feat(studio): the inspector slider and select row on the shared primitives

[miguel-heygen]

Lands the second inspector sweep PR (U10b: the flat slider and select row) of the Studio design-system foundation. Stacked on #3774. Hex ratchet unchanged at 468 (neither file held a literal). Bundle −0.7 KB. Net −423 lines. Adds the R13 native-controls ratchet over components/editor so the remaining section families lower it to zero.

Lands the second inspector sweep PR (U10b) of the Studio design-system foundation. Stacked on the primitives sweep (U10a). The inspector's slider and select row lose their hand-rolled implementations, and the panel has no bare underlined field left.

What

The two controls U10a stopped at now come from the design system.

• FlatSlider drops a ~250-line pointer-capture state machine and wraps the
  shared Slider. Pointer capture, keyboard, ARIA and the thumb are Base UI's.
• FlatSelectRow drops its native <select> and wraps the shared Select, so
  its box is the same box Input, NumberField and the inspector's metric
  field wear, and its popup is Studio's menu surface rather than the operating
  system's.
• The Text section's "Weight" row was a copy of FlatSelectRow written inline
  (same native select, same underline, same caret). It is a FlatSelectRow now.
  It was the last bare underlined field in the panel.
• A ratcheted check fails when a new native <select> or hand-rolled
  type="range" appears under components/editor, and fails again when a
  listed file is converted and not removed from the list, so it only shrinks.

Also fixes a gap in the shared Slider this migration surfaced: it answered
right-click and Escape as drag aborts but had no handler for the platform's own
pointercancel (a scroll or touch takeover, a pen leaving range), so an aborted
gesture left whatever value the pointer last reached applied. The control being
replaced reverted on all three.

Why

R8 wants one implementation per control and no native <select> left. R10 wants
an inspector value to read as a field before it is clicked. The row that carried
"700 · Bold" did neither: it was an unstyled OS dropdown under a 1px line, beside
three siblings that already had a real box.

The slider is the R4 half. The hand-rolled one carried its own pointer capture,
its own lostpointercapture resync, its own aria-valuenow, and its own
keyboard map. All of that is behaviour a headless library already owns and gets
right, and none of it was shared with the other sliders in the app.

How

The wrapper keeps the write rate, and only that. The shared Slider splits
onPreview (every intermediate value) from onCommit (a boundary). This row has
one channel: a caller's onCommit writes the style, and that write is the live
canvas preview. So onPreview is throttled to at most one write per frame rather
than dropped, and the commit boundary flushes it. Dropping it instead would have
been the quiet regression in this PR: every test still passes, and the canvas
stops moving until you let go.

Telemetry keeps its contract: onTrack fires from the commit boundary only, so
one drag is one event no matter how many pixels it crossed.

The abort path goes through onPreview too. Base UI has no cancel API for a
running gesture, but it hands every change an eventDetails.cancel(), so an
aborted drag stops applying moves while the button is still down and the pre-drag
value is re-applied through the same channel the intermediate values used.

Tests. The ~30 tests pinned to the old DOM were sorted by what they assert.
Behaviour tests were re-expressed through the new DOM: the thumb's own
input[type="range"] for value, range and keyboard; [data-slider-control] for
pointer geometry; role="combobox" and role="option" for the select. Tests
that only described the old implementation were deleted, each named below.

Driving a listbox from a test is not a one-liner the way setting select.value
was, so the open-and-choose sequence lives in one helper (flatSelectHarness.ts)
that six section test files share, rather than being rewritten in each.

Test plan

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

• bunx vitest run --poolOptions.forks.maxForks=4 — 441 files, 4843 passed,
  18 todo. Baseline at the branch point: 440 files, 4845 passed, 18 todo. One
  file added (the duplicate-controls ratchet, 3 tests) and five deleted, listed
  below.
• 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 against the base branch — exit 0,
  "No issues in 82 changed files".
• Studio load smoke, the CI job's own script against a dev server on a port
  nobody was holding — PASS: studio loaded with schema-valid API fixtures and no runtime errors.

Five tests deleted, with the reason each is no longer a claim about this code:

Deleted	Why
dim knob at the correct position / filled track and bright knob	The thumb's position and the indicator's width are Base UI's. Replaced by two tests on the value text's tier tint and the thumb's aria-valuenow.
widens the click/drag hit area beyond the thin visible line	The 24px pointer target is the shared Slider's and lives with it.
ignores pointermove once a drag has ended	Pointer-capture bookkeeping the library owns.
still commits the release position when releasePointerCapture fires lostpointercapture synchronously	Describes a synchronisation between two handlers the wrapper no longer has.
resets dragging state / resyncs from the latest value on lostpointercapture	Same.

Three separate abort tests (right-click, Escape, pointercancel) became one
it.each over the three, so the executed count is unchanged there.

One test was rewritten because the swap made it vacuous rather than failing: the
style section asserted that two unset blur sliders drew no fill element, and that
element no longer exists on any slider, so the assertion would have passed
without checking anything. It asserts both blur thumbs read zero now.

Verified in a real browser, because a slider that renders as an invisible
zero-height track passes every test above. Studio booted from source on its own
port, the Style section opened, and the Layer blur slider dragged with a real
mouse: the track and thumb render at the expected geometry, and the canvas
headline blurs during the drag, which is the live preview the throttle exists
to keep.

That same run surfaced one thing worth flagging: the row's numeric readout stays
at its old value through a drag while the canvas updates. It is not a
regression from this change. The identical probe on the base branch produces
the identical result, because the readout has always come from the panel's
styles prop rather than the control's own draft. It is left alone here.

Screenshots, before and after, from the sweep's capture script. The computed
style table is identical in both runs — this PR moves no control the table
measures — and the difference is in the inspector shot: the "Weight" row goes
from bare text under a hairline to the same 28px/6px box as "Size", "Letter
spacing" and "Line height".

Colour literals in the ratchet: unchanged at 468. Neither migrated file had an
entry, so both were already at zero, and this PR does not add one.

Bundle, gzipped across the app's chunks: 1,595,377 bytes before, 1,594,630
after. A delta of -747 bytes. The two shared primitives were already in the
bundle from earlier units, so what this PR removes is larger than what it adds.

Not covered

• The remaining section families keep their own controls: effects, colour
  and colour grading, canvas menu, fx and block params. Ten files still carry a
  native <select> and five a hand-rolled type="range"; they are the ratchet's
  starting list and the later inspector PRs' work.
• The slider no longer says whether a property is explicitly set. The old
  one drew a filled track and a brighter knob for an explicit value. The shared
  Slider has one look. The row's value text still carries the tier tint and
  the reset button still appears only for an explicit value, so the signal moved
  rather than disappearing, but it is a deliberate reduction and not a
  like-for-like port.
• The select trigger's rendered box was not measured. The capture script's
  selector list covers the metric input, not the select, and adding a row for it
  needs a selector that does not also match the font-family combobox beside it.
  The trigger carries fieldBase and the tests assert that; the screenshot is
  the visual evidence.
• This PR is 1,611 changed lines (594 added, 1,017 deleted), over the 1k
  guidance, and was not split.
  Splitting at the slider/select seam does not get under it: the slider's own
  files are about 1,200 of those lines, because 1,017 of the total are deletions
  of the implementation being replaced. The two halves also share the six test
  files, so a slider-only PR would leave the panel with one bare underlined row
  and one migrated row, which is the "only makes sense next to its sibling" case
  a split is supposed to avoid.
• This PR is stacked. A line-count check measured against the trunk reports
  the whole stack instead of this branch's own diff.