test: reduce redundant validation and stabilize GIL checks

.github/scripts/ci_changes.py

@@ -154,6 +154,8 @@ def classify(paths):
             selected.add("python_quality")
         elif path == "tests/test_pyright.py":
             selected.update(("python_quality", "python_build"))
+        elif path == "tests/test_mypy.py":
+            selected.update(("python_quality", "python_build", "wheels"))
         elif path.startswith("tests/python/"):
             selected.update(("python_build", "wheels"))
             if path == "tests/python/conftest.py":

.github/scripts/test_ci_workflow_contracts.py

@@ -1,12 +1,12 @@
 """Check lightweight CI coverage and workflow scheduling contracts."""
 
+import fnmatch
 import json
 import os
 import re
 import subprocess
 import sys
 import tempfile
-import textwrap
 import tomllib
 import unittest
 from pathlib import Path
@@ -23,7 +23,28 @@ def test_installed_typing_checks_are_required(self):
         wheel_dependencies = config["tool"]["cibuildwheel"]["test-requires"]
         checker = [value for value in test_dependencies if value.startswith("mypy==")]
         self.assertEqual(len(checker), 1)
-        self.assertIn(checker[0], wheel_dependencies)
+        self.assertFalse(any(value.startswith("mypy") for value in wheel_dependencies))
+        wheels = config["tool"]["cibuildwheel"]
+        typing_overrides = [
+            o for o in wheels["overrides"] if checker[0] in o.get("test-requires", [])
+        ]
+        self.assertEqual(len(typing_overrides), 1)
+        override = typing_overrides[0]
+        self.assertEqual(override["inherit"]["test-requires"], "append")
+        self.assertEqual(override["inherit"]["test-command"], "append")
+        self.assertIn("pytest {project}/tests/test_mypy.py", override["test-command"])
+        self.assertEqual(
+            {
+                identifier
+                for identifier in wheels["build"].split()
+                if fnmatch.fnmatchcase(identifier, override["select"])
+            },
+            {f"cp{abi}-manylinux_x86_64" for abi in (311, 312, 313, 314)},
+        )
+        self.assertIn(
+            "--ignore={project}/tests/python/test_benchmark_protocol.py",
+            wheels["test-command"],
+        )
         self.assertFalse(
             any(value.startswith("mypy") for value in config["project"]["dependencies"])
         )
@@ -32,7 +53,10 @@ def test_installed_typing_checks_are_required(self):
             "  python-result:\n", 1
         )[0]
         self.assertIn('pip install ".[test,typing]"', build)
-        self.assertIn("python -m pytest tests/python/test_editor_support.py", build)
+        self.assertIn(
+            "python -m pytest tests/python/test_editor_support.py tests/test_mypy.py",
+            build,
+        )
         self.assertNotIn("continue-on-error", build)
         self.assertIn(
             "pytest {project}/tests/python",
@@ -45,33 +69,12 @@ def test_installed_typing_checks_are_required(self):
         for path in (
             "python_bindings/__init__.pyi",
             "tests/python/typing/public_api.py",
+            "tests/test_mypy.py",
         ):
             with self.subTest(path=path):
                 self.assertLessEqual(
                     {"python_build", "python_quality", "wheels"}, classify([path])
                 )
-        # Exercise the actual aggregator script, so a failed selected matrix
-        # cannot be reported as a successful Python validation check.
-        result = workflow.split("  python-result:\n", 1)[1].split(
-            "  python-windows:\n", 1
-        )[0]
-        self.assertIn("needs: [changes, python-build]", result)
-        script = textwrap.dedent(result.split("        run: |\n", 1)[1])
-        for matrix_result in ("success", "failure", "cancelled", "skipped"):
-            with self.subTest(matrix_result=matrix_result):
-                checked = subprocess.run(
-                    ["bash", "-e", "-c", script],
-                    env=dict(
-                        os.environ,
-                        CHANGES_RESULT="success",
-                        PYTHON_SELECTED="true",
-                        MATRIX_RESULT=matrix_result,
-                    ),
-                    capture_output=True,
-                    text=True,
-                    check=False,
-                )
-                self.assertEqual(checked.returncode == 0, matrix_result == "success")
 
     def test_pyright_checks_are_required_and_scoped(self):
         config = tomllib.loads((ROOT / "pyproject.toml").read_text())

