From e231e72eaa590b40931aa89d1d49da11bae830aa Mon Sep 17 00:00:00 2001 From: tannevaled Date: Sun, 4 Oct 2026 21:18:08 +0200 Subject: [PATCH 1/2] Phase 8 named a conformance gate that no longer exists, and a stale total The blurb said the sweep is "gated in CI by a shrink-only floor of 22,475". That floor was replaced by a per-file baseline in go-embedded-ruby/ruby#739, so the sentence described a mechanism that is gone -- in the present tense, which is the part that matters: a dated number going stale is honest, a dated description of how CI works today is not. The total is refreshed with it: 23,481, which is what scripts/conformance/ rubyspec/BASELINE sums to on main today (2,202 of 2,206 spec files produce a result). The repository README reads that figure out of BASELINE mechanically and a test fails if they disagree, so this page now quotes the same source rather than a hand-copied snapshot. The old floor's history is kept rather than deleted, because it is the argument for the current gate: it needed a margin wider than the largest spec file -- 384 examples in core/encoding/compatible_spec.rb -- and so could not see any regression smaller than that, while sitting 211 below the measured total. A drop of that size could not distinguish one file failing to load from hundreds of specs regressing. That is why the gate is now per file. Co-Authored-By: Claude Opus 5 --- hugo.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/hugo.toml b/hugo.toml index 2bd4280..3d5f8af 100644 --- a/hugo.toml +++ b/hugo.toml @@ -40,4 +40,4 @@ disableKinds = ["taxonomy", "term", "RSS", "sitemap", "404"] [[params.phases]] name = "Phase 8 — Conformance & performance" status = "active" - blurb = "Since July the work has been **runtime** conformance, measured against [ruby/spec](https://github.com/ruby/spec)'s `language/` + `core/` suites: **22,488 examples pass** (measured 2026-09-26 on `main`, `3e8e3cc`; three independent runs, identical), gated in CI by a **shrink-only floor of 22,475** — the floor is the level CI refuses to fall below, *not* the score. It has climbed from 6,000 when the ratchet landed on 2026-08-03. **Only `language/` and `core/` are swept** — ruby/spec's `library/`, `optional/` and `security/` trees are not run at all, so this is not a percentage of Ruby. Two of the recent jumps were fixes to the **measurement** rather than the interpreter — the mspec shim, and a parser upgrade that made 13 previously-unparseable spec files visible — and those are gains in what the measurement can *see*, not in what the VM can *do*. Earlier, a conformance campaign brought the pure-Go **front-end** to ~**100% of all valid Ruby** parsed across Rails + Puppet (99.96% of files; 2 misses are syntax-error fixtures MRI rejects) and **parse + compile** of **Rails 99.82% / Puppet 100%**, with **0 over-permissive** — and RuboCop 99.7%, Sinatra/Jekyll/Thor/Kramdown/dry-struct 100%, RSpec DSL 10/10. On the runtime side, that road is now underway: **the real `puppet apply` CLI runs end-to-end** on a pure-Go CGO=0 `rbgo` — `require \"puppet\"` boots Puppet 8.11.0 (all pure-Ruby gem deps on the path), and the genuine `Puppet::Util::CommandLine` → `Puppet::Application::Apply` path (real `OptionParser`, all types + providers, settings catalog applied to disk, then the user catalog via the transaction/RAL) converges `notify`, `file` and `exec` resources (the `exec` provider runs commands via pure-Go process execution with `onlyif`/`unless`/`creates`/`path` guards), emits real Puppet output (`Notice: hi from rbgo cli`) and exits 0 with the YAML run report round-tripped. Broader resource **providers** are the active next milestone (`user`/`group` provider-ready; `package`/`service` need a host package manager/systemd + root). rbgo is also increasingly composed from the standalone pure-Go **go-ruby-* family** — **178 modules across 176 orgs are now bound into rbgo