Design: pose-assisted frame-by-frame animation (rigid bones + AI gap filling)

Status: agreed design, not yet implemented (character sheets — the prerequisite — shipped as M5a). This is deliberately not Spine/Toon Boom puppet animation: the output of every pose is an ordinary drawn keyframe on the timeline, keeping the frame-by-frame feel. Bones exist to draft the next frame, and AI keeps the character complete and on-model where the pose reveals things that were never drawn.

The idea in one scenario

Side-view run cycle. The character is drawn once; the near arm covers part of the torso. The animator swings the arm forward with a bone:

1. The arm's strokes move rigidly with the bone — the artist's own lines, re-posed, as the draft of the next frame.
2. The torso region the arm used to cover is now blank, and part of the character's back has rotated into view that never existed in any frame.
3. The AI is asked to fill exactly those regions with strokes, using the document's character sheet (front/side views) as the authoritative design.
4. The result is committed as a normal keyed frame the artist keeps drawing on. Nothing about playback, inbetweening, or export changes.

M5b — Rig: bones + stroke bindings

Per-layer rig, serialized in the document like everything else:

"rig": {
  "bones": [ { "id": "b1", "name": "upper-arm", "parentId": null,
               "x": 410, "y": 220, "length": 60, "angleDeg": -35 } ],
  "parts": [ { "id": "p1", "name": "arm", "boneId": "b1", "zOrder": 2,
               "region": [ {"x":..., "y":...}, ... ],   // closed spline
               "strokeIds": ["s12", "s13"] } ]
}

• Rig mode in the tool bar: draw bones (click joint, drag tip; a bone started on another bone's tip becomes its child), then draw closed spline regions to cut parts. Strokes are assigned by containment; a stroke crossing a region boundary is split at the intersection so each side belongs to its own part (GeometryOps gains point-in-polygon and segment–polygon intersection).
• A part with no region binds a whole layer — the trivial pre-separated case.
• Part is the extension point for the later mesh/spline deformation feature: a lattice per part can replace the rigid transform without touching bones, bindings, or the AI contract.

M5c — Posing → next-frame draft

• Pose mode on a keyed frame: dragging a bone rotates it (and its children) around its joint; each bound part's strokes get the same rigid transform, rendered in zOrder.
• Commit pose creates a new keyed frame whose stroke record is the transformed copy (role: key). It is a draft: the artist draws over it with the normal tools.
• No automatic pose tweening — the deterministic stroke inbetweener stays the interpolation between drawn frames. (Pose-aware inbetweening can be added later as an option once poses exist on keys.)

M5d — AI gap filling

• After a pose, exposed regions = (previous cover of the moved parts) minus (their new cover), via polygon difference on the part regions; coarse convex hulls are enough — the AI needs a where, not a pixel mask. The same computation naturally captures body areas newly revealed to the observer (the run-cycle back: the torso region no longer covered by the arm).
• Request to the artist (API or MCP): posed frame render + gap hull(s) + character-sheet views (ReferenceImages, shipped in M5a) + the strokes neighbouring each gap. Response: strokes in the existing wire format, inserted through the existing validated, undoable path, and tagged with the covering part id so subsequent poses carry them along.
• One button in the UI: “Pose → fill gaps”. Deterministic fallback when no AI is configured: borrow and rigidly fit strokes from the matching part of a reference view.

Later, separate features

• Mesh/spline deformation per part (bend a limb instead of rotating it) — builds on Part without changing the rest.
• Pose-guide mode: skeleton-only keys; the AI draws complete frames from the stick figure + character sheet.