Say that the runner's disk is the tighter budget

windows.md §7 weighed the Windows leg against the desktop leg in CPU
time only. What actually stopped a release was disk: the one 99 GB disk
is shared with the container images, v0.18.0's desktop run ended at
92 GB used, and v0.18.1's release build ran out, which is why that tag
has no release page. §7 now gives those figures, says what the desktop
leg deletes before its release build since 209ae361 and why that costs
the cache nothing, and names a second target's cache as the first thing
to check if a leg runs out again.

docs/dev/windows.md

@@ -411,6 +411,16 @@ tract, Slint's compiler, wgpu — so with the cargo cache warm it is minutes and
 better part of half an hour. Worth noting because the runner is one machine and the legs run in
 parallel on it; if it starts starving the desktop leg, `needs: desktop` serialises them.
 
+**Disk is the tighter budget.** The runner has one 99 GB disk shared with its container images. At
+rest it holds about 23 GB; the desktop leg's restored target cache, the models and the dependency
+build bring it to about 78 GB before a test runs, and v0.18.0's run ended at 92 GB used. v0.18.1's
+release build then died with "No space left on device", so that tag has no release page. Since
+then the desktop leg deletes its test executables, `target/debug/examples` and
+`target/debug/incremental` after the Test step and before the release build — they are relinked
+whenever a source changes, and the cache exists for the dependency rlibs — and prints the disk
+again beside its Disk before and after lines. A second target's cache on the same disk is the
+first thing to look at if a leg runs out again.
+
 ---
 
 ## 8. Requirements