t2m: composed clips need a PER-FRAME refRoll baseline, not one constant

[fernandotonon]

Follow-up from the review of #1032 (which fixed #1023).

The gap

AnimationMerger::applyMotionClip takes clipRefRoll as one per-role constant and adds it to every frame's twist:

for (int c = 1; c < Jc && c < 22; ++c) {
    const float base = clipRefRoll[c];
    for (int f = 0; f < frames; ++f)
        twistTheta[f][c] += base;
}

That is correct for a single library take, which has exactly one reference frame. It is wrong for a composed clip, whose frames come from several takes with different reference frames.

Measured on the shipped motion-library-v2.json (122 clips, all carrying refRoll):

pair	disagreement
max between any two clips	3.94 rad (226°)
run vs punch	2.13 rad (122°)
walk vs wave	1.01 rad (58°)
walk vs sit	0.56 rad (32°)

"walk then punch" is a prompt MotionComposer really produces, so a 122° wrong baseline over the punch segment is reachable today.

Current state (after #1032)

MotionComposer::compose now emits no refRoll for a multi-take stitch — empty selects the pre-#954 legacy path. That removes the known-wrong constant but also gives up the #954 improvement on every composed clip. Single-take compositions (one action, possibly repeated) keep their baseline, since they have one reference frame.

This mirrors the existing gate on fingers.

What the real fix needs

1. Carry a per-frame roll baseline out of the composer — each Piece already knows its source take, so the baseline can be emitted alongside quats and follow the same cut/blend/trim as every other per-frame channel (rootY is the precedent).
2. Extend applyMotionClip to accept it: either a new std::vector<std::vector<float>> parameter, or promote clipRefRoll to per-frame and keep a one-row form for the single-take case.
3. Blend the baseline across each seam. A step change at a junction would produce exactly the visible roll snap the crossfade exists to avoid — the baseline must ramp with the pose blend.
4. Thread it through all three surfaces (CLI, AnimationControlController, MCPServer), which t2m: GUI 'Generate from text' does not pass refRoll to applyMotionClip (CLI does) #1023 has just aligned.

Why it wasn't done in #1032

That PR is a one-line-per-surface argument fix with a measured before/after. A new retarget parameter plus seam-blended roll is a different change with its own regression surface — and roll is precisely the channel that produced the twisted-arm regressions in #954/#838. It deserves its own before/after on both a Mixamo and a UniRig-style rig.

Verification bar

• "walk then punch" on a Mixamo rig and a UniRig-style rig (Goblin_rigged.glb), before/after render.
• No roll discontinuity at the seam frame (Composition::seamFrames marks them).
• A single-take composition must stay bit-identical to today's output.
• Pin the source clip with variant_index when measuring — "punch" resolves non-deterministically among three takes, and an unpinned comparison measures clip choice rather than the change.