ci: run SQLite regressions on pushes and pull requests

.github/workflows/regressions.yml

@@ -0,0 +1,31 @@
+name: SQLite regressions
+
+on:
+  push:
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  sqlite-regressions:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+        with:
+          fetch-depth: 2
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@eb1897b8dc4b5d5bfe39a428a8f2304605e0983c # v7.0.0
+        with:
+          version: "0.11.19"
+          checksum: "7035608168e106375b36d0c818d537a889c51a8625fe7f8f7cad5e62b947c368"
+          enable-cache: false
+      - name: Install pinned Python
+        run: uv python install 3.13.7
+      - name: Check whitespace
+        run: git diff --check HEAD^ HEAD
+      - name: Check Python syntax
+        run: uv run --python 3.13.7 --no-project --no-sync --offline python -m compileall -q scripts tests
+      - name: Run SQLite regression suite
+        run: uv run --python 3.13.7 --no-project --no-sync --offline python -m unittest discover -s tests -v

README.md

@@ -80,6 +80,29 @@ Only `raw` and non-gold tool-assisted runs should be considered for fair ranking
 
 The July 3 result is a pilot evaluation, not a leaderboard claim.
 
+## Regression tests and CI
+
+The `SQLite regressions` workflow runs on every push and pull request with
+Python 3.13.7 and uv 0.11.19. Actions are pinned to commit SHAs, and the Linux
+uv archive is verified against its release SHA-256. After runtime setup, checks
+run offline with no third-party Python packages or benchmark downloads.
+
+Run the same suite from the repository root with the pinned runtime installed:
+
+```sh
+uv run --python 3.13.7 --no-project --no-sync --offline python -m compileall -q scripts tests
+uv run --python 3.13.7 --no-project --no-sync --offline python -m unittest discover -s tests -v
+```
+
+The 12 tests execute the real SQLite evaluator and comparator against temporary
+fixture databases. They cover denied mutations and unchanged database contents,
+safe literals/comments/identifiers, instruction limits and connection cleanup,
+ordered versus unordered results, duplicate counts, types, NULLs, and the
+versioned evaluation/report contract. A failed test exits nonzero and fails CI.
+These are local fixture regressions, not measured model accuracy or acceptance
+of downloaded benchmarks. CI also checks Python syntax and commit whitespace;
+the repository currently has no configured formatter, linter, or type checker.
+
 ## Documentation
 
 - `docs/HARNESS.md` explains the run and evaluation flow.