feat(skills): motion-blur-streak — per-frame-driven form for elements riding a baked track

[jbernard077]

rules/motion-blur-streak.md teaches two paths (directional SVG blur, echo/ghost trail) that both
ride a tween window. An element driven per frame from a baked track — a tracked insert or
callout, a title matched to camera shake — has no window to share. This PR adds that form as a
carve-in of the same rule, measured on both render lanes.

What's added

• Per-Frame-Driven Carve-In section: the envelope becomes a per-frame function of the track's
  own delta inside the frame-lookup driver — speed = hypot(dx, dy), axis from atan2(dy, dx).
  • Path B, driven form (the default for driven subtrees) — text-free silhouette clones of the
    lead (same box, border, radius, fill; no content; aria-hidden + data-layout-allow-overlap)
    trailing at −0.25 / −0.5 of the per-frame delta, opacity min(0.32, 0.05 × speed) above a
    4 px/frame gate and slaved to the lead's own opacity.
  • Path A, driven form — rotate a filter wrapper onto the travel axis, feGaussianBlur "σ 0"
    with σ = min(SIGMA_MAX, 0.5 × speed), counter-rotate the content.
  • Which path for a driven subtree, with the measurement behind it (below).
• Four Values rows (TRAIL_GATE, TRAIL_K · TRAIL_MAX, TRAIL_OFFSETS, SIGMA_MAX), one
  constraint, an updated description/tags line, the rules-index.md row, and the hook-regenerated
  skills-manifest.json.

Validation

A product-free driven-card composition (one baked, seeded track; a per-frame canvas plate and a
3D-transformed panel loading the rasterizer) rendered as PNG sequences with static-frame dedup off,
hashed frame by frame:

Variant	Lane	Frames	Result
plate + panel only	hardware GL	120 × 2 renders	120/120 identical
driven Path A (per-frame feGaussianBlur)	hardware GL	120 × 2	120/120 identical
driven Path A	software (SwiftShader)	120 × 2	120/120 identical
driven Path B (ghost trail)	hardware GL	120 × 2	120/120 identical
driven Path B	software	120 × 2	120/120 identical
stress: three filtered cards, σ up to 20, 15 s	hardware GL	450 × 3	450/450 identical, both pairs
stress: three filtered cards	software	450 × 2	450/450 identical
stress: three ghost-trail cards, 15 s	hardware GL / software	450 × 2 each	450/450 identical on both lanes
preview (Path A and Path B side by side on one track)	hardware GL / software	180 × 2 each	180/180 identical on both lanes

hyperframes check: 0 errors, 0 warnings on the preview composition (the three-card stress variants
are not check-clean by design — the cards cross the lab's own mode label and GL-renderer HUD
readout, which a real composition would not carry — and are used only for the hash measurement).

The rule therefore ships the driven form as a preference, not a prohibition: Path B is the
default for driven subtrees because it costs no per-frame filter raster and its text-free clones give
the layout and contrast checkers nothing new to flag; Path A
is legitimate when a true blur is the look, with the note that a per-frame-driven filter is the first
thing to re-measure on a new host (an earlier build, 2026-08, saw 1-LSB feGaussianBlur jitter
under per-frame load and shipped Path B for that reason).

Fit against current HEAD

Anchors re-verified against main @ 250afc3: the section sits between the Camera-Travel
Carve-Out and Values; the existing Recipe and Variations are unchanged, Critical Constraints gains
one cross-reference rewording (two carve-outs instead of one) and one new driven-form bullet.

Preview media

(one baked track, two cards: Path A driven blur on the left, Path B text-free ghost trail on the
right; HUD shows per-frame speed and angle; the track stops for the last second so both cards
settle sharp)

pr-c-driven-trail.mp4

🤖 Generated with Claude Code