[ymote]

Adopts OctoSense-org/makepad#116 (10a29aea), the Linux Vulkan follow-up to #87. It changes nothing in the default GL build; it affects Rinx built with MAKEPAD=vulkan on Wayland.

• Windowed Vulkan swapchains prefer mailbox. With FIFO, the Wayland loop's own frame-callback pacing and the driver's wait for the previous image to be presented ran in series, so every frame cost two vblanks: 275 presents under a sustained 8 ms wheel scroll where GL presents 502 (p50 gap 16.7 ms against 9.0 ms). Mailbox leaves pacing to the loop, as GL's swap interval of zero does; the compositor still composes without tearing. MAKEPAD_VULKAN_PRESENT_MODE=fifo|mailbox|immediate overrides.
• The Vulkan window path now reports draw/wait channels to the PerfMonitor (so native_performance.py shows them), times window passes on the GPU under MAKEPAD_VULKAN_PROFILE=1, and summarises presents and busy fence/acquire attempts under MAKEPAD_TRACE=gpu.present.

Validation on Hyprland 0.56 / Intel Arc 140V, 2880x1800@120, release builds with palpo-instrument ux_fixtures, --assert-idle:

• GL (the default build) on this pin: login/home and chat scenarios pass the idle assertions; cargo test --release --locked --features agent_chat --lib passes.
• Vulkan on the same Makepad commit: sustained scroll 502 presents at p50 9.1 ms (GL 502 at 9.0 ms; FIFO override 273 at 16.7 ms); idle, home, account-menu, Moments and typing scenarios pass the idle assertions with no busy attempts; English/Chinese typing preserved. Per frame the Vulkan main thread costs 2.15 ms against GL's 1.78 (encode 0.40 vs 0.19 ms, a Makepad follow-up) with fewer process wakeups (1552 vs 3859).

🤖 Generated with Claude Code