.github/workflows/python.yml

@@ -76,7 +76,7 @@ jobs:
       - name: Check installed Python types and editor metadata
         env:
           OMP_NUM_THREADS: "2"
-        run: python -m pytest tests/python/test_editor_support.py --tb=short -q
+        run: python -m pytest tests/python/test_editor_support.py tests/test_mypy.py --tb=short -q
 
       - name: Check Pyright public API and test code
         run: python -m pytest tests/test_pyright.py --tb=short -q

pyproject.toml

@@ -46,8 +46,17 @@ cmake.args = [
 build = "cp311-manylinux_x86_64 cp312-manylinux_x86_64 cp313-manylinux_x86_64 cp314-manylinux_x86_64 cp311-manylinux_aarch64 cp312-manylinux_aarch64 cp313-manylinux_aarch64 cp314-manylinux_aarch64 cp311-win_amd64 cp312-win_amd64 cp313-win_amd64 cp314-win_amd64 cp311-macosx_arm64 cp312-macosx_arm64 cp313-macosx_arm64 cp314-macosx_arm64"
 build-verbosity = 1
 environment = { CMAKE_ARGS = "-DRABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF" }
-test-requires = ["pytest", "mypy==2.4.0"]
-test-command = "pytest {project}/tests/python --ignore={project}/tests/python/test_compare_with_faiss.py --ignore={project}/tests/python/test_cpp_clustering_examples.py --tb=short -q"
+test-requires = ["pytest"]
+test-command = "pytest {project}/tests/python --ignore={project}/tests/python/test_compare_with_faiss.py --ignore={project}/tests/python/test_cpp_clustering_examples.py --ignore={project}/tests/python/test_benchmark_protocol.py --tb=short -q"
+
+# Static typing runs once per wheel ABI. Every platform still checks its installed
+# declarations, runtime signatures, documentation and package metadata.
+[[tool.cibuildwheel.overrides]]
+select = "*-manylinux_x86_64"
+inherit.test-requires = "append"
+test-requires = ["mypy==2.4.0"]
+inherit.test-command = "append"
+test-command = "pytest {project}/tests/test_mypy.py --tb=short -q"
 
 [tool.cibuildwheel.linux]
 manylinux-aarch64-image = "manylinux_2_28"

scripts/check-python.sh

@@ -18,6 +18,6 @@ if [[ "$($ruff --version)" != "ruff 0.16.1" ]]; then
 fi
 
 cd "$repo_root"
-paths=(python_bindings sample/python tests/python tests/test_pyright.py benchmarks)
+paths=(python_bindings sample/python tests/python tests/test_mypy.py tests/test_pyright.py benchmarks)
 "$ruff" format --check "${paths[@]}"
 "$ruff" check "${paths[@]}"

tests/README.md

@@ -332,12 +332,16 @@ Array shapes and value ranges remain runtime checks.
 After rebuilding and installing with the `test` extra, run:
 
 ```bash
-python -m pytest tests/python/test_editor_support.py -q
+python -m pytest tests/python/test_editor_support.py tests/test_mypy.py -q
 ```
 
 The suite checks installed declarations, signatures, documentation, examples, and
-valid/invalid calls with strict mypy. CI requires it; missing mypy fails the test.
-Keep the mypy pins in the `test` extra and wheel test requirements synchronized.
+runtime metadata on every wheel. Strict mypy valid/invalid-call checks live in
+`tests/test_mypy.py`, run for source installs and each CPython ABI on Linux x86-64
+wheels. Other wheel platforms do not repeat static analysis. CI requires mypy
+for these selected checks; missing mypy fails the test. Keep its pins in the
+`test` extra and Linux x86-64 wheel override synchronized.
+Benchmark protocol tests run in Python quality CI, separately from wheel tests.
 
