run-tests: catch a Windows specific error when testing for a free socket I'm not sure why this only happens with py3, but this error code doesn't map to any of the 3 currently being handled, and kills `run-tests.py` before it can run any tests when it happens: OSError: [WinError 10013] An attempt was made to access a socket in a way forbidden by its access permissions The documentation[1] says this can happen if another process is bound to the address with exclusive access. This seems to keep it happy. [1] https://docs.microsoft.com/en-us/windows/win32/winsock/windows-sockets-error-codes-2 Differential Revision: https://phab.mercurial-scm.org/D9816

run-tests: work around the Windows firewall popup for server processes Windows doesn't have a `python3` executable, so cc0b332ab9fc attempted to work around the issue by copying the current python to `python3.exe`. That put it in `_tmpbindir` because of failures in `test-run-tests.t` when using `_bindir`, which looked like a process was trying to open it to write out a copy while it was in use. (Interestingly, I couldn't reproduce this running the test by itself in a loop for a couple of hours, but it happens constantly when running all tests.) The problem with using `_tmpbindir` is that it is the randomly generated path for the test run, and instead of Windows Firewall remembering the executable signature or image hash when allowing the process to open a server port, it apparently remembers the image path. That means every run will trigger a popup to allow it, which is bad for firing off a test run and walking away. I tried to symlink to the python executable, but that currently requires admin priviledges[1]. This will prompt the first time if the underlying python binary has never opened a server port, but appears to avoid it on subsequent runs. [1] https://bugs.python.org/issue40687 Differential Revision: https://phab.mercurial-scm.org/D9815

hghave: split apart testing for the curses module and `tic` executable ef771d329961 skipped the check for the `tic` executable, because the curses module alone on Windows is enough to pass the `test-*-curses.t` tests. However, `test-status-color.t` uses this same check and explicitly invoked the executable, which fails on Windows. From the cursory searching I did, curses on unix requires `tic`, which I assume is why they were tied together in the first place. So this continues to require both to get past the curses guards on non Windows platforms. Differential Revision: https://phab.mercurial-scm.org/D9814

sqlitestore: disable test with chg There are known issues with transactions not being closed in a timely fashion, making the test flakey. Differential Revision: https://phab.mercurial-scm.org/D9821

heptapod-ci: allow testing with docker image other than :latest The project have a default config of latest for this variable. Differential Revision: https://phab.mercurial-scm.org/D9792

persistent-nodemap: also exchange the nodemap data over the wire It appears that 2 bytes are never equals to 3 bytes. Differential Revision: https://phab.mercurial-scm.org/D9810