feat(transition,shape): silhouette and blob masks, feathered wipes, path trim and morph

[LeadcodeDev]

Closes #383. Refs #377, #388 — #377 is partial by necessity, see below.

#383 — reveal through any silhouette

mask grows an arbitrary polygon or SVG path about its own bounding-box centre; blob builds a smooth closed silhouette from lobes noise-perturbed points joined by quadratic curves. feather and band_color are shared by both and by all four existing wipe_* through one new primitive that paints the boundary into an Alpha8 mask, blurs it, and tints where the alpha is near 0.5.

feather <= 0 keeps the original hard-clip path byte-for-byte, so no existing scenario changes.

Zero at both ends is a hard short-circuit before any silhouette or blur maths, like zoom_blur and whip: byte-identity with the raw frames at from_scale: 0.2, to_scale: 40, at lobes: 12, wobble: 0.6, and at feather: 40 with a band colour.

A bug the tests caught, not the spec. blob first used iris_max_radius unscaled as its target, and a wobble-shrunk lobe could fall short of a frame corner — leaving a sliver of the outgoing scene visible:

just short of the end the blob's automatically solved covering radius must
already reveal the whole frame, whatever origin it grows from

Fixed by dividing by (1 − wobble). That auto-solve is not in the issue at all.

Reused: iris_max_radius directly, the origin type, and the kind: polygon|path vocabulary from clip_path_to_skia. Not reused: clip_path_to_skia itself, because it resolves percentages against a layout box — a concept that does not exist for two already-painted frame buffers.

⚠️ #377 is shape-only, and stays open

shape.draw_start and shape.path_morph work, with two real primitives behind them: trim_path_between using PathMeasure::get_segment (not a dash-effect hack, so it cannot reproduce the zero-length-dash round-cap dot that #376 just fixed in line.rs), and interpolate_path_data lerping point-by-point when both keyframes share a verb sequence. A mismatch prints to stderr and holds the earlier shape — the node-path precedent, not a snap.

svg, line, arrow and connector still have no draw_start. That needs a new AnimatedProperties field, in a file another workstream owned. So the issue's literal JSON — a style.animation keyframe on any drawable — is not what shipped, and #377 stays open for it.

A real Rust bug worth recording

Path::iter() yields PathIterRec, which is self-referential: its .points() slice points into the Rec's own inline storage. Collecting Vec<PathIterRec> and reading .points() afterwards reads dangling data — it surfaced as Rect { left: 10.0, top: 0.0, right: 1.1598642e29, bottom: 10.0 } instead of the path's actual bounds. Fixed by extracting owned points inside the iterator's map.

Verification

34 new tests across two new files plus five inline. cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (1713) all clean.

Two things flagged rather than buried

• mask's origin is pixels, following iris and zoom_blur, although the issue's example looks fractional. Documented as a deliberate deviation rather than a second inconsistent convention.
• Seven pre-existing Shape { .. } test literals in box_builder.rs and legacy_dispatch.rs needed two None fields each — Rust struct literals are exhaustive and none used ..Default::default(). Mechanical, outside the stated ownership, and named by the agent.

Written comment-free, per the codebase-wide rule from #345.