[ihamzatarar]

Fixes #1023.

Under a locale that cannot represent what was typed, IRB aborts on exit and the session's history is thrown away:

$ LANG=en_US.ASCII irb
irb(main):001> puts 'héllo'
héllo
irb(main):002> exit
.../lib/irb/history.rb:68:in 'IO#puts': U+00E9 from UTF-8 to US-ASCII (Encoding::UndefinedConversionError)

save_history opens the file with encoding: IRB.conf[:LC_MESSAGES]&.encoding, which resolves to US-ASCII here, so f.puts(hist) has to transcode and raises on the first character the locale has no room for. The two rescues below it cover Errno::ENOENT, Errno::EACCES and Errno::EPERM, so this one escapes. save_history is called from the ensure block in IRB.run, which is why the whole session goes rather than just the write.

l.scrub on the line above does not help: it repairs invalid bytes, and these bytes are perfectly valid UTF-8. There is simply no US-ASCII character to transcode them to.

This writes the entries in the encoding they already carry, so nothing has to be converted and nothing is lost. The history file is IRB's own file, so there is no external consumer whose expectations the byte encoding has to match.

Scope, since it is worth being explicit about what this does not touch. load_history has the mirror problem one line up, at File.open(history_file, "r:#{IRB.conf[:LC_MESSAGES].encoding}"): an existing UTF-8 history file read back under an ASCII locale comes through with its characters mangled. That path is covered today by test_history_different_encodings, which asserts the mangling, so changing it would mean deciding that the current behaviour is wrong as well. Reading is lossy but does not raise, and the crash and the data loss on exit come entirely from the write. I have kept this change to that and am happy to follow up on the read side separately if you would like it changed.

The new test puts a non-ASCII entry in the history with LC_MESSAGES set to en_US.ASCII and asserts the save does not raise and the entry survives. On master it fails with Encoding::UndefinedConversionError: U+00E9 from UTF-8 to US-ASCII.

The full suite goes from 402 tests and 2483 assertions to 403 and 2484, with the same single pre-existing failure either way (test_use_tracer_enabled_when_gem_is_available, which needs the tracer gem).