as native modules** (parser, regexp, erb, marshal, yaml, format, optparse, strscan, json, date, … a selection is on the landing page), each its own 100%-coverage 6-arch org. (`Set` is not among them: it is implemented in-tree, over rbgo's own Hash.) Of the six arches, amd64 and arm64 are tested on every pull request; riscv64, loong64, ppc64le and s390x are validated on a **nightly schedule** — under qemu, and natively on the GCC Compile Farm plus an IBM LinuxONE s390x host — not on the pull-request path. On performance, a 6-runtime suite (rbgo/rbgo+AOT vs MRI/YJIT/JRuby/TruffleRuby): **rbgo+AOT beats MRI+YJIT on integer-bound kernels — measured here at 8.7× on the tight-loop benchmark and 2.0× on recursive `fib(35)`** (darwin/arm64, whole-process wall time); plus inline method caches, small-integer interning (a small-int loop drops from ~245k allocations to 1) and capture-tracked frame recycling (~14% faster on fib)." + blurb = "Since July the work has been **runtime** conformance, measured against [ruby/spec](https://github.com/ruby/spec)'s `language/` + `core/` suites: **23,481 examples pass** (the total `BASELINE` sums to on `main`, 2026-10-04; 2,202 of 2,206 spec files produce a result), gated in CI by a **per-file baseline** — one record per spec file, so a file that regresses is named rather than absorbed. It replaced a single shrink-only floor, which needed a margin wider than the largest spec file and was therefore blind to any regression smaller than that: the floor had climbed from 6,000 when the ratchet first landed on 2026-08-03 to 22,475, while sitting 211 below the measured total. **Only `language/` and `core/` are swept** — ruby/spec's `library/`, `optional/` and `security/` trees are not run at all, so this is not a percentage of Ruby. Two of the recent jumps were fixes to the **measurement** rather than the interpreter — the mspec shim, and a parser upgrade that made 13 previously-unparseable spec files visible — and those are gains in what the measurement can *see*, not in what the VM can *do*. Earlier, a conformance campaign brought the pure-Go **front-end** to ~**100% of all valid Ruby** parsed across Rails + Puppet (99.96% of files; 2 misses are syntax-error fixtures MRI rejects) and **parse + compile** of **Rails 99.82% / Puppet 100%**, with **0 over-permissive** — and RuboCop 99.7%, Sinatra/Jekyll/Thor/Kramdown/dry-struct 100%, RSpec DSL 10/10. On the runtime side, that road is now underway: **the real `puppet apply` CLI runs end-to-end** on a pure-Go CGO=0 `rbgo` — `require \"puppet\"` boots Puppet 8.11.0 (all pure-Ruby gem deps on the path), and the genuine `Puppet::Util::CommandLine` → `Puppet::Application::Apply` path (real `OptionParser`, all types + providers, settings catalog applied to disk, then the user catalog via the transaction/RAL) converges `notify`, `file` and `exec` resources (the `exec` provider runs commands via pure-Go process execution with `onlyif`/`unless`/`creates`/`path` guards), emits real Puppet output (`Notice: hi from rbgo cli`) and exits 0 with the YAML run report round-tripped. Broader resource **providers** are the active next milestone (`user`/`group` provider-ready; `package`/`service` need a host package manager/systemd + root). rbgo is also increasingly composed from the standalone pure-Go **go-ruby-* family** — **178 modules across 176 orgs are now bound into rbgo as native modules** (parser, regexp, erb, marshal, yaml, format, optparse, strscan, json, date, … a selection is on the landing page), each its own 100%-coverage 6-arch org. (`Set` is not among them: it is implemented in-tree, over rbgo's own Hash.) Of the six arches, amd64 and arm64 are tested on every pull request; riscv64, loong64, ppc64le and s390x are validated on a **nightly schedule** — under qemu, and natively on the GCC Compile Farm plus an IBM LinuxONE s390x host — not on the pull-request path. On