mruby-regexp: String#split steps by a byte through a binary subject - #7085

An empty match leaves split where it started, so the loop steps forward
before searching again. The step read the rest of the subject as UTF-8 and
cleared a whole character. A string marked by String#b has one position per
byte and __byte_match finds an empty match at every one of them, so the
fields came back joined.

"\u{1F600}".b.split(//)  # ["\xF0\x9F\x98\x80"]    <- CRuby: ["\xF0", "\x9F", "\x98", "\x80"]
"a\u{E9}b".b.split(//)   # ["a", "\xC3\xA9", "b"]  <- CRuby: ["a", "\xC3", "\xA9", "b"]
"\u{1F600}".b.split("")  # ["\xF0", "\x9F", "\x98", "\x80"]   the core split, all along

The same subject came apart two ways depending on which spelling of the same
request was used.

The change

split reads the flag off the subject with Regexp.__binary_string? and
steps by a byte, which is what gsub a few lines above already does at the
same point and for the same reason. Nothing else in the loop moves: a subject
read as UTF-8 still steps by a character, and the limit still counts fields.

Why the step cannot read it off the slice

The step asks byteslice for the rest of the subject and takes its first
element. byteslice builds that string with mrb_str_byte_subseq(), which
copies MRB_STR_SINGLE_BYTE and not MRB_STR_BINARY, so the slice arrives
unmarked and its first element is a whole character again however the subject
was marked. Reading the flag off the subject asks nothing of the slice.
Whether a slice of a binary string should itself be binary is a question about
src/string.c and is left alone here.

Testing

rake test is green on full-core with gcc on x86_64-linux, with and
without MRB_REGEXP_UNICODE_CASE: 2215 and 2214 asserts, no failures. The new
test fails without the change:

Fail: String#split - a byte-indexed subject is split by byte
  KO: 1   ->   KO: 0

Checked against CRuby 4.0.6, including the limit forms split(//, 2) and
split(//, -1), an empty subject, and an ASCII subject marked by String#b.

Summary by CodeRabbit

• Bug Fixes

  • Improved String#split behavior with zero-length regular-expression matches.
  • Binary strings are now split byte by byte, while UTF-8 and other text strings continue to split by character.
  • Corrected handling at the end of strings and when a maximum field count is specified.

• Tests

  • Added coverage for binary-string splitting, UTF-8 character splitting, and field-count limits.