[tannevaled]

Consumes go-ruby-parser v0.9.0, released this morning with two fields added for exactly this: ast.HashLit.Braced (upstream #40/#42) and ast.Call.Paren (#41/#43).

Measured

2,206 *_spec.rb under language/ + core/, SPEC_SHA 87b1631, swept from a real path — a symlinked corpus answers differently, which is issue #741 — with the installed shim verified byte-identical to this branch's.

base	23,471
branch	23,482
	+11 raw, +10 attributable, ZERO regressed

file
language/keyword_arguments_spec.rb	14 → 20	+6 — Braced
language/method_spec.rb	157 → 160	+3
language/send_spec.rb	71 → 72	+1 — Paren
core/random/bytes_spec.rb	8 → 9	the known 8/9 flake, not a mover

Every witness, both directions

def f(*a, **k) = [a, k]
f({x: 1})  # [[{x: 1}], {}]   positional
f(x: 1)    # [[], {x: 1}]     keywords
f(h)       # [[{y: 2}], {}]   a variable was always fine
f(**h)     # [[], {y: 2}]
f({})      # [[{}], {}]       an EMPTY POSITIONAL Hash
f()        # [[], {}]         nothing at all

nope_xyz      # NameError:     undefined local variable or method 'nope_xyz' for main
nope_xyz()    # NoMethodError: undefined method 'nope_xyz' for main
nope_xyz!     # NoMethodError  — tFID is an FCALL
nope_xyz?     # NoMethodError
nope_xyz { }  # NoMethodError  — a literal block makes an FCALL
1.nope_xyz    # NoMethodError

All byte-for-byte against ruby 4.0.5.

One guard removed

vm.vcallNameShape re-checked the name half of the VCALL rule inside the VM, because the compiler could not see the parentheses and set FlagSendVCall for nope! as well. With Paren the compiler states the whole rule, so the VM reads the flag alone. Two guards for one rule is how they drift apart.

Two things the change uncovered, fixed here because they are its consequences

The MSpec shim's complain took a keyword verbose: where upstream mspec takes a positional options Hash. It worked only because the compiler could not tell complain(re, {verbose: true}) from complain(re, verbose: true). Read correctly the braced form is positional, a keyword-only signature rejects it, and core/regexp/{new,compile}_spec.rb each went −2 with ArgumentError.

One signature covers all 46 call sites — the 18 written bare and the 2 braced — because Ruby converts keyword arguments to a positional Hash when the callee declares none, and rbgo does the same (measured against the oracle before relying on it).

Worth naming how this surfaced: the sweep reported those −2 while a direct run of the same spec reported 38. run.sh:66 copies its own spec_helper.rb over the corpus on every run, so the judge tree's shim was overwriting the corrected one. Same hazard as #741's path frame: the measurement carries state the pin does not cover.

cmd/freeze-prelude took its paths relative to the working directory, so go generate ./internal/vm — which is what TestEmbeddedPreludeMatchesSource tells you to run — could not work at all, since go:generate runs in the directory of the file carrying the directive. It now locates the module root itself, and a missing go.mod is fatal rather than a silent write to the wrong place. Both spellings verified and idempotent.

The frozen prelude is regenerated: 11 lines, all send flags — the compiler recording what it previously could not.

Gates

go test ./... green on Go 1.27.1 with CGO_ENABLED=0; gofmt -l and go vet clean; GOOS=wasip1, GOOS=js GOARCH=wasm and -tags rbgo_closed ./internal/... all build.

For whoever merges #739

The shim fix moves core/regexp/{new,compile}_spec.rb up by 2 each on any binary, so that PR's per-file baseline wants re-recording once this lands.

🤖 Generated with Claude Code