performance, a 6-runtime suite (rbgo/rbgo+AOT vs MRI/YJIT/JRuby/TruffleRuby): **rbgo+AOT beats MRI+YJIT on integer-bound kernels — measured here at 8.7× on the tight-loop benchmark and 2.0× on recursive `fib(35)`** (darwin/arm64, whole-process wall time); plus inline method caches, small-integer interning (a small-int loop drops from ~245k allocations to 1) and capture-tracked frame recycling (~14% faster on fib)." From 090ac766fedc30c2b74410797ba039ed2fbe6f9d Mon Sep 17 00:00:00 2001 From: tannevaled Date: Sun, 4 Oct 2026 21:18:35 +0200 Subject: [PATCH 2/2] fix the wording of the sentence I just rewrote "the total `BASELINE` sums to on `main`" was missing its relative pronoun. Co-Authored-By: Claude Opus 5 --- hugo.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/hugo.toml b/hugo.toml index 3d5f8af..096d003 100644 --- a/hugo.toml +++ b/hugo.toml @@ -40,4 +40,4 @@ disableKinds = ["taxonomy", "term", "RSS", "sitemap", "404"] [[params.phases]] name = "Phase 8 — Conformance & performance" status = "active" - blurb = "Since July the work has been **runtime** conformance, measured against [ruby/spec](https://github.com/ruby/spec)'s `language/` + `core/` suites: **23,481 examples pass** (the total `BASELINE` sums to on `main`, 2026-10-04; 2,202 of 2,206 spec files produce a result), gated in CI by a **per-file baseline** — one record per spec file, so a file that regresses is named rather than absorbed. It replaced a single shrink-only floor, which needed a margin wider than the largest spec file and was therefore blind to any regression smaller than that: the floor had climbed from 6,000 when the ratchet first landed on 2026-08-03 to 22,475, while sitting 211 below the measured total. **Only `language/` and `core/` are swept** — ruby/spec's `library/`, `optional/` and `security/` trees are not run at all, so this is not a percentage of Ruby. Two of the recent jumps were fixes to the **measurement** rather than the interpreter — the mspec shim, and a parser upgrade that made 13 previously-unparseable spec files visible — and those are gains in what the measurement can *see*, not in what the VM can *do*. Earlier, a conformance campaign brought the pure-Go **front-end** to ~**100% of all valid Ruby** parsed across Rails + Puppet (99.96% of files; 2 misses are syntax-error fixtures MRI rejects) and **parse + compile** of **Rails 99.82% / Puppet 100%**, with **0 over-permissive** — and RuboCop 99.7%, Sinatra/Jekyll/Thor/Kramdown/dry-struct 100%, RSpec DSL 10/10. On the runtime side, that road is now underway: **the real `puppet apply` CLI runs end-to-end** on a pure-Go CGO=0 `rbgo` — `require \"puppet\"` boots Puppet 8.11.0 (all pure-Ruby gem deps on the path), and the genuine `Puppet::Util::CommandLine` → `Puppet::Application::Apply` path (real `OptionParser`, all types + providers, settings catalog applied to disk, then the user catalog via the transaction/RAL) converges `notify`, `file` and `exec` resources (the `exec` provider runs commands via pure-Go process execution with `onlyif`/`unless`/`creates`/`path` guards), emits real Puppet output (`Notice: hi from rbgo cli`) and exits 0 with the YAML run report round-tripped. Broader resource **providers** are the active next milestone (`user`/`group` provider-ready; `package`/`service` need a host package manager/systemd + root). rbgo is also increasingly composed from the standalone pure-Go **go-ruby-* family** — **178 modules across 176 orgs are now bound into rbgo as native modules** (parser, regexp, erb, marshal, yaml, format, optparse, strscan, json, date, … a selection is on the landing page), each its own 100%-coverage 6-arch org. (`Set` is not among them: it is implemented in-tree, over rbgo's own Hash.) Of the six arches, amd64 and arm64 are tested on every pull request; riscv64, loong64, ppc64le and s390x are validated on a **nightly schedule** — under qemu, and natively