 CI also checks the public type examples and test code with Pyright. With Node 24
 and the `test,typing` extras installed in the same environment, run:

tests/common/test_data.cpp

@@ -1,6 +1,5 @@
 #include "test_data.hpp"
 
-#include <cmath>
 #include <cstddef>
 #include <random>
 #include <vector>
@@ -20,76 +19,4 @@ std::vector<float> TestDataGenerator::GenerateRandomVector(
     return vec;
 }
 
-std::vector<float> TestDataGenerator::GenerateNormalizedVector(
-    size_t dim, unsigned int seed
-) {
-    auto vec = GenerateRandomVector(dim, -1.0f, 1.0f, seed);
-
-    // Calculate L2 norm
-    float norm = 0.0f;
-    for (float val : vec) {
-        norm += val * val;
-    }
-    norm = std::sqrt(norm);
-
-    // Normalize
-    if (norm > 1e-10f) {
-        for (float& val : vec) {
-            val /= norm;
-        }
-    }
-
-    return vec;
-}
-
-std::vector<std::vector<float>> TestDataGenerator::GenerateRandomVectors(
-    size_t num_vectors, size_t dim, float min, float max, unsigned int seed
-) {
-    std::vector<std::vector<float>> vectors;
-    vectors.reserve(num_vectors);
-
-    for (size_t i = 0; i < num_vectors; ++i) {
-        vectors.push_back(GenerateRandomVector(dim, min, max, seed + i));
-    }
-
-    return vectors;
-}
-
-std::vector<float> TestDataGenerator::GenerateGaussianVector(
-    size_t dim, float mean, float stddev, unsigned int seed
-) {
-    std::mt19937 rng(seed);
-    std::normal_distribution<float> dist(mean, stddev);
-
-    std::vector<float> vec(dim);
-    for (size_t i = 0; i < dim; ++i) {
-        vec[i] = dist(rng);
-    }
-    return vec;
-}
-
-std::vector<float> TestDataGenerator::GenerateSimpleVector(size_t dim) {
-    std::vector<float> vec(dim);
-    for (size_t i = 0; i < dim; ++i) {
-        vec[i] = static_cast<float>(i % 10) / 10.0f;
-    }
-    return vec;
-}
-
-std::vector<float> TestDataGenerator::GenerateZeroVector(size_t dim) {
-    return std::vector<float>(dim, 0.0f);
-}
-
-std::vector<float> TestDataGenerator::GenerateOnesVector(size_t dim) {
-    return std::vector<float>(dim, 1.0f);
-}
-
-std::vector<float> TestDataGenerator::GenerateIncrementalVector(size_t dim) {
-    std::vector<float> vec(dim);
-    for (size_t i = 0; i < dim; ++i) {
-        vec[i] = static_cast<float>(i);
-    }
-    return vec;
-}
-
 }  // namespace rabitq_test

tests/common/test_data.hpp

@@ -12,35 +12,6 @@ class TestDataGenerator {
     static std::vector<float> GenerateRandomVector(
         size_t dim, float min = -1.0f, float max = 1.0f, unsigned int seed = 42
     );
-
-    // Generate random normalized vector (unit length)
-    static std::vector<float> GenerateNormalizedVector(size_t dim, unsigned int seed = 42);
-
-    // Generate multiple random vectors
-    static std::vector<std::vector<float>> GenerateRandomVectors(
-        size_t num_vectors,
-        size_t dim,
-        float min = -1.0f,
-        float max = 1.0f,
-        unsigned int seed = 42
-    );
-
-    // Generate Gaussian distributed vector
-    static std::vector<float> GenerateGaussianVector(
-        size_t dim, float mean = 0.0f, float stddev = 1.0f, unsigned int seed = 42
-    );
-
-    // Generate a simple test vector with known values
-    static std::vector<float> GenerateSimpleVector(size_t dim);
-
-    // Generate zero vector
-    static std::vector<float> GenerateZeroVector(size_t dim);
-
-    // Generate vector with all ones
-    static std::vector<float> GenerateOnesVector(size_t dim);
-
-    // Generate vector with incremental values [0, 1, 2, 3, ...]
-    static std::vector<float> GenerateIncrementalVector(size_t dim);
 };
 
