window minimize qualification
The compositor implements native normal and dialog minimize/restore using GNOME Shell 51's 400 ms ease-out-expo curve, motion policy, slowdown, monitor-sized fade, leading-corner fallback and GTK-published dash icon bounds. Full restore stays opaque. Opacity follows Clutter's integer truncation.
An attached transient inherits logical hiding from minimized ancestors without changing its own flag. Explicit activation restores ancestors. Independently minimized children remain minimized when the parent is restored. Hidden child map/close does not restore its parent. Managed modal children continue to dim their parent even when independently minimized. This follows Shell 51 _checkDimming, which uses has_attached_dialogs(), and Mutter 51 has_attached_foreach_func, which checks attached and not unmanaging without checking visibility or minimize state. The native proof includes restored-parent brightness with an independently minimized child, then real unset_modal and set_modal requests as brightness controls; qualification of these new checks remains pending. Focus history is tracked separately from family stacking: a detached modal that closes falls back to a visible MRU window, without using its former parent.
Minimize renders current native surface buffers and cancels earlier map/size-change effects. A surface destroyed during minimize is copied to an independent texture and finishes the original clock. Retained textures share the existing 64 MiB / 32-effect budget; overflow or copy failure is diagnosed and skips retention. Snapshot rendering explicitly samples the full texture when scaling its destination. A native surface destroyed during restore retains its last rendered geometry and baked opacity before starting the existing destroy curve. The interrupted-restore proof compares the frozen live and retained pixels and rejects a full-size jump. This qualifies presentation continuity only; GNOME interruption curves and event timing still require comparison.
Reproducible checks
dbus-run-session -- /usr/bin/python3 scripts/tuna-minimize-proof --bundle-root <outside-checkout-directory> consumes one source-identified frozen/ directory directly and locks a free X display unless --display names an unused one. It contains the exact Cargo-JSON-selected compositor, GTK shell and native proof client executables, SOURCE_COMMIT, Cargo.lock and manifest.json; neither direction copies, links, strips or rebuilds those executables. LTR evidence belongs only to <bundle-root>/ltr/; --rtl selects <bundle-root>/rtl/ and GTK's inverted text direction. Each mode requires a fresh directory and owns separate runtime, settings, accessibility buses, captures and receipts. Reusing a mode directory fails rather than overwriting evidence.
CI uses scripts/tuna-minimize-ci-proof <outside-checkout-directory> with a previously nonexistent bundle root to build once from clean committed source and freeze all three executables, Cargo.lock and SHA-256 hashes. The consumer rejects missing/nonregular/nonexecutable roles, altered hashes, dirty or different source, and a mismatched lockfile before launching. After the unchanged live pixel checks and verified owned cleanup, it validates the entire freeze again against its original snapshot, including manifest and source-marker bytes; only then can it write PASS assertions. The existing wellbeing proof still consumes frozen/ and the same build.json receipt. Failure logs, captures and partial manifests remain in each mode, with the separate bounded failure-only diagnostic subset unchanged.
For authorized exact-binary replay, extract the full artifact outside the checkout, check out its frozen/SOURCE_COMMIT with a clean matching lockfile, and use --bundle-root with a private unused display under dbus-run-session. Preserve any existing mode directory elsewhere before deliberately replaying that mode; the CLI will not erase it. Keep frozen/ intact: it is the sole retained executable set and its original manifest must validate without rewriting provenance. Replay still requires the documented graphical dependencies and does not qualify other revisions or installed hardware.
Consumer regression coverage exercises missing, changed and mismatched bundles, source/lock binding, post-cleanup mutation (including a coherently rewritten role manifest), and LTR/RTL state isolation. A controlled executable CLI smoke also admitted a valid fixture freeze, rejected stale mode directories and a mutated role before launch, and refused the removed CLI options; it is not native pixel qualification.
The retained original artifact from CI run 41214937117, source 70e723617eb17e47aa68afd2a391dc84cb0d3385, contains three role sets with SHA-256 hashes matching its canonical frozen manifest (stream-verified for all nine executable members). One role set is 899,596,608 uncompressed bytes. Removing only the six redundant LTR/RTL executable members projects uncompressed member bytes from 2,704,436,996 to 905,243,780, and compressed member bytes from 582,801,450 to 195,117,580. These are original-ZIP metadata subtraction figures, not a newly generated ZIP size or an upload-time measurement. All unique evidence and all three exact executables remain; the smaller byte volume alone does not establish the cause or resolution of the original roughly 60-minute upload. Exact-head CI must still exercise the unchanged native pixel/cleanup gates and observe actual transport timing.
The harness checks actual pixels for fallback, dash, restore, reduced/Off, fullscreen, interrupted-map reset, live SHM commits, safe destroyed-surface retention, attached-parent hiding and independent child flags. These checks are prepared; they have not yet passed on this source revision. Cargo coverage includes clutter_guint_opacity_truncates_quarter_expo_samples, live_minimize_frame_uses_latest_buffer_geometry_and_owns_no_snapshot, shrinking_snapshot_samples_full_texture_instead_of_destination_crop, the transient graph tests and three native protocol family regressions. The unchanged native regression native_modal_unset_parent_immediately_before_close_has_no_stale_focus_return guards ancestor/MRU focus fallback and actual keyboard delivery; the new focus-history correction awaits CI qualification.
Qualification remains partial: this is not a paired GNOME benchmark. Installed-session, native GPU, multi-output/fractional-scale, X11, shadows, buffer resize during minimize, and event ordering within one dispatch batch require runtime comparison. No GNOME performance or complete v1 parity claim is made.
Primary references: GNOME Shell 51 window manager, Mutter window actor, Mutter workspace focus/MRU, Mutter plugin manager, Clutter interval opacity.