[tannevaled]

Three more of the 27 examples that went red on the move to Ruby 4.0 (#799),
again one cause. core/marshal/dump_spec.rb goes 191 → 194, fail 18 → 15.

The reference

marshal.c at v4.0.9 brackets a class path with the same encoding machinery
a String goes through:

VALUE encname = w_encivar(path, arg);   /* emits 'I' when path is non-ASCII */
w_byte(TYPE_CLASS, arg);
w_bytes(RSTRING_PTR(path), RSTRING_LEN(path), arg);
w_encname(encname, arg);                /* 1 ivar: :E => true               */

w_encivar returns nil for an ASCII-only path — that single line is what keeps
every ordinary class name dumping exactly as before. encoding_name maps UTF-8
to Qtrue, so the trailer is the short :E => true form and never the long
:encoding => "…" one: a constant name here is a Go string, hence UTF-8, and
there is no way to name a constant in another encoding.

Measured, not read

rubydiff against MRI 4.0.7 on this host: AGREE, byte for byte, on all
eleven probes. The three the specs assert:

"\x04\bIc,MarshalSpec::Multibyteぁあぃい Class\x06:\x06ET"
"\x04\bIm-MarshalSpec::Multibyteけげこご Module\x06:\x06ET"
"\x04\bIc+MarshalSpec::Multibyteぁあぃい Time\x06:\x06ET"

⛔ The controls matter more than the three rows

control	why
Marshal.dump(String) is still "\x04\bc\vString"	a wrapper emitted unconditionally passes all three spec rows and breaks every existing dump
all three round-trip to the identical object	the reader had never been handed an 'I' in front of a 'c' — and the specs do not assert this: their second line has no .should
Marshal.dump([K, K]) still ends @\x06, loading to [true, true] on .equal?	the link must point at the start of the wrapped form, not at the 'c' inside it

specdelta, provenanced, over core/marshal:

control   rbgo-rmaxC  — built from fdf0939840e8
treatment rbgo-mshlT  — built from 6a7944b6bc40
MOVE none declared: both engines read the same corpus, so every delta is the ENGINE

core/marshal/dump_spec.rb    191   194   +3
core/marshal/load_spec.rb    230   230   +0      <- the reader side did not move
6 file(s): control 442 passing, treatment 445 passing, +3
1 file(s) moved, 0 regressed

Coverage: writeClassRef and classPathEncIvar at 100.0%.

Independent of #892 — different file, different cause, same issue.

🤖 Generated with Claude Code