nx_nxa_posewrite_lib.nx

nx_nxa_posewrite_lib.nx -- THE ONE POSE-SECTION WRITER for the NXA format (spec: knowledge/nxa_format_spec.md POSE). WHY (measured 2026-09-05, /compare/modding MD3): the estate had exactly ONE writer of a POSE section -- inline inside nx_nxa_pose freeze, an organ that is SOURCE-ONLY -- and nx_nxa_retarget, which computes a complete pose (a world-delta quat and a dt per target joint, the exact POSE entry shape), published it by BAKING the posed VERT in place instead. The baked donor is a statue with a different shape: nx_nxa_play reads poses_in_library=0 joints_posed=0 on every retargeted output while the house rig reads 104/104, so nothing downstream can replay, hold, or sequence the pose. A writer that lives in one organ and is re-typed in the next is the duplicate-ruler defect; this lib is that writer, composed by both (nx_nxa_retarget emit=pose today; nx_nxa_pose freeze is the migration owed). Layout written (nx_nxa.nx + the nx_nxa_pose incumbent): header 4 words [magic, NXA_VER, nsections, toc_check]; TOC 4 words per section [tag, byte_offset, word_len, payload_check]; payloads follow in TOC order. POSE payload: [count, then per pose: pose_id, nentries, entries of NPW_WPE=8 words (joint, qx, qy, qz, qw q12, dtx, dty, dtz MODEL units)]. Every existing section EXCEPT an old POSE is copied unchanged; the new POSE payload keeps every old pose whose id differs from the one written (so re-emitting an id REPLACES, never DUPLICATES) and appends the new one LAST. q and dt are indexed by JOINT id (q[joint*4], dt[joint*3]) and jl lists the joints to emit, so a dense per-joint table (the retarget) and a tracked subset (the freeze) share one signature. license_tier: ORIGINAL No hw writes (Rule 26).