fix: resolve pre-commit failures and clean up runner/utils

- runner.py: extract call_attr variable to avoid repeated dict lookup; return call_attr instead of module.__call__ to prevent accidental binding to module-level dunder
- utils.py: guard against division by zero in is_variance_stable when both variances are 0
- tests/test_utils.py: strengthen assertion to `is True` instead of isinstance check
- ci.yml: remove redundant PYTEST_ADDOPTS env var (--randomly-seed already in addopts)
- pyproject.toml: reformatted by pyproject-fmt (pytest section key style)

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: ternaus

.github/workflows/ci.yml

@@ -33,5 +33,3 @@ jobs:
       run: |
         uv pip install numpy
         pytest -n auto --cov=benchmark --cov=tools --cov-report=xml --randomly-seed=random
-      env:
-        PYTEST_ADDOPTS: "--randomly-seed=random"

benchmark/runner.py

@@ -432,7 +432,8 @@ def load_from_python_file(specs_file: Path) -> tuple[str, Callable[[Any, Any], A
     if "__call__" not in module.__dict__:
         raise TypeError(f"Python file {specs_file} must define __call__ function")
 
-    if not callable(module.__dict__["__call__"]):
+    call_attr = module.__dict__["__call__"]
+    if not callable(call_attr):
         raise TypeError("__call__ must be a callable function")
 
     if not hasattr(module, "TRANSFORMS"):
@@ -447,7 +448,7 @@ def load_from_python_file(specs_file: Path) -> tuple[str, Callable[[Any, Any], A
         if missing:
             raise ValueError(f"TRANSFORMS[{i}] missing keys: {missing}")
 
-    return module.LIBRARY, module.__call__, module.TRANSFORMS
+    return module.LIBRARY, call_attr, module.TRANSFORMS
 
 
 # ------------------------------------------------------------------

benchmark/utils.py

@@ -254,8 +254,10 @@ def is_variance_stable(
 
     # Check if all variance ratios are below threshold
     for i in range(len(variances) - 1):
-        var_ratio = abs(variances[i] - variances[i + 1]) / max(variances[i], variances[i + 1])
-        if var_ratio > threshold:
+        denom = max(variances[i], variances[i + 1])
+        if denom == 0:
+            continue
+        if abs(variances[i] - variances[i + 1]) / denom > threshold:
             return False
 
     return True

tests/test_utils.py

@@ -109,7 +109,7 @@ def test_exact_minimum_length(self) -> None:
         # Exactly window * min_windows → should evaluate, not short-circuit
         data = [500.0] * 15
         result = is_variance_stable(data, window=5, threshold=0.05, min_windows=3)
-        assert isinstance(result, bool)
+        assert result is True
 
     @pytest.mark.parametrize("n", [0, 1, 5, 14])
     def test_short_inputs_always_false(self, n: int) -> None: