sigpipe-remote: simply delegate pipe forwarding to subprocess we can kill Instead of using sophisticated logics with thread a non blocking pipes, we simply spawn two new process in charge of reading the pipe and sending the result to the client. When it is time to cut the pipe we violently kill them without any remorse. This close the pipe regardless of any in progress `os.read` call. Ironically this is the very same things as what the initial shell setup was doing, but in Python. This makes the test pass run properly on Windows. This also reveal that the Windows behavior is broken as the transaction is not properly rollback. However this is an adventure for another time. Making the test behave properly was enough effort. Differential Revision: https://phab.mercurial-scm.org/D11087

sigpipe-remote: verify the script is IO are unbuffered We don't want to get stuck anywhere by buffers. Differential Revision: https://phab.mercurial-scm.org/D11086

sigpipe-remote: check environment variable sooner Lets make sure we have everything before doing anything. Differential Revision: https://phab.mercurial-scm.org/D11085

test-sigpipe: run the hook using `sh` This make the hook properly run on Windows. Differential Revision: https://phab.mercurial-scm.org/D11084

windows: add windows behavior on broken pager Apparently, Windows has "better" behavior than Unix in this case. This is an edge case that led me down a rabbit hole, only to find a bug in the Python documentation... I am not planning on trying to reproduce the same behavior on Unix systems since it's not really useful, but other people are welcome to! Differential Revision: https://phab.mercurial-scm.org/D11000

mergestate: use dirstate.update_file API in recordupdates() Let's use the new single function to replace multiple different functions. This is part of dirstate API refactor to make it more clearer and generic so that it can be replicated in Rust. Differential Revision: https://phab.mercurial-scm.org/D11077