 }  // namespace rabitq_test

tests/common/test_helpers.hpp

@@ -4,7 +4,6 @@
 #include <gtest/gtest.h>
 
 #include <cmath>
-#include <cstddef>
 
 namespace rabitq_test {
 
@@ -13,74 +12,11 @@ inline bool FloatNearlyEqual(float a, float b, float epsilon = 1e-5f) {
     return std::abs(a - b) < epsilon;
 }
 
-inline bool DoubleNearlyEqual(double a, double b, double epsilon = 1e-10) {
-    return std::abs(a - b) < epsilon;
-}
-
-// Vector comparison
-inline bool VectorsNearlyEqual(
-    const float* a, const float* b, size_t size, float epsilon = 1e-5f
-) {
-    for (size_t i = 0; i < size; ++i) {
-        if (!FloatNearlyEqual(a[i], b[i], epsilon)) {
-            return false;
-        }
-    }
-    return true;
-}
-
-// Calculate relative error
-inline float RelativeError(float actual, float expected) {
-    if (std::abs(expected) < 1e-10f) {
-        return std::abs(actual - expected);
-    }
-    return std::abs((actual - expected) / expected);
-}
-
-// Calculate mean squared error
-inline float MeanSquaredError(const float* a, const float* b, size_t size) {
-    float mse = 0.0f;
-    for (size_t i = 0; i < size; ++i) {
-        float diff = a[i] - b[i];
-        mse += diff * diff;
-    }
-    return mse / static_cast<float>(size);
-}
-
-// Calculate dot product
-inline float DotProduct(const float* a, const float* b, size_t size) {
-    float result = 0.0f;
-    for (size_t i = 0; i < size; ++i) {
-        result += a[i] * b[i];
-    }
-    return result;
-}
-
-// Calculate L2 distance
-inline float L2Distance(const float* a, const float* b, size_t size) {
-    float sum = 0.0f;
-    for (size_t i = 0; i < size; ++i) {
-        float diff = a[i] - b[i];
-        sum += diff * diff;
-    }
-    return std::sqrt(sum);
-}
-
-// Custom assertion macros
-#define ASSERT_FLOAT_NEARLY_EQUAL(a, b, epsilon)                    \
-    ASSERT_TRUE(rabitq_test::FloatNearlyEqual((a), (b), (epsilon))) \
-        << "Expected: " << (a) << " to be nearly equal to " << (b)  \
-        << " (epsilon: " << (epsilon) << "), but difference was " << std::abs((a) - (b))
-
 #define EXPECT_FLOAT_NEARLY_EQUAL(a, b, epsilon)                    \
     EXPECT_TRUE(rabitq_test::FloatNearlyEqual((a), (b), (epsilon))) \
         << "Expected: " << (a) << " to be nearly equal to " << (b)  \
         << " (epsilon: " << (epsilon) << "), but difference was " << std::abs((a) - (b))
 
-#define ASSERT_VECTORS_NEARLY_EQUAL(a, b, size, epsilon)                      \
-    ASSERT_TRUE(rabitq_test::VectorsNearlyEqual((a), (b), (size), (epsilon))) \
-        << "Vectors are not nearly equal (epsilon: " << (epsilon) << ")"
-
 }  // namespace rabitq_test
 
 #endif  // RABITQ_TEST_HELPERS_HPP