[tannevaled]

Two README sentences had fallen behind the code they describe. Both claims
below come from running the real interpreter, not from reading the binding.

NDArray — the axis keyword (#883)

The line read (axis, keepdims: true), so the positional form was the only
documented way to reduce along an axis, and cumsum was absent from the
sentence. rbgo built at 6eb7e4ad:

a.sum                   => 15.0
a.sum(axis: 0).to_a     => [3.0, 5.0, 7.0]
a.sum(0).to_a           => [3.0, 5.0, 7.0]
a.cumsum(axis: 1).to_a  => [0.0, 1.0, 3.0, 3.0, 7.0, 12.0]
a.argmax(axis: 1).to_a  => [2.0, 2.0]
a.sum(0, axis: 0)       => ArgumentError: axis given both positionally and as a keyword

Predefined globals — $/ is not $, (#886)

The list claims each global gets "the setter MRI's variable table gives it
rather than one untyped slot", and then groups $;, $, and $/ as though
one setter covered them. Init_IO wires $, and $\ to rb_str_setter and
$//$-0 to deprecated_rs_setter; #886 made that split real, and $/ now
stores a frozen plain String of the bytes it was handed rather than the object.
That difference is precisely what the sentence promises, so it is now spelled
out.

rubydiff on a program exercising all four, against MRI 4.0.7:

AGREE   exit 0
  A rbgo-main   ruby 4.0.9 [arm64-darwin]
  B ruby@4.0    ruby 4.0.7 (2026-09-15 revision 229531a6cf) +PRISM [arm64-darwin25]

The control: the same program under the system ruby 2.6.10 diverges on
exactly the two $/ lines and agrees on the two $,/$\ lines. The freezing
is 4.0 behaviour that rbgo now matches, not something rbgo invented — a run
that agreed with 4.0.7 and with 2.6.10 alike would have proved nothing.

🤖 Generated with Claude Code