[BUG] mx.pad with mode="reflect"/"symmetric" hangs in an infinite loop when a padded axis is empty

[codeAnqiang-ma]

Describe the bug

mx.pad with mode="reflect" or mode="symmetric" never returns when a padded axis has size 0: it spins in an infinite loop at ~100% CPU during graph construction (no eval involved), so the process hangs unrecoverably. numpy raises a ValueError for the same input.

To Reproduce

import mlx.core as mx

mx.pad(mx.array([]), 2, mode="reflect")     # never returns
mx.pad(mx.array([]), 2, mode="symmetric")   # never returns
mx.pad(mx.zeros((0, 3)), [(1, 1), (0, 0)], mode="reflect")  # any padded size-0 axis hangs

Each of these was verified with a 10 s watchdog: the process is still alive and spinning at ~100% CPU when the watchdog fires.

Expected behavior

Raise an error the way numpy does:

>>> np.pad(np.array([]), 2, mode="reflect")
ValueError: can't extend empty axis 0 using modes other than 'constant' or 'empty'

mode="constant" on an empty axis already works in mlx and matches numpy, so only reflect/symmetric misbehave. A zero-width pad on an empty axis (e.g. mx.pad(mx.zeros((0, 3)), [(0, 0), (2, 1)], mode="reflect")) works today and should keep working — numpy allows it too.

Desktop (please complete the following information):

• OS Version: macOS 26.5.2 (Apple Silicon)
• Version: CPU-only source build at 7729d58 (0.32.1.dev); the loop is in graph construction (reflect_pad in mlx/ops.cpp), so it is backend-independent

Additional context

In reflect_pad, a padded axis with n = 0 gives tile = n - offset = 0, so chunk = std::min(remaining, tile) is always 0 and remaining -= chunk never decreases — both the low-side and high-side fill loops spin forever. n == 1 is special-cased (n > 1) but n == 0 is not. Introduced with the new reflect/symmetric modes (#3608).

While reading this code I also noticed two adjacent issues that are only reachable from the C++ pad(a, axes, low_pad_size, high_pad_size, ...) overload (the Python binding always passes normalized, full axes), flagging them here for maintainers rather than bundling them into my fix: (1) pad normalizes negative axes for the output shape but passes the raw axes through, and reflect_pad / edge_pad index starts[ax] with them, so a negative axis writes out of bounds; (2) edge_pad ignores its axes argument entirely and iterates all dims by position, which goes wrong when axes is a strict subset or reordered. Happy to file these separately if useful.

I have a minimal fix (raise std::invalid_argument for a non-zero pad on an empty axis, matching numpy) with a regression test ready and will open a PR shortly.