Take the machine out of the comments

The draw-cost notes described the machine they were measured on, which is not
something this repository can check or keep true. What is left is what the
file measures and how to read it; the rest lives with the machine.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/default/render.rs

@@ -24,8 +24,6 @@ impl UiRenderer {
     pub fn draw(&mut self) {
         let output = match self.surface.get_current_texture() {
             CurrentSurfaceTexture::Success(texture) => texture,
-            // Still drawable; the surface has just changed under us, and
-            // reconfiguring is what the next frame wants rather than this one.
             CurrentSurfaceTexture::Suboptimal(texture) => {
                 self.surface.configure(&self.device, &self.config);
                 texture
@@ -34,7 +32,6 @@ impl UiRenderer {
                 self.surface.configure(&self.device, &self.config);
                 return;
             }
-            // Nothing to draw into this frame.
             CurrentSurfaceTexture::Timeout
             | CurrentSurfaceTexture::Occluded
             | CurrentSurfaceTexture::Validation => return,
@@ -81,8 +78,6 @@ impl UiRenderer {
     pub fn new(window: Arc<Window>) -> Self {
         let size = window.inner_size();
 
-        // The display handle is what GLES needs to present on Wayland, and it
-        // has to be the one the surface is made from.
         let instance = Instance::new(InstanceDescriptor {
             backends: Backends::PRIMARY,
             display: Some(Box::new(window.clone())),

tests/draw_cost.rs

@@ -4,20 +4,19 @@
 //!
 //!     cargo test --release --test draw_cost -- --ignored --nocapture
 //!
-//! **Read the instruction count, not the clock.** Wall time here swings by 2x
+//! Wall time is the wrong number to read for anything under a few percent --
-//! between runs of one binary on this machine -- more under `cargo test` than
+//! it varied by 2x between runs of one unchanged binary where instructions
-//! run directly -- while instructions retired are stable to 0.1%:
+//! retired varied by 0.1%. Count those instead:
 //!
 //!     perf stat -e instructions:u target/release/.../draw_cost-* --ignored
 //!
-//! Measured that way on 2026-09-13, drawing each primitive through its own
+//! That is how `PrimitiveRender` was measured against a match in the renderer:
-//! `PrimitiveRender` rather than a match in the renderer costs **6
+//! 6 instructions per list drawn, against the ~5,400 wgpu spends recording
-//! instructions per list drawn**, which is 0.1% of a frame at both 256 and
+//! one.
-//! 1024 layers. Recording one list into the pass costs wgpu ~5,400.
 //!
-//! The instance is leaked on purpose. Dropping the last one makes the Vulkan
+//! The instance is leaked deliberately. A Vulkan loader may unload the driver
-//! loader unload Mesa's ICD, which faults when a thread that touched Vulkan
+//! when the last one drops, which can fault as a thread that used it exits --
-//! exits -- and libtest runs every test on a spawned thread.
+//! and every test runs on a spawned thread.
 
 use std::time::Instant;
 
@@ -30,13 +29,13 @@ use wgpu::{Color as GpuColor, *};
 
 const SIZE: u32 = 1024;
 const FRAMES: u32 = 200;
-/// Reported as the best of this many batches. The mean moves by 15% between
+/// Reported as the best of this many batches, since the mean moves by more
-/// runs on this machine, which is more than the thing being measured.
+/// than the thing being measured.
 const BATCHES: u32 = 8;
 
 fn gpu() -> Option<(Device, Queue)> {
-    // Probed rather than assumed: this machine's Vulkan device comes and goes,
+    // Probed rather than assumed: there may be no Vulkan adapter, and GL is
-    // and GL is what is left when it is gone.
+    // what is left when there is not.
     let all = Instance::new(InstanceDescriptor::new_without_display_handle());
     let instance = match pollster::block_on(all.request_adapter(&RequestAdapterOptions::default()))
     {