Changelog

0.3.0

Codepagex.Backend: use Codepagex.Backend, encodings: [...] in a module of your own compiles those encodings into it, with the functions of Codepagex for them. A library can so have the encodings it needs without the application configuring Codepagex, and without recompiling it.
Aliases for the Windows code pages: :windows_874, :windows_932, :windows_936, :windows_949, :windows_950 and :windows_1250 to :windows_1258, for "VENDORS/MICSFT/WINDOWS/CP874" and so on. Like the other aliases, Codepagex has them only for the encodings it is configured with.

:ascii is now plain US-ASCII, where every byte 0..127 maps to the same codepoint. It used to be an alias for "VENDORS/MISC/US-ASCII-QUOTES", which maps ' and ` to the curly quotes U+2019 and U+2018. To keep the old behavior, add "VENDORS/MISC/US-ASCII-QUOTES" to :encodings and use that name.
to_string/4 and from_string/4 now return {:error, reason, acc} when the outer missing_fun returns {:error, reason}. Previously the 2-tuple was returned as it was, contradicting the typespec. Code that matched {:error, reason} on that result must be updated.
Codepagex.Error now always has a string message. A reason that is not a string is converted with inspect/1. Before, Exception.message/1 raised for such exceptions.
Encoding CP932 now uses the standard byte sequences for the 23 codepoints that have several, for example U+2252 is 81E0 and not the NEC duplicate 8790. Decoding is unchanged.
The alias :iso_8859_12 is removed. It never worked, as there is no ISO 8859-12 mapping.
An entry in config :codepagex, :encodings that matches no encoding now gives a compile time warning. Such entries used to be ignored silently. Projects compiled with --warnings-as-errors will fail on a mistyped entry.

replace_nonexistent/1 no longer crashes with FunctionClauseError on input that is not valid UTF-8. Each invalid byte is replaced.
to_string!/4 and from_string!/4 raise Codepagex.Error when the outer missing_fun fails, instead of CaseClauseError.
An ArgumentError raised inside a missing_fun is no longer reported as Unknown encoding when the encoding is given as a string.
Which encodings are compiled no longer depends on where the project is located on disk. A path containing for example next, readme or unicode could result in no encodings, or wrong encoding names.
The test configuration and the README example used the encoding names MISC/CP424 and MISC/CP856, which do not exist. The names are VENDORS/MISC/CP424 and VENDORS/MISC/CP856.
Typespecs corrected: the accumulator is any term and not an integer, the inner missing_fun returns {:error, reason, acc}, and the argument and return types of from_string/4.
Benchmarks: the 1M ISO8859-1 input was about 42 times too large, and some results were labeled from_string when to_string was measured. The Benchee task no longer keeps all results in memory.

Continuous integration moved from Travis CI to GitHub Actions, testing Elixir 1.16 to 1.20 with Erlang/OTP 24 to 29, and running format checking, Credo and Dialyzer.
README documents how regexes in :encodings behave on Erlang/OTP 28.