Nightcap v0.3.1

Added

• Each release carries a DEB, an RPM, and a pacman package beside the
  AppImage, with a .sha256 for every file. They install /usr/bin/mocktail
  and conflict with each other and with upstream's mocktail package.
  Nightcap is in no distribution repository or the AUR; the recipes under
  packaging/aur/nightcap{,-bin,-git} build from this repository locally.
• The AppStream metadata lists the Nightcap releases, so software centres
  show the release history instead of upstream's.

Fixed

• ETC2 texture decoding no longer blocks the thread that submits the frame.
  Uploads decode on a background dispatcher and the submit carries a timeline
  semaphore wait, so the GPU waits for the texture instead of the render
  thread. Reading the application's compressed bytes moves to that dispatcher
  too whenever they are already mapped. In Blade Ball the worst
  vkQueueSubmit fell from 140ms to 1ms, its p99 from 4.9ms to 0.05ms, and
  the total time spent in submits from 2534ms to 309ms. Frame interval is
  unchanged: the remaining tail is the engine's own main-thread step. A device
  without the timelineSemaphore feature decodes inline as before, as does a
  submit whose pNext already sizes arrays by its semaphore counts.
• The mouse pointer jumped when a right-click camera drag ended, and drifted
  to the viewport edge during one. The input router inferred pointer capture
  from SDL reporting zero absolute coordinates, which holds on X11 but not on
  Wayland, where the reported position accumulates the relative deltas and
  leaves the window. Motion, button, and wheel events now carry the window's
  relative mouse mode. The right-click camera fallback also only captures
  inside an experience, so the cursor stays free over the app surface.
• Shutting the window down while an experience was starting or leaving could
  free the pointer capture owner underneath the thread using it. Callers off
  the SDL thread now hold a reference for the length of their call.
• The TextBox caret did not move when a trailing space was typed in a
  single-line field. The final cluster rect SDL_ttf reports stops at the last
  inked glyph, so the laid-out width is used there instead. Wrapped layouts
  keep the clamp, because their width is the widest line rather than the last.
• Every TextBox keystroke re-ran TTF_Init, a fontconfig sort, six
  TTF_OpenFont calls, and TTF_Quit on the render thread. The overlay holds
  its opened fonts and reopens them only when the Roblox font or the point
  size changes, keeping the current set when a reopen fails. Measured over 100
  rasters, 1.59ms to 0.12ms per keystroke.
• Staged ETC2 copies now carry an explicit host-write barrier. A queue
  submission only makes host writes from before it visible to the device, and
  an asynchronous decode writes after the submit, so without the barrier those
  texels were not guaranteed to reach the copy that reads them.

Changed

• docs/PERFORMANCE.md documents the GameMode settings to run with. A
  desiredgov=powersave in ~/.config/gamemode.ini overrides the packaged
  performance default and clocks the CPU down for the whole session while
  still reporting an active performance request.