fix(linux_window): separate the surface scale from the stage factor in mutter's PHYSICAL layout

[dfattal]

Companion to DisplayXR/displayxr-runtime#1785 (no merge-order dependency: this PR builds against today's runtime headers, and the runtime PR does not touch the vendored u_wayland_geom.h).

Problem

mutter has two monitor layout modes. In LOGICAL (fractional scaling; mutter 50's default) the stage is logical px. One number, the monitor scale, is then both the surface scale and the stage factor. In PHYSICAL (Ubuntu 24.04 / GNOME 46 at an integer scale, out of the box) the stage already is device px. At 200 % the stage factor is 1 while clients still draw at 2.

The Wayland window used one number for both:

• Sizing. mode / xdg_output.logical_size is the stage factor (1 in PHYSICAL), and it was used as the surface scale. request_initial_rect() therefore declared a 720x400 rect as 1440x800 and held it. The deferred fullscreen's first configure guess was also twice the panel.
• Positions. The window-geometry extension's rects were converted by monitor.scale (2 in PHYSICAL). The initial rect's landing check and the drag-lattice map therefore used the wrong factor.

Fix

The new dxr_wl_scale.h (pure) keeps the two factors apart:

• dxr_wl_surface_scale_estimate() handles sizing. It returns the integer wl_output.scale when the stage factor is 1 (the PHYSICAL signature), else the stage factor.
• dxr_wl_stage_factor() converts published rects. It tries, in order:
  1. The extension's own answer (monitor.device_scale / layout_mode, extension v11 in runtime#1785).
  2. This client's output at the monitor's origin (mode / logical_size, right in both modes).
  3. The monitor scale.

The runtime's new u_wayland_layout.h is vendored verbatim in xrt_aux/util. get_own_geometry() parses the two new fields.

Verified

This ran on a private headless mutter 50.1 (3840x2160 virtual monitor, gdctl set --layout-mode physical --scale 2), using linux_window_test with DXR_LW_TEST_WAYLAND=1 DXR_LW_TEST_WAYLAND_PANEL=0,0,3840,2160:Meta-0.

	before	after
PHYSICAL 200 %, extension v11	declared 1440x800, wanted 720x400; landing read as (600, 400) 2880x1600	initial rect: LANDED — (300, 200) 720x400 device px, off by (+0, +0)
PHYSICAL 200 %, extension v10	—	LANDED, off by (+0, +0) (client-output fallback)
LOGICAL 200 %	LANDED	LANDED (unchanged)

The panel-fullscreen check (fullscreen once mapped) fails under mutter in both modes, before and after. It is written against weston and is not part of this change.

The existing lattice, drag and window tests pass. The new tests/linux_window_scale_test.cpp covers both modes with the measured inputs.

After merge

Tag it. When the runtime's test-app pin moves to that tag, u_wayland_layout.h can join the byte-identity guard in test_apps/common/dxr_linux_window_aux_guard.cmake.

🤖 Generated with Claude Code