CI: fix both failures from the Test workflow's first run

Wiring the suites into CI immediately found two environment
assumptions that a single dev box had been hiding.

glxinfo is in mesa-utils, not x11-utils, so it was simply absent on the
runner -- and hasGL() raised FileNotFoundError and took the whole apps
suite down before a single app started. It is a diagnostic, not a
requirement, so a missing tool now reports "unknown" instead of
crashing, and mesa-utils is installed. Also dropped the "|| true" that
was hiding apt failures in that step, which is how a missing package
got this far unnoticed.

broadwayd binds its port before it can serve clients, and the suite
treated "port bound" as ready. On a slower machine that gap matters:
the first two codes alphabetically (Amica, DirDyVTST) failed with
"Unable to access the X Display" and every later one passed, which is
exactly the shape of a startup race. It now probes with a real wx
client and retries for up to 30s.

Neither was reproducible locally, which is the point of running these
somewhere other than the machine they were written on.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EFMUNbw2UpRPmgsqviAowb

Author: andy

.github/workflows/test.yml

@@ -82,13 +82,14 @@ jobs:
           # libgl1-mesa-dri: llvmpipe, the software renderer that actually
           #   answers those GLX calls.  Without it the apps get no GL and
           #   the suite says so on startup.
-          # x11-utils: xdpyinfo/xwininfo/glxinfo, used to see whether a
-          #   window actually appeared.
+          # x11-utils: xdpyinfo/xwininfo, used to see whether a window
+          #   actually appeared. glxinfo is in mesa-utils, separately.
+          # mesa-utils: glxinfo, only to report whether GL is working.
           # python3-wxgtk4.0: pulled in because the installed package's
           #   codereg dialogs need it at runtime.
           sudo apt-get install -y \
             apache2 apache2-utils xvfb libgl1-mesa-dri x11-utils \
-            activemq python3-wxgtk4.0 || true
+            mesa-utils activemq python3-wxgtk4.0
 
       - name: Build
         run: |

tests/apps/run_tests.py

@@ -251,9 +251,13 @@ def main():
     before = apps.serviceState()
     results = Results()
     try:
+        gl = display.hasGL()
         print("display: %s (GL: %s)"
-              % (display.name, "yes" if display.hasGL() else "NO -- the "
-                 "viewer apps will not render"))
+              % (display.name,
+                 "yes" if gl else
+                 ("unknown -- glxinfo not installed (mesa-utils)"
+                  if gl is None else
+                  "NO -- the viewer apps will not render")))
         serviceLog = []
         apps.startServices(display, serviceLog)
         for line in serviceLog:

tests/apps/xdisplay.py

@@ -124,9 +124,20 @@ def windows(self):
         return found
 
     def hasGL(self):
-        result = subprocess.run(["glxinfo", "-display", self.name, "-B"],
-                                stdout=subprocess.PIPE,
-                                stderr=subprocess.DEVNULL)
+        """True when GLX is answering. Unknown counts as "probably fine".
+
+        glxinfo lives in mesa-utils, which is not installed everywhere, and
+        this is a diagnostic rather than a requirement -- the apps
+        themselves will fail loudly if they really cannot get a context. A
+        missing tool must not take the whole suite down, which is exactly
+        what it did on its first CI run.
+        """
+        try:
+            result = subprocess.run(["glxinfo", "-display", self.name, "-B"],
+                                    stdout=subprocess.PIPE,
+                                    stderr=subprocess.DEVNULL)
+        except (FileNotFoundError, OSError):
+            return None
         return b"direct rendering: Yes" in result.stdout
 
 

tests/dialogs/harness.py

@@ -137,6 +137,35 @@ def _startBroadway(self):
             "BROADWAY_DISPLAY": ":%d" % number,
             "DISPLAY": "",
         }
+        self._waitUntilUsable()
+
+    def _waitUntilUsable(self):
+        """Wait until a GTK client can actually connect, not just until the
+        port is bound.
+
+        broadwayd binds its port before it is ready to serve, and on a slow
+        machine the gap is big enough to matter: the first CI run of this
+        suite failed the first two codes alphabetically with "Unable to
+        access the X Display" and then succeeded for all the rest. Probing
+        with a real wx client is the only check that means anything here.
+        """
+        environment = dict(os.environ)
+        environment.update(self.env)
+        probe = "import wx; wx.App(); raise SystemExit(0)"
+        deadline = time.time() + 30
+        lastError = ""
+        while time.time() < deadline:
+            result = subprocess.run([sys.executable, "-c", probe],
+                                    env=environment,
+                                    stdout=subprocess.DEVNULL,
+                                    stderr=subprocess.PIPE)
+            if result.returncode == 0:
+                return
+            lastError = result.stderr.decode("utf-8", "replace").strip()
+            time.sleep(0.5)
+        raise HarnessUnavailable(
+            "broadwayd is listening but no wx client can connect: %s"
+            % lastError[-300:])
 
     def __exit__(self, *exc):
         if self._proc is not None: