Read the pieces partition cuts the way the whole was read - #7146

Follow-up to #7136, which carried the byte reading into every cut of a string but this one.

String#partition and String#rpartition build their pieces with mrb_str_new() and never look at what the string they cut was read as, so the pieces come back reading UTF-8 over bytes that refuse to read as it:

b = "a\xABb".b

b.partition("a")   #=> ["", "a", "\xABb"]  head and tail UTF-8, invalid  (CRuby: ASCII-8BIT)
b.rpartition("b")  #=> ["a\xAB", "b", ""]  head and tail UTF-8, invalid  (CRuby: ASCII-8BIT)
b.partition("x")   #=> the two empty pieces UTF-8                        (CRuby: ASCII-8BIT)

b.split("a")       #=> ASCII-8BIT   (since #7136)
b.chars            #=> ASCII-8BIT
b[1, 2]            #=> ASCII-8BIT

Since 80c781b refuses a subject whose bytes do not read as UTF-8, this is no longer a mis-indexing but a refusal: b[1, 2] =~ /b/ answers, while b.partition("a")[2], the same two bytes, raises ArgumentError.

The rule

The three pieces do not all come from one place, and CRuby reads each accordingly.

• The head and the tail are cut out of the receiver. They hold nothing but bytes of it, so they are read the way it is, ASCII bytes and all, which is what every other cut already does.
• The middle piece is the separator that was handed in. It is read the way that was, which mrb_str_dup() already gives it. An empty separator now comes back through that same dup rather than as a fresh literal, so it too answers the way it was handed in.
• Where the separator is found nowhere there is none to hand back. The two empty pieces stand for places in the receiver, so they carry the receiver's reading.

str_cut_piece() says the first and the last of these once for both methods: cut the bytes, then read them the way the string they were cut from was read. An empty piece goes through it as a cut of no bytes, which is what it is, rather than through mrb_str_new_lit() where no reading is at hand to carry.

This reproduces CRuby's answer for every pair it accepts. The pairs CRuby refuses with Encoding::CompatibilityError are separators found nowhere here, as they already were, so "aあb".partition(171.chr) hands the receiver back whole and says nothing about the reading rather than something false, the same choice #7137 made.

Tests

The first commit pins where every answer stands before any of it moves; the second flips exactly the pins it changes. Both sit in mruby-encoding's test file next to the cuts #7136 pinned, since that gem already takes mruby-string-ext as a test dependency and the reading is only visible through it.

Verification

Full suite green at every commit, on full-core with MRB_UTF8_STRING (2281 tests), the same with the C++ ABI (2281), and the default gembox, where mruby-encoding is absent and the new test skips (2066). 0 failures, 0 crashes, no new compiler warnings.

Still left out

Array#pack / String#unpack and the Symbol round trip, the two #7143 named. Each is its own change.

Summary by CodeRabbit

• Bug Fixes
  • Improved String#partition and String#rpartition to preserve byte-oriented string behavior across extracted pieces.
  • Corrected encoding propagation for found, missing, and empty separators.
  • Ensured separator encoding affects only the appropriate partition result.