Fix KDE portal Freeze and Zoom capture on mixed-scale displays

[devmobasa]

Summary

Related to #398.

KDE captures the entire desktop at the highest output scale. Wayscriber expected a raster assembled from each monitor's native pixel density, causing valid screenshots to fail geometry validation.

• Share verified portal crop geometry between Freeze and Zoom, and resample the active monitor's crop to its native dimensions.
• Preserve output identity, raster bounds, layout generation, topology, and dimension checks.
• Retry once after a confirmed layout change, retaining the original request and repeating the GTK and main-surface transparency barriers.
• Wait for complete output metadata and 100 ms without another portal layout change before retry admission, bounded to one second.
• Log screenshot bounds, crop coordinates, output identity, layout generations, and retry phases.
• Document the Plasma/KWin screenshot-portal requirement in the support matrix.

Review follow-up

Seven focused commits address the accepted review findings:

• Check live Freeze topology before interpreting a worker crop error, including the interval before new_output refreshes geometry.
• Preserve buffer-release and frame-cap pacing instead of repeatedly waking dispatch with a zero timeout while no buffer is available.
• Separate portal desktop generations from active-output source generations, so another monitor's metadata change does not invalidate direct captures or installed screen-modal sources.
• Add bounded layout settling before the one allowed retry.
• Remove the obsolete activation-time desktop crop path and its cropped flag; validate and install the worker's completed crop.
• Split activation, direct capture resources, and retry coordination into focused modules; update scoped guides and logical blank-line grouping.
• Cover Freeze activation/cancel/output-switch/stable-error behavior, coordinator failure branches, retry deadlines, rotated crop pixels, and Zoom terminal categories.

Validation

• ./tools/lint-and-test.sh: passed at 8b124b76, including strict workspace Clippy and all-feature/no-default-feature builds and tests.
• Root unit harnesses: 5221 all-feature and 5037 no-default-feature tests passed; 10 intentionally ignored in each. Configurator: 511 / 510 passed.
• ./tools/test-gtk-widgets.sh: 84 parent-harness tests passed in private headless Weston, with both actual GTK execution markers.
• Follow-up KWin 6.7.5/KDE portal 6.7.5 acceptance: 64 Wayscriber portal requests passed with builtin and GTK toolbars. Covers mixed/equal/fractional scales, both initial monitors, successful retries, retry exhaustion, two-batch layout updates, and two-buffer capture with VSync on/off and a 10 FPS cap when off.
• Pixel observers and regressions verify the correct monitor region and resampling. Four regressions were demonstrated failing before repair and passing afterward.
• Follow-up diff and working-tree whitespace checks passed. The unrelated local docs/CONFIG.md edit is excluded.

The KWin lab binary was built immediately before a final whitespace-only amend; its code behavior matches the final head. The full local and GTK gates used the final head.

Limitations

The request stays bound to its original wl_output. Recreating that object remains terminal: connector name/make/model fields are not a guaranteed unique physical identity for safely redirecting a pending capture. The reporter's physical Samsung EDID behavior and specific GTK popover timeout were not reproduced. The 100 ms quiet period is bounded; a later update during the retry barrier remains terminal.

A separate GTK output-cycling failure and the restricted KWin ScreenShot2 backend remain outside this change. The lab pacing cases use normal compositor buffer releases; forced withholding of all buffers is covered by the wakeup policy regression and source guard, not a runtime CPU measurement. Portal chooser dismissal is controller-tested; lab acceptance uses a private consent store. No host desktop was used.