[tannevaled]

Closes issue 756. Closes the kind-A half of issue 759.

Two issues that meet on one line of output

756 — Kernel#to_s is MRI's rb_any_to_s, which always prints the
address. Object.new.to_s is #<Object:0x…>, not #<Object>. rbgo printed the
short form, so #to_s and #inspect disagreed with each other inside rbgo
while #inspect agreed with MRI.

759 (kind A) — Thread::Mutex and three siblings answered their bare
name. In MRI the class has one name, the qualified one, and top-level Mutex is
a second constant pointing at the same class: Mutex.name is "Thread::Mutex"
however it was reached, and Mutex.equal?(Thread::Mutex) is true in both.

They are coupled: RMutex.ToS hardcoded "#<Thread::Mutex>", so once
Kernel#to_s builds its string from classOf(self).name the bare name becomes
visible as #<Mutex:0x…> — a correct address with a wrong name. The first
version of this branch did exactly that. Fixing either alone trades one
divergence for another.

⚠ The scope is narrower than MRI's, deliberately

The fix applies to a plain *RObject only. Thirteen core classes reach
Kernel#to_s here and MRI gives all thirteen an address — but most are Go types
whose own ToS() is their #to_s: rbgo implements many objects' to_s as
the Go method rather than as a registered Ruby one, and they arrive at
Kernel#to_s relying on the delegation.

A version that applied the identity form to everything reaching Kernel#to_s
broke 78 tests, and every one of them was right to break:

receiver	its #to_s must be
Capistrano::Server	"db1" — the hostname
Capistrano::Task	"solo" — the task name
Net::SMTP::Response	"250 2.0.0 queued"
OpenIDConnect::ProviderMetadata	#<… issuer=https://…>
the top-level self	"main" — which is what irb's prompt is built from

I had transposed a rule that is true of MRI — "reaching rb_any_to_s means
the receiver has no useful representation" — into an engine whose dispatch works
differently. My seven targeted probes were all green; only the full suite said
otherwise. The count is in the code comment so the next person to widen this
knows what they will meet.

The three tests that needed updating, each checked against MRI

• builtin_subclass_test.go and phase1_test.go pinned "#<Plain>" and
  "#<E>" for a plain object. MRI 4.0.7 answers "#<Plain:0xA>" and
  #<E:0xA>, so those were green only while rbgo omitted the address.
• wave59_puts_to_s_test.go's wrapper block already normalised addresses to
  0xA — it was built for this. That block is now byte-identical to MRI on 13
  of its 14 lines.

The fourteenth is worth naming rather than folding in: rbgo's Enumerator#to_s
answers MRI's inspect form, #<Enumerator: [1, 2]:each>, where MRI's
#to_s inherits Kernel's and answers #<Enumerator:0x…>. A separate
divergence, recorded in the test.

Measured

Reach established by census, not assumption: of 31 core classes asked
instance_method(:to_s).owner, thirteen answer Kernel in rbgo and MRI
answers Kernel for the same thirteen. Everything else — including a subclass
of String or Array — defines or inherits its own, and a control case asserts
each of those is untouched.

Full suite: 0 failures.

🤖 Generated with Claude Code