Improve layout reliability, personal exceptions, and benchmark coverage

.github/workflows/layout-benchmark.yml

@@ -0,0 +1,67 @@
+name: Layout Reliability Benchmark
+
+on:
+  push:
+    branches:
+      - master
+      - develop
+      - "feat/**"
+  pull_request:
+    branches:
+      - master
+      - develop
+
+permissions:
+  contents: read
+
+jobs:
+  benchmark:
+    name: Layout Reliability Check (macos-15)
+    runs-on: macos-15
+    timeout-minutes: 20
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Show toolchain environment
+        run: |
+          swift --version
+          sw_vers
+          uname -a
+
+      - name: Run TestRunner
+        run: swift run -c release TestRunner
+
+      - name: Run InputPipelineTestRunner
+        # Hosted runners pass the Accessibility-dependent TextCorrector test (release CI runs
+        # the same runner); every failure here is a real failure.
+        run: swift run -c release InputPipelineTestRunner
+
+      - name: Run LayoutBenchmark self-check
+        run: swift run -c release LayoutBenchmark self-check
+
+      - name: Validate corpora and frozen layout tables
+        run: |
+          swift run -c release LayoutBenchmark validate --corpus Benchmarks/layout-corpus/v2 --tables Benchmarks/layout-tables/v1 --strict
+          # v1 is frozen with known label conflicts; it must still validate, warnings allowed.
+          swift run -c release LayoutBenchmark validate --corpus Benchmarks/layout-corpus/v1 --tables Benchmarks/layout-tables/v1
+
+      - name: Development smoke run gated against baseline v2
+        run: |
+          mkdir -p Benchmarks/results/ci
+          swift run -c release LayoutBenchmark run \
+            --adapter switchfix \
+            --split development \
+            --profile boundary-layout-only \
+            --corpus Benchmarks/layout-corpus/v2 \
+            --baseline Benchmarks/baselines/v2/corpus-v2/baseline.json \
+            --output Benchmarks/results/ci
+
+      - name: Upload benchmark artifacts
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-results-ci
+          path: Benchmarks/results/ci/
+          if-no-files-found: ignore

.github/workflows/release.yml

@@ -96,6 +96,9 @@ jobs:
       contents: write
 
     steps:
+      - name: Checkout release notes
+        uses: actions/checkout@v4
+
       - name: Download all release artifacts
         uses: actions/download-artifact@v4
         with:
@@ -112,4 +115,4 @@ jobs:
         with:
           files: |
             release-assets/*.zip
-          generate_release_notes: true
+          body_path: docs/releases/${{ github.ref_name }}.md

.gitignore

@@ -25,3 +25,4 @@ temp.txt
 *.key
 brag-output/
 brag-output-*/
+Benchmarks/results/

Benchmarks/baselines/v1/SUPERSEDED.md

@@ -0,0 +1,11 @@
+# Baseline v1 is superseded
+
+Do not compare against this baseline. Use `Benchmarks/baselines/v2/corpus-v2/baseline.json`.
+
+Verification on 2026-10-19 found that the harness that recorded v1 did not measure the app's behavior:
+
+- It configured the detector with the `.standard` Ukrainian variant for every trial. All trials use `com.apple.keylayout.Ukrainian`, whose frozen table (key 1 = и, key 11 = і) the app classifies `.legacy`. The four EN → UK misses (for example `добрий день` → `добрій день`) come from this setting alone. Rerunning the same harness with `.legacy` gave 40/40. They were not caused by static mapping, as `baseline.md` states.
+- It fed every key to the state machine as a letter, while the app treats `/`, `!`, `%`, `(` and other hard punctuation as word boundaries. This hid a real false positive: `/usr/local/bin` on ABC becomes `/гик/local/bin`.
+- Its scorer dropped missing results, ignored error status and checkpoints, and scored layout-only switches as text damage. Its runner counts and analysis text were hardcoded.
+
+The files here are kept unchanged as the historical record. `Benchmarks/baselines/v2/corpus-v1` is the corrected measurement on the same corpus (40/40 recall, 1/160 text-damage false positives).