feat(extract): the extraction window is an open sheet (#336)

[Maxaubert]

Closes #336.

The owner picked variant B, "Open sheet" (mockups: research/prism/mockups/2026-10-09-extract-dialog/round2/index.html, overview.png beside it). Brief: "clean, fits every theme, minimal text": the bar, the percentage, one live line naming the file being extracted right now (WinRAR style), and Cancel.

What changed

• ExtractWindow: no title, no caption, no destination line. The archive is the box's title and its aria-label ("Extracting X to D" / "Couldn't extract X").
• Top row, 40px fixed: the percentage at 34px weight 400 with a small % in --p-text-soft (blank until the engine speaks), Cancel at its end.
• The bar is a 12px recessed well: --p-control ground, --p-dim2 inset edge, --p-accent-solid fill, corners var(--p-radius). It is the widest and thickest thing in the window. The old pulse before the first number is gone (the well is simply empty, per the mockup).
• File line under it: UI face, tabular figures, 12.5px, fixed 18px, --p-dim. Fitted by a canvas measure in the MIDDLE so the member's own name is kept (fitPath, pure, unit-tested; replaces the character-count middleEllipsis). It is written in a layout effect and refits on resize or zoom. Main's existing 60ms progress cap (PROGRESS_EVERY_MS) keeps renders under one per frame, so no extra throttle was needed.
• Cancelling: the file line says "Cancelling" and Cancel dims (disabled, 50%).
• Failed: "Couldn't extract" (17px semibold, --p-text) takes the number's place, the well keeps the fill where it stopped, hatched in --p-dim (the reducer now keeps pct on failure), the file line is the reason alone (failureText no longer repeats "That couldn't be extracted."), and Close is --p-sel-bg with --p-on-accent.
• Padding 20px top and bottom, 22px at the sides, width 456px.
• Unchanged: cannot be dismissed (scrim, inert root, key guard), focus on the box at every phase, Cancel is a request that waits for the clean-up, success closes silently, the 700ms minimum showing, failure turns the same window into the error, keyed Cancel/Close.
• CLAUDE.md: a note under the extraction window rule. Version 0.100.0 (minor, feature; main is 0.99.1 and fix(explorer): coming back out of a zip shows it in the preview again (#334) #335 takes 0.99.2).

Tests

• npm run typecheck: clean. npm run lint: 0 errors (7 existing warnings). npm test: 2657 passed, 2 skipped.
• e2e (parked runner, one process at a time): extract (extract, extractWindow, extractCancel) all passed; zip =treeVerbs =drag (sevenZip, flatZip, zipMenus, zipWrites, zipProject, zipRestore, zipLocked, zipFolder, drag, treeVerbs) all passed.
• New e2e assertions: every route's window has no heading and no drawn archive name, its label names the archive and destination and its tooltip the archive; through a real 7-Zip run the box and the well are sampled every 25ms and never change size or move; the file line ticks through at least three members; the number is only ever a percentage, 34px at 400; the well is 12px and the widest thing, --p-control with an --p-accent-solid fill; padding 20/20; the file line is in the UI face; the cancelling state is caught at its commit (line says Cancelling, Cancel disabled at 0.5); the failure shows "Couldn't extract", the hatched stopped well, the reason alone, and Close on --p-sel-bg.
• Screenshots in .e2e/shots/: extract-sheet-running-aurora.png, extract-sheet-running-sand.png, extract-sheet-failed-{aurora,sand,jade,midnight-hc}.png, plus the full-window extract-window-running.png and extract-window-error.png. Compared with the mockup's B; they match.

Not machine-testable, for the hands-on pass: how the ticking file line reads zoomed in on a real large archive.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHHaWKR4M5QtW7Wecyuk4t