Fix window_scatter_max/min crash on f64 tensors (#1711)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>

Author: 3 people

nx/lib/nx/binary_backend.ex

@@ -1631,11 +1631,7 @@ defmodule Nx.BinaryBackend do
 
         # Compute absolute index in padded space, then adjust back to
         # original (unpadded) coordinates by subtracting low-padding
-        padded_absolute_index =
-          anchor
-          |> Enum.zip(offset_from_anchor)
-          |> Enum.map(fn {x, y} -> x + y end)
-
+        padded_absolute_index = Enum.zip_with(anchor, offset_from_anchor, &+/2)
         absolute_index = Enum.zip_with(padded_absolute_index, low_pads, &-/2)
 
         source_consumed = i * source_size
@@ -1664,7 +1660,8 @@ defmodule Nx.BinaryBackend do
       |> Enum.group_by(&elem(&1, 1), &elem(&1, 0))
       |> Enum.map(fn {index, value} ->
         offset = weighted_offset(output_weighted_shape, index)
-        {offset, Enum.reduce(value, init_value, scatter_fn)}
+        tensor = Enum.reduce(value, init_value, scatter_fn)
+        {offset, scalar_to_number(tensor)}
       end)
       |> Enum.sort_by(&elem(&1, 0))
 
@@ -1675,7 +1672,12 @@ defmodule Nx.BinaryBackend do
         {acc_offset, acc_binary} ->
           num_vals_before = div(offset - acc_offset, output_size)
           vals_before = List.duplicate(init_binary, num_vals_before)
-          source_val = to_binary(value)
+
+          source_val =
+            match_types [output_type] do
+              <<write!(value, 0)>>
+            end
+
           new_binary = :erlang.list_to_bitstring([vals_before, source_val])
 
           {offset + output_size, <<acc_binary::bitstring, new_binary::bitstring>>}

nx/test/nx_test.exs

@@ -1059,6 +1059,24 @@ defmodule NxTest do
                ])
     end
 
+    test "computes window scatter max with f64" do
+      t = Nx.iota({6}, type: :f64)
+      s = Nx.iota({3}, type: :f64)
+      init = Nx.tensor(0.0, type: :f64)
+      result = Nx.window_scatter_max(t, s, init, {2}, strides: [2], padding: :valid)
+      assert Nx.type(result) == {:f, 64}
+      assert Nx.shape(result) == {6}
+    end
+
+    test "computes window scatter min with f64" do
+      t = Nx.iota({6}, type: :f64)
+      s = Nx.iota({3}, type: :f64)
+      init = Nx.tensor(0.0, type: :f64)
+      result = Nx.window_scatter_min(t, s, init, {2}, strides: [2], padding: :valid)
+      assert Nx.type(result) == {:f, 64}
+      assert Nx.shape(result) == {6}
+    end
+
     test "computes window reduce (sum of squares) with same padding" do
       t = Nx.iota({4, 4}, type: {:f, 32})