cockpit-ci: Update container to 2026-05-16; Use cockpit.location.go() for URL-based container switching

[github-actions]

• test: Fix Browser.go() for recent Firefox cockpit#23283
• Pull in that new testlib and fix the fallout: Makefile: Update Cockpit lib to 8ddf198aeeb5df9373b00be359724492 #2545

[martinpitt]

The firefox failure reproduces effortlessly with

TEST_BROWSER=firefox test/check-application TestApplication.testBasicUser -stv $RUNC

• I don't get the "AllowWebgl2:false restricts context creation on this system" error message locally, that's apparently a red herring and only happens in CI.
• The "TypeError: WeakMap key null must be an object or an unregistered symbol" is also a red herring, it happens way before the time of error.
• There are no browser errors with the failing command (b.click(busybox_sel + " td.pf-v6-c-table__toggle button")
• The screenshot is misleading -- it's the cleanup that happens to kill the session, at the time of failure I still see the cockpit UI.
• At the time of the failure, the "Images" card is not expanded. The test expects that. When I expand it manually (in TEST_SHOW_BROWSER=1), the test continues normally and succeeds. It is expanded when running with Chromium.

The critical step is

        b.go("#/?container=all")

in line 842. With the new firefox this triggers a page reload, which resets the expanded state of "Images".

I'll push a fix, but looking at the other tests still.

[martinpitt]

This affects more tests. Editing the URL in the interactive browser does not reload the page. Our ph_go() implementation sets window.location.hash which somehow triggers this.

I noticed that we call ph_go() in frame context, which strikes me as odd:

info: XXX ph_go #/?owner=0 setting window.location.hash {"href":"http://127.0.0.2:9091/cockpit/@localhost/podman/index.html#/","origin":"http://127.0.0.2:9091","protocol":"http:","host":"127.0.0.2:9091","hostname":"127.0.0.2","port":"9091","pathname":"/cockpit/@localhost/podman/index.html","search":"","hash":"#/","ancestorOrigins":{"0":"http://127.0.0.2:9091"}}

This isn't something that a user could ever do interactively, and is cheating. Instead of fixing/replacing b.go() calls, I tried changing go() to change the actual top-level URL:

    def go(self, url_hash: str) -> None:
        with self.driver.restore_context():
            self.switch_to_top()
            self.call_js_func('ph_go', url_hash)

This passes test/check-application TestApplication.testBasicUser $RUNC for both Firefox and Chromium, but still fails testBasicSystem, as the top-level URL change does not propagate to the frame properly.

This phenomenon also explains this kind of failure in cockpit-project/cockpit#23277 -- the unexpected page load causes the "ph_texts not defined" error, as at that point we don't wait for a page load.

So let's instead try to fix Browser.go() properly. I have a local implementation which fixes both the podman and the cockpit tests.

[martinpitt]

Still several test failures now. But we really need to fix these screenshots, they are absolutely useless when they happen after cleanup.

Update: fixed in #2547

[martinpitt]

The TestApplication.testRunImageUser failure reproduces locally, and it's 100% stable. Once again, clicking on the container link in the images table with firefox reloads the page. The browser bug is well known by now, it updates the URL to add an anchor /podman#123abc). This even happens with an interactive click in the browser, so this time it's not a test issue any more.