ci(WO71): conformance matrix — frozen fixtures byte-exact in CI

.github/workflows/rake.yml

@@ -101,3 +101,39 @@
           ruby-version: "3.4"
           bundler-cache: true
       - run: bundle exec standardrb
+
+  conformance-matrix:
+    name: Conformance matrix (frozen fixtures)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: ruby/setup-ruby@v1
+        with:
+          ruby-version: "3.3"
+          bundler-cache: true
+      - name: Download fixtures (GitHub Releases channel)
+        run: |
+          mkdir -p fixtures && cd fixtures
+          BASE="https://github.com/interscript/interscript-models/releases/download"
+          curl -fsSL -o models-index.yaml "$BASE/index-v12/models-index.yaml"
+          curl -fsSL -o models-index.yaml.sha256 "$BASE/index-v12/models-index.yaml.sha256"
+          shasum -a 256 -c models-index.yaml.sha256
+          for v in urd-diac-base fas-diac-base urd-g2p-base; do
+            curl -fsSL -O "$BASE/index-v12/$v-vectors.json"
+          done
+          # the two single-zip models; urd-g2p (split parts) resolves
+          # via the runtime resolver in a later matrix pass
+          for m in urd-diac-base-1.0 fas-diac-base-1.0; do
+            url=$(ruby -ryaml -e "puts YAML.load_file('models-index.yaml')['models']['$m']['url']")
+            curl -fsSL -o "$m.zip" "$url"
+          done
+      - name: Run the matrix
+        run: |
+          for pair in "urd-diac-base-1.0 urd-diac-base" "fas-diac-base-1.0 fas-diac-base"; do
+            set -- $pair
+            echo "::group::$1"
+            SECRYST_E2E_ZIP="$PWD/fixtures/$1.zip" \
+              CONFORMANCE_VECTORS="$PWD/fixtures/$2-vectors.json" \
+              bundle exec rspec spec/interscript/ml/conformance_spec.rb
+            echo "::endgroup::"
+          done

lib/interscript/ml/byt5_onnx.rb

@@ -38,7 +38,9 @@ def id
     def build_session(graphs, name)
       path = File.join(@tmpdir, name)
       File.binwrite(path, graphs.fetch(name))
-      OnnxRuntime::Model.new(path)
+      # single-thread decode: the conformance contract is byte-exact
+      # across environments (WO79); multi-thread fp16 ties flip rows
+      OnnxRuntime::Model.new(path, intra_op_num_threads: 1)
     end
 
     def zero_pasts

spec/interscript/ml/conformance_spec.rb

@@ -0,0 +1,28 @@
+# frozen_string_literal: true
+
+require "json"
+require "interscript/ml"
+
+# The frozen-fixture contract (interscript-models/conformance/) — the
+# ruby runtime must reproduce a model's vectors byte-exactly under the
+# declared decode mode, or the change does not merge.
+RSpec.describe "frozen conformance fixtures" do
+  before do
+    skip "set CONFORMANCE_VECTORS and SECRYST_E2E_ZIP" unless
+      ENV["CONFORMANCE_VECTORS"] && ENV["SECRYST_E2E_ZIP"]
+  end
+
+  it "reproduces the vectors byte-exactly (single-row-greedy)" do
+    vectors = JSON.parse(File.read(ENV["CONFORMANCE_VECTORS"]))
+    expect(vectors["decode"]["mode"]).to eq("single-row-greedy")
+    model = Interscript::ML::Translator.new(model_file: ENV["SECRYST_E2E_ZIP"])
+    failures = []
+    vectors["vectors"].each do |v|
+      got = model.translate(v["src"])
+      failures << "#{v["src"].inspect} -> #{got.inspect} (expected #{v["out"].inspect})" if got != v["out"]
+    end
+    expect(failures).to be_empty,
+      "#{vectors['model']}: #{failures.length}/#{vectors['vectors'].length} mismatched\n" +
+      failures.first(3).join("\n")
+  end
+end