fix(registry): animate mk card offsets with transforms, not top/left - #2960

What

Converts the top/left tweens in mk-background and mk-clone-wall-transition
to x/y transforms, and adds a script that lints registry items the way users
actually receive them.

Why

Both blocks animate the card's top/left. Layout properties snap to integer
device pixels, so the move stutters under the seek-by-frame capture engine
(lint rule: gsap_non_transform_motion).

These shipped in #2939 because registry items had never been linted.
hyperframes lint <dir> expects an index.html, but items ship as
<name>.html, so pointing it at an item directory fails with "No composition
found". The errors only surfaced when I installed all 30 items with
npx hyperframes add and linted the resulting project.

How

Both cards sit at top: 0; left: 0 in CSS, so the tween values carry over to
x/y unchanged. In mk-clone-wall-transition the later scale tween
composes cleanly with the translate (same transform-origin, same visual).

scripts/lint-registry-items.mjs (bun run lint:registry-items) mounts each
item into a throwaway project at the path hyperframes add would use, lints it,
and reports only findings for the item's own file. Verified it reproduces the
original failure on the pre-fix source.

Rules that assume a standalone composition (root_missing_composition_id,
root_missing_dimensions, multiple_root_compositions) are filtered out —
many components ship as paste-able fragments with no root by design, and those
rules fire on ~47 long-standing items.

Test plan

• Manual testing performed

• Re-rendered both blocks and compared frames against the previous output at the
  docked position — identical. This changes how the motion is computed, not
  how it looks.

• bun run lint:registry-items mk-background mk-clone-wall-transition — clean.

• Confirmed the script fails on the pre-fix source (stashed the fix, re-ran, got
  the original gsap_non_transform_motion error back).

• format:check clean.

Not covered: no unit tests (these are content fixes; the render is the gate).

Why this isn't a CI gate yet

Two things have to land first, otherwise the job fails on main:

1. A linter false positive. svg_measure_before_path_d fires on
   hw-path-text and hw-text-cloud. Both assign d via a chained
   document.getElementById("…").setAttribute("d", …); the rule's regex only
   matches <var>.setAttribute("d", …), so it can't see the assignment. The
   items are correct — verified by rendering.
2. Pre-existing debt. A full sweep reports 38 items with real findings
   (font_family_without_font_face, caption_transcript_parse_error) across
   long-standing blocks and components.

Once those are addressed, this should run scoped to the items a PR changes —
the same pattern the file-size check already uses to avoid blaming PRs for
pre-existing offenders.