fix(scipy): relax STFT test tolerances

overlays/patches/scipy/README.md

@@ -0,0 +1,70 @@
+# SciPy STFT test tolerances
+
+The x86-64-v3 build can produce small floating-point residuals in inverse-STFT
+comparisons and scaling round trips. The original bounds reject these results,
+including residuals around `4e-17` where a round trip expects zero for a signal
+with amplitude 2.
+
+## Scope and behavior
+
+`stft-test-tolerances.patch` changes only the signal tests:
+
+- The inverse-STFT comparison in `_scipy_spectral_test_shim.py` uses
+  `max(1e-7, 2 * np.finfo(x.dtype).eps)` as its relative tolerance. Float64
+  keeps the original bound, and the existing i686 override remains.
+- Three scaling round trips in `test_spectral.py` gain an absolute tolerance
+  of one epsilon for the input dtype, allowing small residuals near zero.
+
+The tests remain enabled, and the production STFT implementation is unchanged.
+
+## Reproduction and focused checks
+
+From this directory, apply the patch to a disposable SciPy 1.18.0 checkout:
+
+```sh
+patch --fuzz=0 -d /path/to/scipy -p1 < stft-test-tolerances.patch
+```
+
+Build and install that tree with SciPy's test dependencies. From outside the
+source directory, run the installed tests:
+
+```sh
+python -m pytest --pyargs scipy.signal.tests.test_spectral \
+  -k 'roundtrip_float32 or roundtrip_scaling' -q
+```
+
+Use the same compiler flags and numerical libraries for before/after runs.
+The earlier reproduction called `TestSTFT.test_roundtrip_float32` and
+`TestSTFT.test_roundtrip_scaling` against the x86-64-v3 libraries, then loaded
+patched copies of the test modules. Both failed with the original bounds
+and passed with the adjusted bounds.
+
+## Upstream status
+
+[SciPy issue #25488](https://github.com/scipy/scipy/issues/25488) records
+related test failures with architecture-specific compiler flags. It is
+context for the local tolerance repair; this exact patch has not been
+submitted upstream during this work.
+
+## Local NixOS integration and build results
+
+[`../default.nix`](../default.nix) loads [`default.nix`](default.nix) through
+`pythonPackagesExtensions`, preserving the package's existing patches.
+The override also covers SciPy used to test other Python dependencies,
+including pgvector in portal's shared Python environment.
+
+From the repository root:
+
+```sh
+nix build --no-link -L .#nixosConfigurations.portal-1.pkgs.python314Packages.scipy
+```
+
+The original remote build of patched SciPy 1.18.0 passed 87,723 tests, with
+8,342 skips, 300 expected failures, and 22 unexpected passes. The patch and
+override have been restored byte-for-byte from commit `24cbf74f`; those counts
+describe the earlier full build.
+
+Restoration checks confirmed that the patch applies to the pinned source
+without fuzz, portal's evaluated SciPy retains its existing patch and install
+checks, and pgvector uses the patched SciPy. A full package or system rebuild
+was not repeated for this restoration.

overlays/patches/scipy/default.nix

@@ -0,0 +1,5 @@
+{ scipy }:
+scipy.overridePythonAttrs (old: {
+  # Keep the STFT tests enabled with tolerances for x86-64-v3 rounding.
+  patches = (old.patches or [ ]) ++ [ ./stft-test-tolerances.patch ];
+})

overlays/patches/scipy/stft-test-tolerances.patch

@@ -0,0 +1,51 @@
+Subject: [PATCH] signal: allow floating-point rounding in STFT tests
+
+Keep the STFT tests enabled for x86-64-v3 builds. Allow two float32
+epsilons of relative error when comparing inverse-STFT implementations;
+float64 and the existing i686 override remain unchanged. Allow one
+float64 epsilon of absolute error in all three scaling round trips,
+which otherwise require exact zeros (observed residual: 4e-17 for a
+signal with amplitude 2).
+
+Upstream issue: https://github.com/scipy/scipy/issues/25488
+
+--- a/scipy/signal/tests/_scipy_spectral_test_shim.py
++++ b/scipy/signal/tests/_scipy_spectral_test_shim.py
+@@ -294,7 +294,7 @@
+ 
+     # Adapted tolerances to account for resolution loss:
+     atol = np.finfo(x.dtype).resolution*2  # instead of default atol = 0
+-    rtol = 1e-7  # default for np.allclose()
++    rtol = max(1e-7, 2 * np.finfo(x.dtype).eps)
+ 
+     # Relax atol on 32-Bit platforms a bit to pass CI tests.
+     #  - Not clear why there are discrepancies (in the FFT maybe?)
+--- a/scipy/signal/tests/test_spectral.py
++++ b/scipy/signal/tests/test_spectral.py
+@@ -2044,7 +2044,7 @@
+ 
+         # Test round trip:
+         x1 = istft(Zs, boundary=True, scaling='spectrum')[1]
+-        assert_allclose(x1, x)
++        assert_allclose(x1, x, atol=np.finfo(x.dtype).eps)
+ 
+         # For a Hann-windowed 256 sample length FFT, we expect a peak at
+         # frequency 64 (since it is 1/4 the length of X) with a height of 1
+@@ -2074,7 +2074,7 @@
+ 
+         # Test round trip:
+         x1 = istft(Zp, input_onesided=False, boundary=True, scaling='psd')[1]
+-        assert_allclose(x1, x)
++        assert_allclose(x1, x, atol=np.finfo(x.dtype).eps)
+ 
+         # The power of the one-sided psd-scaled STFT can be determined
+         # analogously (note that the two sides are not of equal shape):
+@@ -2094,7 +2094,7 @@
+ 
+         # Test round trip:
+         x1 = istft(Zp0, input_onesided=True, boundary=True, scaling='psd')[1]
+-        assert_allclose(x1, x)
++        assert_allclose(x1, x, atol=np.finfo(x.dtype).eps)
+ 
+ 
+ class TestSampledSpectralRepresentations: