feat(studio): inspector fields, toggles and sections on the shared primitives

[miguel-heygen]

Lands the first inspector sweep PR (U10a: primitives, layout, text, sections, the 3D Transform heading) of the Studio design-system foundation. Stacked on the Renders sweep (#3772). Ratchet 470 to 468. Bundle +7.0 KB gzipped. Stops at a seam: the flat slider and select row keep their hand-rolled implementations and move in the next PR.

What

Lands U10a, the first of the inspector sweep PRs: the inspector's shared value
controls move onto the design system's primitives, and the doubled "3D Transform"
heading is gone.

• CommitField is a boxed field. It wears the same fieldBase as Input,
  NumberField and Select, so the inspector's metric input and every other
  Studio field are one control with one look.
• MetricField and DetailField put their label outside that box.
• SliderControl becomes the shared Slider, SegmentedControl the shared
  Tabs, SelectField the shared Select, FlatToggle the shared Toggle.
• FlatRow drops the underline it drew around the value; the field's own box is
  the boundary now, and the value tier tints it.
• The Layout section stops printing a "3D Transform" label above a component that
  already renders one.
• Section and the 3D Transform header expand on the duration-expand token.
• Colour literals in the ratchet: 470 to 468, with two files reaching zero.

Also fixes a defect in the shared Toggle that this migration surfaced: Base UI
renders the switch as a <span role="switch">, so disabled:opacity-40 compiled
to a real rule that could never match and a disabled switch drew at full opacity
with a pointer cursor. It is data-[disabled]: now, with a test beside the
primitive that fails without the fix.

Why

Two requirements from the design-system contract.

R10 asks that inspector inputs read as inputs, with a boundary that differs from
their label. They did not: a value was bare text on the panel background, and the
metric input measured 17px high with no radius and no background. You could not
tell a value from its own label until you clicked one.

R9 asks that "3D Transform" render once for a selected element. It rendered twice:
the Layout section printed a static label directly above a component whose own
header carries the same words and is the one that collapses.

R8 and KTD7 want one implementation per control. The hand-rolled range input,
segmented strip and native <select> were three more. The segmented strip in
particular carried aria-pressed buttons with no keyboard handling, so the only
way through it was Tab, one segment at a time; on Tabs it answers arrow keys,
Home and End.

How

The public prop signatures of the migrated primitives do not change, so their
twenty-odd consumers across the inspector pick up the new look and the new
keyboard behaviour without a line of their own changing. That is what keeps this
PR small while the visible change is panel-wide.

Telemetry keeps its contract: one event per commit boundary, never per
intermediate value. Each migrated control passes the tracker the section already
holds as onTrack, and the assertions run through a real section rather than the
primitive alone, because the section is what holds the tracker.

data-flat-row-value is deleted rather than moved. The boxed field already
carries data-testid="inspector-field", so a second hook on the same element was
one more thing to keep in step.

Test plan

Studio suite, from packages/studio, with a capped worker pool:

• bunx vitest run --poolOptions.forks.maxForks=4 — 440 files, 4845 tests, green.
  Baseline on the branch point was 440 files / 4841 tests, so the delta is the
  four tests added here and no test removed or weakened.
• bun run typecheck — clean.
• bun run --filter @hyperframes/studio build — succeeds.
• bunx oxlint and bunx oxfmt --check on the changed files — clean.
• bunx fallow audit --fail-on-issues — exit 0 against both the base branch and
  the trunk; no finding in a file this PR touches.
• Studio load smoke, the CI job's own script against a local dev server —
  PASS: studio loaded with schema-valid API fixtures and no runtime errors.

New tests, each proven to fail against the unfixed code:

• The Layout section renders "3D Transform" exactly once. Fails with the removed
  label put back.
• Section expands on duration-expand. Fails with a hard-coded duration.
• The migrated Select trigger classifies the way the native <select> did on
  both hotkey selector lists, asserted true-and-equal so two falses cannot agree
  and prove nothing.
• The segmented control renders a real tablist and answers ArrowRight.
• The shared Toggle wears its disabled look on the attribute Base UI sets.

Screenshots and computed styles, before and after, from the sweep's capture
script. The row this unit is judged on:

Control	Height	Radius	Font size	Background
Inspector metric input, before	17px	0px	11px	transparent
Inspector metric input, after	28px	6px	11px	rgb(22, 22, 24)

28px and 6px are what both Export buttons already measure, so the inspector field
now sits inside the same three-height, three-radius set the sweep is aiming at.
Every other row in the table is identical before and after.

Bundle, gzipped across the app's chunks: 1,571,013 bytes before, 1,578,200 after.
A delta of 7,187 bytes (about 7.0 KB) for the first inspector consumers of
Slider and Toggle, against the 100 KB budget for the set. Reported, not
trimmed.

Not covered

• FlatSlider and FlatSelectRow keep their hand-rolled implementations.
  Both are large pointer-capture and native-select state machines with about
  thirty tests each pinned to their exact DOM, and converting them would have
  taken this PR well past a reviewable size. They are the next sweep PR's, and
  the flat inspector therefore still shows one underlined row (Weight) beside the
  boxed ones until then.
• The remaining section families are untouched: effects, colour and colour
  grading, canvas menu, media, motion and overlays keep their own colours and
  controls. Those are the later inspector PRs.
• Four colour literals in files this PR touches stay. One is an
  rgb(0, 0, 0) CSS fallback for a computed style, which is a browser default
  rather than a design value. The other three are shadow presets that have no
  matching token yet; adding tokens belongs to the unit that owns the token file.
• A pre-existing flake in the screenshot capture, not fixed here. The canvas
  motion-path toggle appears in some captures and not others, on the base branch
  as well as this one (present in one of two base runs, absent in two runs here).
  Its visibility is gated on an async runtime probe of the preview iframe, so it
  races the screenshot. Anyone diffing the inspector screenshots should expect it
  and not read it as a regression.
• This PR is stacked. Its own diff is 583 changed lines; a line-count check
  measured against the trunk will report the whole stack instead.