ci: give the race job a ceiling that matches the work it now does

Raises the race detector's limits from 25m under a 30m job to 40m under 45m,
which is the owner's call and the other half of the change beside it.

Not because a run went red. Because the margin had already gone and the reds
were the symptom. Measured on the runner across four consecutive runs: 21m35s
on main before JPEG XL, a timeout on the JPEG XL branch, 20m15s on main after
it merged, and a timeout on the dependency bump. Two out of four, and no data
race reported in any of them - a limit that decides on how busy the runner
happens to be is telling you about the runner, not about the code.

21m35s against 25m was never a margin, and that predates JPEG XL. What this
package holds today is twenty four formats, two of them running a borrowed
encoder, twenty five screens and two binaries, and the 25m was chosen when it
held less.

The gap between the two numbers is kept on purpose. Go's limit stays below the
job's so a slow run fails as a test, with the list of what was still running,
rather than as a killed job with no output - which is the reason the number was
stated here in the first place, on 2026-08-25.

The cheap half of the answer is the commit before this one: the JPEG XL ladder
sweep stopped spending the budget on arithmetic that finds no races, 102s to
54s under -race. This half stops the work being shaved to fit a ceiling nobody
had remeasured.

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

Author: donislawdev

.github/workflows/ci.yml

@@ -625,7 +625,9 @@ jobs:
       github.event_name == 'workflow_dispatch' ||
       needs.touched.outputs.concurrency == 'true'
     runs-on: ubuntu-latest
-    timeout-minutes: 30
+    # Raised from 30 on 2026-08-31, with the Go timeout below, and the reason is
+    # measured rather than "it went red". See that step for the numbers.
+    timeout-minutes: 45
     env:
       # The one thing in this project that needs a C toolchain. Linux runners
       # ship one, so this job carries the cost and the matrix above stays on
@@ -678,8 +680,27 @@ jobs:
         #
         # Race instrumentation costs five to twenty times the wall clock, and
         # this package renders twenty five screens and builds two binaries. The
-        # job's own thirty minutes is the ceiling that means something.
-        run: go test -tags "$(cat .github/build-tags)" ./... -count=1 -race -timeout 25m
+        # job's own ceiling is the one that means something, and this number
+        # stays below it so a slow run fails as a test rather than as a killed
+        # job with no output.
+        #
+        # Both were raised on 2026-08-31, from 25m under a 30m job. Not because
+        # a run went red, but because the margin had already gone and the reds
+        # were the symptom. Measured on the runner across four consecutive runs:
+        # 21m35s on main before JPEG XL, a timeout on the JPEG XL branch, 20m15s
+        # on main after it merged, and a timeout on the dependency bump. Two out
+        # of four, with no data race reported in any of them - a limit that
+        # decides on how busy the runner is tells you nothing about the code.
+        #
+        # 21m35s against 25m was never a margin, and that predates JPEG XL. What
+        # this package holds now is twenty four formats, two of them running a
+        # borrowed encoder, twenty five screens and two binaries, and the 25m
+        # was chosen when it held less. The cheap half of the answer is in
+        # internal/guard/jxlladder_test.go, which stopped spending the budget on
+        # arithmetic that finds no races - 102s to 54s under -race. This is the
+        # other half: the ceiling now matches the work rather than the work
+        # being shaved to fit a ceiling nobody remeasured.
+        run: go test -tags "$(cat .github/build-tags)" ./... -count=1 -race -timeout 40m
 
   coverage:
     name: coverage gate