (X = ∓264) that static had overlapping; engines correctly aft; extent 872×670×2181. Only brg_01 (bridge) missing — culled at that camera angle, needs a bridge-facing capture. HANDOFF: next = re-capture bridge, bake a per-ship placement table the viewer prefers over assemble_ship, then capture the cruiser / ACROPOLIS / TCAF ships. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>