Summary

keras.ops.pad() works for a Keras symbolic input with pad_width=0, but raises a TensorFlow PadV2 rank error for an eager tensor.

Reproduction

import keras
import tensorflow as tf

x = tf.constant([[[1., 2.], [3., 4.]]], dtype=tf.float64)

# Works
y1 = keras.ops.pad(keras.Input(shape=(2, 2), batch_size=1), 0, mode="constant", constant_values=0.5)
print(y1.shape)

# Fails
y2 = keras.ops.pad(x, 0, mode="constant", constant_values=0.5)
print(y2)

Expected behavior

pad_width=0 should be a no-op and return the input tensor unchanged for both eager tensors and Keras symbolic tensors.

Actual behavior

The eager TensorFlow path fails with:

The first dimension of paddings must be the rank of inputs [...] [Op:PadV2]

while the symbolic Keras input returns the original shape correctly.