on the GCC Compile Farm plus an IBM LinuxONE s390x host — not on the pull-request path. On performance, a 6-runtime suite (rbgo/rbgo+AOT vs MRI/YJIT/JRuby/TruffleRuby): **rbgo+AOT beats MRI+YJIT on integer-bound kernels — measured here at 8.7× on the tight-loop benchmark and 2.0× on recursive `fib(35)`** (darwin/arm64, whole-process wall time); plus inline method caches, small-integer interning (a small-int loop drops from ~245k allocations to 1) and capture-tracked frame recycling (~14% faster on fib)." + blurb = "Since July the work has been **runtime** conformance, measured against [ruby/spec](https://github.com/ruby/spec)'s `language/` + `core/` suites: **23,481 examples pass** — the total that `BASELINE` sums to on `main` (2026-10-04; 2,202 of 2,206 spec files produce a result), gated in CI by a **per-file baseline** — one record per spec file, so a file that regresses is named rather than absorbed. It replaced a single shrink-only floor, which needed a margin wider than the largest spec file and was therefore blind to any regression smaller than that: the floor had climbed from 6,000 when the ratchet first landed on 2026-08-03 to 22,475, while sitting 211 below the measured total. **Only `language/` and `core/` are swept** — ruby/spec's `library/`, `optional/` and `security/` trees are not run at all, so this is not a percentage of Ruby. Two of the recent jumps were fixes to the **measurement** rather than the interpreter — the mspec shim, and a parser upgrade that made 13 previously-unparseable spec files visible — and those are gains in what the measurement can *see*, not in what the VM can *do*. Earlier, a conformance campaign brought the pure-Go **front-end** to ~**100% of all valid Ruby** parsed across Rails + Puppet (99.96% of files; 2 misses are syntax-error fixtures MRI rejects) and **parse + compile** of **Rails 99.82% / Puppet 100%**, with **0 over-permissive** — and RuboCop 99.7%, Sinatra/Jekyll/Thor/Kramdown/dry-struct 100%, RSpec DSL 10/10. On the runtime side, that road is now underway: **the real `puppet apply` CLI runs end-to-end** on a pure-Go CGO=0 `rbgo` — `require \"puppet\"` boots Puppet 8.11.0 (all pure-Ruby gem deps on the path), and the genuine `Puppet::Util::CommandLine` → `Puppet::Application::Apply` path (real `OptionParser`, all types + providers, settings catalog applied to disk, then the user catalog via the transaction/RAL) converges `notify`, `file` and `exec` resources (the `exec` provider runs commands via pure-Go process execution with `onlyif`/`unless`/`creates`/`path` guards), emits real Puppet output (`Notice: hi from rbgo cli`) and exits 0 with the YAML run report round-tripped. Broader resource **providers** are the active next milestone (`user`/`group` provider-ready; `package`/`service` need a host package manager/systemd + root). rbgo is also increasingly composed from the standalone pure-Go **go-ruby-* family** — **178 modules across 176 orgs are now bound into rbgo as native modules** (parser, regexp, erb, marshal, yaml, format, optparse, strscan, json, date, … a selection is on the landing page), each its own 100%-coverage 6-arch org. (`Set` is not among them: it is implemented in-tree, over rbgo's own Hash.) Of the six arches, amd64 and arm64 are tested on every pull request; riscv64, loong64, ppc64le and s390x are validated on a **nightly schedule** — under qemu, and natively on the GCC Compile Farm plus an IBM LinuxONE s390x host — not on the pull-request path. On performance, a 6-runtime suite (rbgo/rbgo+AOT vs MRI/YJIT/JRuby/TruffleRuby): **rbgo+AOT beats MRI+YJIT on integer-bound kernels — measured here at 8.7× on the tight-loop benchmark and 2.0× on recursive `fib(35)`** (darwin/arm64, whole-process wall time); plus inline method caches, small-integer interning (a small-int loop drops from ~245k allocations to 1) and capture-tracked frame recycling (~14% faster on fib)."