[tannevaled]

What this locks in

Waves 33–36 merged today: TracePoint-era work aside, the movers were

PR		corpus effect
#674	instance_variable_set on native kinds (#672)	+3
#675	warnings reach stderr, Tempfile ORs its mode	+8
#677	parser v0.4.0 — do…end as a paren-less argument	+223
#679	encoding-error objects, String#encode	+37
#683	parser v0.5.0 — the COND half	0 (see below)
#685	Dir.pwd raises on a removed cwd (#681)	−122
#692	the mspec shim repaired	+357
#694	parser v0.6.0 — over-acceptance in arg_value	0 (see below)

Two of those numbers deserve their sign explained

#685 is −122 and that is the point. Dir.pwd discarded os.Getwd()'s error,
so the Dir specs — which delete the working directory in teardown — kept running
where MRI's failure cascades. rbgo scored 92 on core/dir/glob_spec.rb against
MRI's 16 through our own shim: roughly 145 examples counted as passing with no
reference able to reach them. Removing them made the measurement stricter, not
worse. Same shape as #660, where repairing ShouldProxy correctly lowered the
floor by 175.

#683 and #694 each moved zero examples, and are still conformance wins. The
parser now accepts 27 loop-condition shapes MRI accepts, and refuses 40 it
refuses; real Ruby writes none of them, so no corpus can witness either change.
Both were measured by enumerating shapes against MRI's two parsers
(--parser=prism and --parser=parse.y), which is the only instrument that can
see them.

Why this PR carries a provisional number

22650 is known-good rather than predicted: it is what CI's ratchet read on
#685 before #692 merged. #685 and #692 interact — #685 makes Dir.pwd raise when
the working directory has been removed, and #692 restores the after :all hook
whose absence was what removed it — and neither branch's CI ran with the other
(#692's lanes completed at 16:54Z, #685 merged at 17:55Z).

Rather than guess the combined total, this PR measures it: its own ratchet lane
prints the real figure, and a second commit here sets the floor to exactly that.
The number is taken from the lane that applies it; a development host reads a
different one, and a floor set below what the judge sees is slack rather than
safety.

The judge did not become lenient

#692 is a large change to our own measuring instrument, so the control that
matters is what it did to the reference, not to us:

	before	after
MRI 4.0.5 through our shim	23926	24642 (+716)
rbgo	22757	23114 (+357)

MRI rose twice as much as rbgo, MRI's unscoreable set fell 531 → 112, no file
moved down for MRI, and nowhere did rbgo rise while MRI did not. The repaired
instrument also makes headroom legible for the first time: 1528 reachable
examples, with 85 files scoring 0 for rbgo where MRI scores above 0.

🤖 Generated with Claude Code