fix(transition): stop the feather leaking, give the pill its corners, add clip-path blob

[LeadcodeDev]

Closes #432. Part of #438.

The feather leaked onto three fixed edges. A feathered wipe built its mask
from the exact reveal rect, whose three non-moving edges sit on the frame
boundary, and MaskFilter::blur softens a shape's own edge wherever it is. The
feathered path now extends those three sides by feather * 6 + 16, comfortably
past the Gaussian's reach at sigma = feather / 3, and leaves the moving edge
exactly where it was. The hard-clip fast path for an unfeathered wipe is
untouched. Reverting it paints [255, 140, 198] — the band colour — onto pixel
(640, 3), which is the issue's own report.

The pill was a rounded rectangle. IRIS_PILL_CORNER_FRACTION was 0.2. At
1.0 the corner radius equals the smaller half-dimension, which is a true
stadium — and a circle at aspect: 1.0, which iris-transition.md already
documented. The doc was right and the constant was wrong.

clip-path gains blob. It reuses the generator the blob transition
already has rather than growing a second one: blob_local_path and
scaled_mask_path become pub(crate) and nothing else moves. The documented
boundary holds — mask_transition and blob_transition still do not depend on
ClipPath, because a transition has no layout box to resolve a percentage
against.

A transition that ignores a field now says so. The applicability table is
written from apply_transition's own match arms, and a field is reported only
when it differs from its default — so an unset field is never flagged, only an
explicit override that does nothing.

The fifth item, which the workstream's briefing left out

A mask with no silhouette, or one that does not resolve to a path, falls back
to a plain fade and says so on stderr under a zero exit code — which reads as
the transition having been ignored rather than refused. validate names both
cases now, reusing mask_shape_to_local_path so the check and the render agree by
construction rather than by a second opinion.

origin on mask, blob, iris and zoom_blur is in frame pixels. Writing
{ "x": 0.5, "y": 0.5 } for the centre is the natural mistake, and it places the
origin half a pixel from the top-left corner with no complaint. A warning now says
which unit the field takes and what the centre of the frame would be.

That item was missing from the brief I wrote, not from the work: the other four
each carried a reproduction and a file pointer and this one carried neither.

Tests

Nine, each reverted and watched go red. Beyond the two quoted above:

a_pill_with_aspect_one_degenerates_into_a_true_circle_not_a_barely_rounded_square
clip_path_blob_reuses_the_wobbling_lobe_generator_instead_of_drawing_a_plain_circle
a_fade_transition_names_every_field_it_ignores_in_one_warning
a_mask_with_no_silhouette_is_named_rather_than_falling_back_to_a_fade_at_render
a_fractional_origin_is_flagged_because_the_field_is_in_frame_pixels

cargo test -p rustmotion-core --lib engine::transition 22 passed · engine::paint_pass 101 · -p rustmotion-core --tests all green · -p rustmotion --lib cli::commands::validate_schema 56 passed.