fix numba stiching blend error and add numpy blend fallback

src/orisys_blend/kernel.py

@@ -0,0 +1,51 @@
+"""Fused stitch blend kernel (Numba JIT when available)."""
+
+from __future__ import annotations
+
+import numpy as np
+
+try:
+    from numba import njit
+
+    _NUMBA_AVAILABLE = True
+except ImportError:
+    _NUMBA_AVAILABLE = False
+
+    def njit(*args, **kwargs):
+        def decorator(fn):
+            return fn
+
+        return decorator
+
+
+if _NUMBA_AVAILABLE:
+
+    @njit(cache=True)
+    def blend_kernel(blend_patch, wsum_patch, cov_patch, src_patch, w_patch):
+        """Fused incremental weighted-average composite (numba-accelerated)."""
+        sh = src_patch.shape[0]
+        sw = src_patch.shape[1]
+        for y in range(sh):
+            for x in range(sw):
+                w = w_patch[y, x]
+                s0 = np.float32(src_patch[y, x, 0])
+                s1 = np.float32(src_patch[y, x, 1])
+                s2 = np.float32(src_patch[y, x, 2])
+                if cov_patch[y, x] == 0:
+                    blend_patch[y, x, 0] = s0
+                    blend_patch[y, x, 1] = s1
+                    blend_patch[y, x, 2] = s2
+                    wsum_patch[y, x] = w
+                else:
+                    ws = wsum_patch[y, x] + w
+                    if ws < 1e-6:
+                        ws = 1e-6
+                    a = w / ws
+                    blend_patch[y, x, 0] += (s0 - blend_patch[y, x, 0]) * a
+                    blend_patch[y, x, 1] += (s1 - blend_patch[y, x, 1]) * a
+                    blend_patch[y, x, 2] += (s2 - blend_patch[y, x, 2]) * a
+                    wsum_patch[y, x] = ws
+                cov_patch[y, x] = 1
+
+else:
+    blend_kernel = None