Clone scroll transform pin and time tracks

[hongnoul]

Summary

• extend scroll-effect capture to serialize non-text visual scrub tracks, sticky-safe pin tracks, and threshold-triggered time tracks
• add direct replay runtime with support gating and honest report-only fallback for unsupported shapes
• add deterministic fixture and behavioral script that drives the generated clone at scroll positions and validates visual style changes, pin stability, and time state flips

Closes #46

Validation

• scripts/test-clone-scroll-tracks.sh
• pgrep -af "python3 -m http.server.*hwatu-issue46-scroll-tracks" || true
• cargo test -q
• cargo clippy -q --all-targets --all-features

[hongnoul]

Review verdict: changes requested (comment-only self-review).

P1 – tracks are marked replayed while changed properties are dropped. direct_style_supported returns true when any supported field changes, but does not reject other changed fields. The included timed fixture changes color; the capture key records it and the generated transition names it, yet writeStyle never writes color. The report still says replayed, and the behavioral test checks only state/opacity/transform. Implement color/background/text-shadow replay or classify tracks with any unhandled changed field as report-only, then assert computed color in the fixture.

P1 – the matrix support gate accepts shapes the runtime cannot preserve. supported_matrix accepts every matrix(...)/matrix3d(...), while matrix() retains only sx/sy/tx/ty, discarding rotation, skew, and most 3D terms. It also uses p[i] || 1, turning a legitimate zero scale into 1. Restrict the gate to axis-aligned scale/translate matrices or interpolate all coefficients, and add rotation/skew/scale-zero tests.

Validation: CI is green; both new focused Rust tests passed independently. Those tests currently encode only the happy-path subset above.

[hongnoul]

Follow-up findings from an independent pass:

P1 – new tracks can suppress the existing text-effect capture. The text-group loop still stops on if (effects.length >= 3), but effects is now pre-seeded by pin/visual/time tracks. Three of those cause the loop to capture zero scroll-coupled-text-style effects, a regression from the old cap of three text groups. Count text-style effects separately.

P1 – retagging can invalidate an earlier replay selector. The time-track loop assigns c.el.dataset.hwatuScrollEffect = effectIndex without excluding an element already tagged as a visual/pin track. The old serialized selector then matches nothing while the report still marks it replayed. Skip already-tagged elements or use independent per-kind attributes.

P2 – worst-case capture latency jumps by ~42 seconds. Up to 80 text candidates each incur a 520 ms sequential dwell, even when no time effect exists. Gate probes on evidence from the coarse sweep or substantially reduce/batch the candidate set.