edits conftest to run single thread for hac (#1070)

Author: Dpananos

tests/conftest.py

@@ -1,50 +1,34 @@
 "Pytest configuration for pyfixest tests."
 
 import os
-
-import pytest
-
-
-@pytest.fixture(scope="session", autouse=True)
-def single_thread_blas():
-    """
-    Force single-threaded BLAS for deterministic HAC standard errors.
-
-    What Claude says:
-
-    Multi-threaded BLAS libraries (OpenBLAS, MKL, Accelerate) can produce
-    slightly different numerical results due to different parallel reduction
-    orders when computing matrix multiplications. This causes sporadic test
-    failures in HAC variance calculations even though both R and Python
-    implementations are mathematically correct.
-
-    The differences arise because floating-point arithmetic is not associative:
-    (a + b) + c ≠ a + (b + c) in IEEE 754. Different thread scheduling can
-    change the order of operations, leading to different rounding errors.
-
-    By forcing single-threaded execution, we ensure deterministic results
-    that match R's fixest package exactly.
-    """
-    # Store original values to restore after tests
-    original_values = {}
-
-    env_vars = [
-        "OMP_NUM_THREADS",
-        "OPENBLAS_NUM_THREADS",
-        "MKL_NUM_THREADS",
-        "VECLIB_MAXIMUM_THREADS",
-        "NUMEXPR_NUM_THREADS",
-    ]
-
-    for var in env_vars:
-        original_values[var] = os.environ.get(var)
-        os.environ[var] = "1"
-
-    yield
-
-    # Restore original values after all tests complete
-    for var, value in original_values.items():
-        if value is None:
-            os.environ.pop(var, None)
-        else:
-            os.environ[var] = value
+import sys
+
+# Force single-threaded BLAS for deterministic HAC standard errors.
+#
+# Multi-threaded BLAS libraries (OpenBLAS, MKL, Accelerate) can produce
+# slightly different numerical results due to different parallel reduction
+# orders when computing matrix multiplications. This causes sporadic test
+# failures in HAC variance calculations even though both R and Python
+# implementations are mathematically correct.
+#
+# The differences arise because floating-point arithmetic is not associative:
+# (a + b) + c ≠ a + (b + c) in IEEE 754. Different thread scheduling can
+# change the order of operations, leading to different rounding errors.
+#
+# By forcing single-threaded execution, we ensure deterministic results
+# that match R's fixest package exactly.
+#
+# This must be set at module level (before numpy/scipy imports) because
+# BLAS libraries read these environment variables when first initialized.
+
+# Check if HAC tests are being run
+_run_hac_tests = any(
+    arg in sys.argv for arg in ["test_hac_vs_fixest.py", "-m hac", "--markers=hac"]
+) or any("test_hac_vs_fixest" in arg for arg in sys.argv)
+
+if _run_hac_tests:
+    os.environ["OMP_NUM_THREADS"] = "1"
+    os.environ["OPENBLAS_NUM_THREADS"] = "1"
+    os.environ["MKL_NUM_THREADS"] = "1"
+    os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
+    os.environ["NUMEXPR_NUM_THREADS"] = "1"