Rebuild the Studio as a single-voice Liquid Glass app

[MattModeCode]

The Studio worked, but it looked like a local dev tool: two columns, bundled
mono type everywhere, a flat dark background, and a voice picker you had to
answer every time despite there only ever being one voice that mattered.

This rebuilds the interior. The desktop shell, the engine, and all eight
/api/* routes are untouched — nothing here changes what Double Chin does,
only what it is like to use.

The interface

Single voice. The app opens straight into composing. The voice is resolved
at boot and never appears as a choice on the main surface; enrolment, switching
between enrolled voices, and the environment report all live in a Settings
sheet. The choice persists in localStorage, so it stops depending on
whichever voice happens to sort first. The toolbar shows the voice rather than
repeating the window title.

Progressive disclosure. At rest: a script editor and a Generate button.
Delivery controls and the prosody reference sit behind panes that summarize
their own state when closed — "Default", or "Expression 0.70 · fixed seed".

Apple's design system, not an impression of it. The first pass wore Liquid
Glass without earning it: every panel was translucent, and behind them sat an
animated four-pool radial gradient with a noise overlay — a synthetic backdrop
invented so the blur would have something to show. That is exactly the tell.
Apple's materials blur the app's own content; a gradient placed behind glass to
make it look like glass is decoration pretending to be a material.

So the ambient field, the grain and the drift animation are gone, and what is
left is Apple's grouped-content system: a recessed page (#F2F2F7 / #1C1C1E)
with opaque cards on top (#FFFFFF / #2C2C2E), divided by .5px separators
and carrying no shadow, because grouped cards in AppKit don't have one.

Glass now appears in the two places Apple puts it, and only there:

• the toolbar, which the page genuinely scrolls beneath
• the settings sheet, which sits over the page

Both blur real content.

Colour and type are semantic AppKit/UIKit values — label, secondaryLabel,
tertiaryLabel, separator, the fill ramp, controlAccentColor, systemGreen /
Orange / Red — at the macOS scale (13px body, 11px secondary, 15/17px titles).
Two departures, documented in the file: the secondary label alpha is raised
because Apple's own value measures 3.5:1 on white and fails WCAG AA, and the
accent has a defined pressed shade instead of a brightness filter.

Controls use AppKit metrics: 36px prominent button, 24px secondary, 8px
control radius, 12px card radius, the three-point accent focus halo, a
disclosure triangle that rotates 90°, and a slider whose track is a two-stop
fill. History is one inset grouped list with hairline dividers rather than
eight floating cards. Nothing pulses, glows or drifts.

I also tried real window vibrancy — pywebview 6.2.1 does give a genuine
NSVisualEffectView behind the WebView with transparent=True, vibrancy=True
— and did not ship it: it costs the window shadow, which is a bigger Mac
regression than desktop-sampling translucency is a win.

Details. ⌘↵ generates (scoped so it does not fire behind the Settings
sheet). Reduced motion, increased contrast with its own dark variant, visible
focus rings, labelled controls, aria-live on status regions, and no
horizontal scroll down to 520px.

Structure

static/ was three files and is now six: tokens.css (design system),
style.css (components), and ES modules api.js (endpoint wrapper),
format.js (pure presentation helpers), app.js (state and wiring). The
bundled JetBrains Mono files are deleted — system type for prose, ui-monospace
for numerics.

Desktop shell

The native window painted pywebview's white default before the page loaded — a
bright flash on the way into a dark interface. It now launches on the
interface's own background colour, read from the macOS appearance, and the
default size follows the content column (820×860). Since that colour lives in
two files, a test asserts desktop.py and tokens.css cannot drift apart.

Fixes found while building

• History rows rendered right-to-left: .score carried an explicit grid-row,
  which took it out of auto-flow and let it claim column 1.
• A single-chunk take sat at 0% progress for its whole run — the bar was showing
  "chunks finished", which is zero until the only chunk lands.
• <dialog> lost its centring to the global * { margin: 0 } reset.
• "1 chunks".
• Re-enrolling an existing voice always failed with a 409; the form had no
  overwrite field even though the API has always accepted one.
• Increased contrast whitened the glass without redefining --ink, so dark
  mode plus increased contrast was white-on-white. (Found by /code-review.)

Test plan

• pytest — 110 passed, 2 skipped (3 new tests for the launch colour)
• Browser at 1120px and 560px, light and dark, no console errors
• Real pywebview window in both system appearances
• Rebuilt dist/Double Chin.app — styling survives PyInstaller, and the
  whole static/ directory is already in the spec's datas, so no
  packaging change was needed
• End-to-end generation: idle → progress → verifying → result (0.855,
  strong match), plus replay from history
• Contrast measured against the shipped tokens

Not included

The README prose is updated, but demo/studio/double-chin-studio-demo.gif and
the linked session video still show the old interface. Re-recording them is a
separate pass.

https://claude.ai/code/session_01FdWprv4TKfVzapYFJnRffz