[AlexZ005]

Roadmap 57, lane 57-chars: M2, M4, M5 (Avatars source), M6, M7, M9. Packs half: theprototype-app/packs#18.

M2 humanoid map + retarget

• Every rig's bones are mapped onto VRM 1.0's humanoid bone set (src/lib/humanoid/humanoidMap.js, a leaf). Detection works from bone names (Mixamo, KayKit, Meshy, VRoid/UniRig, Unreal, Rigify/Quaternius UAL), then from the hierarchy (spine order), then from the rig's shape (numbered UniRig bones). A VRM brings its own map.
• retarget.js: a rest-pose and facing corrected retarget; the hips' travel is scaled to the character's size.
• Apply clip… from a dialog, Inspector ▸ Character, the object menu, and the Animation window's "Apply to…". Every peer computes the fitted clip for itself, so the model's bytes never change and a late joiner computes the same clip. Removing an applied clip undoes cleanly.
• The map can be fixed in Skeleton ▸ Humanoid map. A fix is a replicated config (userData.character); the file is untouched.

M4 VRM characters

• @pixiv/three-vrm 3.5.5 (exact version) loads as a lazy chunk. It reads VRM 0.x and 1.0, including a .glb that carries the VRM extension. Spring bones and expressions work, there's a Set Expression flow node, and the licence meta shows in the Inspector and becomes the model's provenance licence. VRM 0.x models are turned to face +Z.

M6 rig fix-ups (lazy)

• Rename, move in the rest pose, add, delete and mirror bones. Weights: automatic (heat diffusion in a worker), smooth, normalize, plus a weight brush and a test pose.
• Each operation rewrites the GLB, sends it to peers once (objectfile replace:true) and undoes as one step.

M7 provenance + licence warnings

• provenance.js records where every asset came from. Licence warnings show in Publish, Export, Community gallery and Download; they warn and never block. Exports and download zips include a CREDITS.txt.

M9 Download…

• Layouts: (a) one GLB with the clips embedded (default), (b) model + one .tpanim per clip, (c) model only.
• A single clip downloads as .tpanim, .glb, .vrma or .bvh (lazy BVH writer).
• FBX/USD are listed but switched off, with the reason shown. Several selected objects download as one zip.

M5 Asset Browser ▸ Avatars

• Reads the Open Source Avatars registry, pinned to a commit via jsDelivr. CC0 collections are listed by default and CC-BY ones only on request; a CC-BY avatar keeps its credit. Production sends no CSP header, so no cloud change is needed.

Proof

• Unit: vitest 207 files / 2,532 tests pass. New tests cover humanoidMap, retarget, provenance, characterDownload (including a BVH round trip), rigFixes and osaRegistry.

• svelte-check ratchet held at 299/39.

• e2e (from a snapshot, port 5422):

  Suite	Result
  character-apply	ALL PASS, 22 (three peers incl. a late joiner, undo)
  character-vrm	ALL PASS, 26 (two peers)
  character-download	ALL PASS, 28
  character-skeleton	ALL PASS, 24 (two peers)
  character-avatars-source	ALL PASS, 8

• Production build: three-vrm, springs, rig fix-ups, BVH writer, retarget and VRMA writer all stay off the boot graph (boot is 368 chunks / 8,076,062 bytes).

For the 57-clips union

• This branch doesn't contain 57-clips' M1 yet. The planned merge (QUESTIONS-57-chars Q5):
  • register this detector through their setHumanoidDetector
  • wire setClipAssetResolver to their decodeTpanim
  • use retargetClip inside their bindClip
  • fold the duplicate rest-pose readers into one
  • write .tpanim through their encodeTpanim
• The character config moved to userData.character because they use userData.humanoid as a plain bone map.

🤖 Generated with Claude Code

[AlexZ005]

5e5e9d4 — merged 57-clips M1 (#425) and joined the two halves. This PR now carries #425 too; merging it first is fine, as is merging this one alone.

• Their humanoid clip assets (.tpanim) now bind through this lane's retarget: bindClip → new retargetTpClip (corrects for rest pose and facing; bindHumanoid is kept as the fallback). The character's own map, including the user's fixes, is registered through setHumanoidDetector.
• Apply clip… lists the Explorer's .tpanim files under Clip files and applies them as their clip assets (assignClipAsset, one history batch). Rigs that aren't humanoid are refused.
• Download: .tpanim and .vrma are now written by encodeTpanim(toHumanoid(...)), so every clip file the app writes has one shape. The single-clip .glb keeps the rig's own skeleton.
• Merge trap: both lanes added hasSkin() to fileHandler, and git merged them into a duplicate declaration. Only npm run build caught it; I dropped one copy.
• Left as is: restRowsOf and restPoseOf both still exist (folding them is cleanup).

Checks on 5e5e9d4:

• Unit: 2,562 of 2,563 passed. The one failure is a tankSurfaces timeout under load average about 30; it passes 12/12 alone. New: tpanimRetarget (2 tests).
• svelte-check ratchet: 299/39.
• Build: OK, 369 chunks / 8,089,493 bytes, +1 chunk / +13.4 KB, all from M1. The VRM, retarget and encoder code all stay lazy.
• e2e: all pass, 146 checks. character-clip-assets 13 (new: two peers, a Mixamo Kick as a humanoid .tpanim played on the KayKit knight, undo on both), tpanim-clips 25, character-apply 22, character-download 28, character-vrm 26, character-skeleton 24, character-avatars-source 8.