src/skimage2/feature/_peak.py (1)

1-6: Align imports with repo conventions (relative + SciPy alias).

Prefer relative imports from skimage2 internals and the import scipy as sp convention; then replace ndi.* usages with sp.ndimage.* throughout the file.

✅ Suggested update

-import scipy.ndimage as ndi
-
-from skimage._shared._warnings import warn_external
-from skimage._shared.coord import ensure_spacing
+import scipy as sp
+
+from .._shared._warnings import warn_external
+from .._shared.coord import ensure_spacing

As per coding guidelines, use relative module imports (e.g., from .._shared import xyz) and the import convention: import scipy as sp.

src/skimage/feature/peak.py (1)

131-144: Consider deprecate_func for standardized deprecation handling.

Since this is a public API deprecation, using skimage._shared.utils.deprecate_func would align with the project’s standard deprecation utilities while still allowing a custom warning category.

As per coding guidelines, use skimage._shared.utils.deprecate_parameter and skimage._shared.utils.deprecate_func utility functions for removing keywords or entire functions.

tests/skimage2/feature/test_peak.py (1)

1-559: Missing test coverage for p_norm parameter.

Per PR objectives and Issue #6608, a key change is setting p_norm=2 (Euclidean distance) as the new default instead of p_norm=np.inf (Chebyshev distance). The commit messages also note that "existing tests do not cover p_norm and a test should be added."

Consider adding a test that verifies:

1. The default p_norm=2 behavior (Euclidean distance)
2. Explicit p_norm values produce expected results for peak suppression based on distance

💡 Example test for p_norm coverage

def test_p_norm_euclidean_default(self):
    """Test that p_norm=2 (Euclidean) is the default for distance calculation."""
    image = np.zeros((10, 10))
    # Place two peaks at distance sqrt(8) ≈ 2.83 apart (Euclidean)
    # but distance 2 apart (Chebyshev/infinity norm)
    image[3, 3] = 2
    image[5, 5] = 1
    
    # With min_distance=2, Chebyshev would suppress the second peak
    # but Euclidean (sqrt(8) > 2) should keep both
    peaks = peak_local_max(image, min_distance=2, threshold_rel=0)
    assert len(peaks) == 2

def test_p_norm_infinity(self):
    """Test explicit p_norm=inf uses Chebyshev distance."""
    image = np.zeros((10, 10))
    image[3, 3] = 2
    image[5, 5] = 1
    
    # With p_norm=inf, distance is max(|5-3|, |5-3|) = 2
    # With min_distance=3, should suppress second peak
    peaks = peak_local_max(image, min_distance=3, threshold_rel=0, p_norm=np.inf)
    assert len(peaks) == 1

tests/skimage/feature/test_peak.py (1)

1-662: Consider adding a p_norm deprecation/behavior change test for the wrapper.

Similar to the skimage2 test file, this file should also verify that the skimage wrapper correctly handles or documents the p_norm default change behavior. This is important since the old skimage implementation used p_norm=np.inf while skimage2 uses p_norm=2.

src/skimage2/feature/_peak.py (1)

98-103: Use TypeError for type validation in tuple elements.

When a tuple element is not an integer, this is a type error rather than a value error. The static analysis correctly identifies this.

♻️ Proposed fix

         for exclude in exclude_border:
             if not isinstance(exclude, int):
-                raise ValueError(
+                raise TypeError(
                     "`exclude_border`, when expressed as a tuple, must only "
                     "contain ints."
                 )

tests/skimage2/feature/test_peak.py (2)

154-178: Consider seeding random number generator for test reproducibility.

test_reorder_labels uses np.random.uniform without setting a seed, which could lead to flaky tests. Consider using np.random.seed() or a numpy.random.Generator for deterministic behavior.

♻️ Proposed fix

     def test_reorder_labels(self):
+        np.random.seed(42)
         image = np.random.uniform(size=(40, 60))

---

13-560: Add tests for p_norm parameter to verify Euclidean distance behavior.

The PR changes the default p_norm from np.inf (Chebyshev distance) to 2 (Euclidean distance) per issue #6608. However, there are no tests verifying:

1. The default p_norm=2 behavior differs from p_norm=np.inf
2. Different p_norm values produce expected results

Consider adding a test that demonstrates the difference between distance metrics.

💡 Example test for p_norm parameter

def test_p_norm_euclidean_vs_chebyshev(self):
    """Test that p_norm=2 (Euclidean) differs from p_norm=inf (Chebyshev)."""
    image = np.zeros((10, 10))
    # Place peaks at positions where Euclidean and Chebyshev distances differ
    image[2, 2] = 2  # highest peak
    image[4, 5] = 1  # Euclidean dist ~3.6, Chebyshev dist = 3
    
    # With min_distance=3 and Chebyshev (inf-norm), both peaks kept (dist=3)
    peaks_chebyshev = peak_local_max(
        image, min_distance=3, p_norm=np.inf, exclude_border=False
    )
    
    # With min_distance=3 and Euclidean (2-norm), second peak excluded (dist~3.6>3)
    # Actually with ensure_spacing, higher intensity wins, so behavior depends on implementation
    peaks_euclidean = peak_local_max(
        image, min_distance=3, p_norm=2, exclude_border=False
    )
    
    # Verify default uses Euclidean
    peaks_default = peak_local_max(image, min_distance=3, exclude_border=False)
    assert_equal(peaks_default, peaks_euclidean)

Based on learnings: "Tests should be added to cover all modifications in behavior".

doc/source/user_guide/skimage2_migration.md (1)

77-89: Migration guidance is less concrete than the binary_blobs example.

The binary_blobs migration section (lines 55–66) spells out exact values and formulas. Here, old_p_norm and new_exclude_border are placeholders with only a prose explanation. Consider making this more actionable — e.g., showing concrete old defaults (p_norm=np.inf, exclude_border=min_distance) with a brief comment, similar to how binary_blobs does it.

src/skimage2/feature/_peaks.py (1)

54-62: _exclude_border mutates its input in-place — safe given callers always pass a copy.

In the labels path (line 248), labels.astype(int, casting="safe") creates a new array before passing it here. In the mask path (line 240), the mask is freshly created by _get_peak_mask. Both are safe, but the in-place mutation could be surprising if this helper is reused elsewhere. A brief docstring note (e.g., "Modifies label in-place") would help.