are unable to cleanly restore the state.

a network port, we could have guessed.

+ Changed port to 51914 (rationale: it's 0xCACA; so okay, it's outside
      the registered ports area, but we'd never be accepted anyway).
    + Retry network sends in caca_get_event(), too.
    + Have a per-client buffer in case of network congestion. If that buffer
      gets full, just drop it and start again at the next frame.
    + Set the window title to "caca for the network".
    + Do not send the ANSI buffer's terminating \0.
    + Restore the SIGPIPE handler on close.
    + Set window size to 80x24 instead of 80x25.

Note to developers: use that chunk in your ~/.subversion/config:

[miscellany]
enable-auto-props = yes

[auto-props]
*.m = svn:keywords=Id
*.h = svn:keywords=Id
*.mm = svn:keywords=Id
*.c = svn:keywords=Id
*.cpp = svn:keywords=Id
*.pch = svn:keywords=Id
*.command = svn:keywords=Id
*.pl = svn:keywords=Id
*.pm = svn:keywords=Id
*.py = svn:keywords=Id
*.sh = svn:keywords=Id
*.txt = svn:keywords=Id
*.html = svn:keywords=Id
*.php = svn:keywords=Id

writes. Currently works very badly with more than one client.

network_display() so that new clients immediately get the contents
    of the canvas and do not need to wait for a caca_display() call.

cucul_putchar32() function but I'm not convinced it will be useful.

they can force a resize. We need to find a way to make those specific
    cucul - caca interactions unavailable to the clueless user.

80x25 upon initialisation.

* Print a small message at kernel boot time just to check that the kernel
    was properly loaded.

as multiboot kernels.

good to have something to start with.

by simply ignoring everything non ASCII :-)