Fix simcheck source and portable interpolation checks

pyproject.toml

@@ -53,7 +53,7 @@ test = [
     "pytest-cov>=6",
     # Monte Carlo test machinery. A dependency-group, not an extra: PyPI rejects
     # direct URL references in published project metadata.
-    "simcheck @ git+https://github.com/finite-sample/simcheck",
+    "simcheck",
 ]
 docs = [
     "sphinx>=8",
@@ -134,3 +134,7 @@ source = ["src"]
 
 [tool.uv.build-backend]
 module-name = "streamcal"
+
+[tool.uv.sources]
+# The statistical test helper is unrelated to the simcheck package on PyPI.
+simcheck = { git = "https://github.com/finite-sample/simcheck" }

src/streamcal/calibrators.py

@@ -51,8 +51,9 @@ def _grid_interp(
     """``np.interp(values, grid, fitted)`` for an evenly spaced ``grid``.
 
     Finds each interval with :func:`_grid_index` instead of a binary search,
-    then evaluates the same expression ``np.interp`` does, so the result is
-    the same to the last bit.
+    then evaluates the same expression as ``np.interp``. For nondecreasing
+    probability values, results differ by at most one floating-point step:
+    NumPy may fuse the multiply-add while separate ufuncs round twice.
 
     Args:
         values: Points to interpolate at.
@@ -71,8 +72,8 @@ def _grid_interp(
 
 
 # Batch sizes from which the grid kernels beat NumPy's binary searches. Below
-# them, NumPy's lower fixed cost wins. Both paths give identical results, so
-# the switch changes speed only. Measured with ``benchmarks.speed``.
+# them, NumPy's lower fixed cost wins. Index choices are identical; interpolation
+# can differ by one floating-point step. Measured with ``benchmarks.speed``.
 _GRID_INDEX_FROM = 1024
 _GRID_INTERP_FROM = 2048
 

tests/test_kernels.py

@@ -1,4 +1,4 @@
-"""Faster code paths must reproduce the straightforward ones exactly."""
+"""Grid kernels preserve bin choices and interpolation within one float step."""
 
 import subprocess
 import sys
@@ -30,7 +30,7 @@ def _adversarial_points(n_bins, rng):
 
 
 @pytest.mark.parametrize("n_bins", [*range(1, 70), 100, 997, 1000, 4096])
-def test_grid_kernels_match_numpy_bit_for_bit(n_bins):
+def test_grid_kernels_match_numpy(n_bins):
     rng = np.random.default_rng(n_bins)
     edges, centers, points = _adversarial_points(n_bins, rng)
     reference = np.clip(np.digitize(points, edges) - 1, 0, n_bins - 1)
@@ -39,7 +39,9 @@ def test_grid_kernels_match_numpy_bit_for_bit(n_bins):
         fitted = np.sort(rng.random(n_bins))
         fast = _grid_interp(points, centers, fitted)
         slow = np.interp(points, centers, fitted)
-        assert np.array_equal(fast.view(np.int64), slow.view(np.int64))
+        # NumPy may fuse its multiply-add; separate ufuncs round twice.
+        np.testing.assert_array_max_ulp(fast, slow, maxulp=1)
+        assert np.array_equal(_grid_interp(centers, centers, fitted), fitted)
 
 
 @pytest.mark.parametrize("n_bins", [1, 2, 20, 1000])
@@ -54,8 +56,10 @@ def test_size_switches_agree_on_both_sides_of_the_threshold(n_bins):
             _bin_index(points, edges),
             np.clip(np.digitize(points, edges) - 1, 0, n_bins - 1),
         )
-        assert np.array_equal(
-            _interpolate(points, centers, fitted), np.interp(points, centers, fitted)
+        np.testing.assert_array_max_ulp(
+            _interpolate(points, centers, fitted),
+            np.interp(points, centers, fitted),
+            maxulp=1,
         )