window: don't allocate a swapchain for 0x0 windows, handle reconfigure/next failures

[kallegrens]

Problem

A Wayland compositor may send wp_fractional_scale_v1.preferred_scale before the first xdg configure, i.e. while the window's logical size is still 0x0. On such an event, CWaylandWindow's scale handler calls onScaleUpdate() → configure(m_waylandState.logicalSize) → resizeSwapchain({0,0}), which:

1. calls CSwapchain::reconfigure() with .length = 2 and a zero size — under current aquamarine this clears the swapchain but leaves options.length = 2 with zero buffers,
2. ignores the reconfigure() return value, and
3. unconditionally feeds next() results into CWaylandBuffer, which would also dereference a null buffer at buffer->dmabuf() if next() ever returned nullptr.

The result is a guaranteed std::out_of_range crash on the first render attempt:

terminate called after throwing an instance of 'std::out_of_range'
  what():  vector::_M_range_check: __n (which is 1) >= this->size() (which is 0)

#8  Aquamarine::CSwapchain::next(int*)
#9  Hyprtoolkit::IWaylandWindow::resizeSwapchain(Hyprutils::Math::Vector2D const&)
#10 Hyprtoolkit::IWaylandWindow::configure(Hyprutils::Math::Vector2D const&, unsigned int)
#11 Hyprtoolkit::_CWpFractionalScaleV1PreferredScale(void*, void*, unsigned int)

Reproduced 100% of the time with hyprpolkitagent under niri on a fractional-scaled (150%) output.

Fix

In IWaylandWindow::resizeSwapchain():

• return early when the pixel size is zero — there is nothing to allocate yet; the subsequent xdg configure with the real size goes through the full path as before,
• check the reconfigure() return value and bail out on failure instead of proceeding with a stale swapchain,
• check next() for nullptr before constructing a CWaylandBuffer (which would null-deref otherwise).

After this, the previously guaranteed crash under niri + 150% fractional scaling is gone and the window renders/authenticates correctly.

Note: there is also an aquamarine-side inconsistency (the swapchain clearing branch leaves options.length > 0 with zero buffers), which I've fixed separately in hyprwm/aquamarine — see hyprwm/aquamarine#379. The two are independent: either one alone stops the crash, but both should hold the invariant.

[github-actions]

Hi @kallegrens,

Thanks for your contribution! Unfortunately, Hyprland and hypr* no longer
accept contributions from unvouched contributors.

To get vouched, please see our contributing guide at wiki.hypr.land/Contributing-and-Debugging

This PR has been closed automatically and will not be re-opened.