Fix Settings switch geometry and stabilize Model Pad chrome - #101

ℹ️ No critical issues — one robustness suggestion inline. The geometry refactor reads coherently: the fixed chrome reservation and grid-only observation remove the label-wrap feedback loop, the measure math is consistent with the CSS caps, the source-reading contracts match the actual CSS/TSX, and all 98 focused unit tests pass.

Reviewed changes

• Model Pad measurement — measureModelPadAvailableSize extracted into renderer/lib/model-pad-layout.ts with exported MIN_SIZE/CHROME/GUTTER constants; settings layout effect now observes only grid + scrollport and reserves 128px of fixed chrome instead of measuring live labels/legend, so legend wrap and marker labels can no longer resize the square outline.
• Model Pad CSS — .model-pad gains aspect-ratio: 1/1, flex-shrink: 0, align-self: start, min-width: 0; .model-pad-legend reserves min-height: 2.5rem; square is additionally capped to the visible scrollport minus 1px.
• Composer picker details card — ModelHoverDetails is now a fixed-height (min(22.5rem, 70vh)) column with an internal scroll area and pointer-events-auto, so popover height no longer shifts while model copy or catalog text changes.
• Switch geometry — shared Switch root is inline-flex items-center shrink-0 overflow-visible with a pointer-events-none thumb; settings Field rows carry has-[[role=switch]] trailing-switch columns; group/appearance cards drop overflow: hidden so the Plugins-style thumb is not clipped.
• Trailing alignment — items-center on git-push and shortcut rows, removed max-[540px]:justify-start in remote-access settings, new narrow-container :has([role="switch"]) appearance rules.
• Docs & tests — settings-design-system.md row updated to the reserved-chrome contract; unit tests for the measure function, settings source/CSS contracts, memory-settings copy placement, and picker chrome; E2E waits for in-flight view-transition animations before measuring.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

✅ No new issues found. The previously flagged unbounded animation wait is now genuinely bounded: infinite-iteration animations are excluded via the finite-endTime filter, anything still pending is capped by the 500 ms race, and the infinite sentinel makes the old hang reproduce deterministically on regression rather than only when a spinner happens to be animating.

Reviewed changes

• Bounded Model Pad E2E wait — the pre-measure wait now snapshots document.getAnimations(), filters to animations whose computed endTime is finite (endTime: Infinity for infinite iterations, so perpetual spinners drop out), and races Promise.allSettled(finished) against a 500 ms cap — paused or never-ending animations can no longer stall the merge-gate spec until the Playwright action timeout.
• Infinite-animation sentinel — a 1px, opacity-0, infinite-rotation element is appended to the page for the whole test, so the original unbounded-wait failure fails loudly on regression; the sentinel also exercises the filter on every panel iteration.
• Papercut documented — .papercuts/troubleshooting.md records the animation.finished never-settles rule and keeps geometry polling authoritative for E2E readiness.

  ^{｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}