[Build] Pin nanobind below 2.14.0

[hughperkins]

Summary

main is red on Python 3.11+ for every architecture. A single test fails,
tests/python/test_ast_refactor.py::test_single_compare:

TypeError: write_uint(): incompatible function arguments. The following argument types are supported:
    1. write_uint(self, arg0: collections.abc.Sequence[int], arg1: int, /) -> None

Invoked with types: quadrants._lib.core.quadrants_python.SNodeCxx, tuple, bool

Nothing in this repository changed. Two unpinned dependencies intersected:

• nanobind 2.14.0, released 2026-08-07: "Implicit conversion to
  integer-typed arguments now requires that the input implement Python's
  __index__ protocol." Our requirement was nanobind>=2.0.0, so builds
  picked it up automatically.
• NumPy >= 2.3 removed __index__ from numpy.bool (deprecated in 2.2).

The test assigns into a field from Python scope with c[i * 6] = a[i] == b[i].
qd.Vector stores its entries as a NumPy array, so the comparison yields
np.False_, and SNodeHostAccessor hands that straight to write_uint:

https://github.com/Genesis-Embodied-AI/quadrants/blob/main/python/quadrants/lang/field.py#L751-L755

Python 3.10 escapes because it caps NumPy at 2.2.6, where the deprecated
__index__ still exists. Confirmed locally:

NumPy	operator.index(np.False_)
2.2.6	returns 0, with a DeprecationWarning
2.4.6	raises TypeError: 'numpy.bool' object cannot be interpreted as an integer

This pins nanobind below 2.14.0 in both places that declare it, and marks both
bounds FIXME.

Why a pin rather than a fix

The pin unblocks CI immediately. The real fix is coercing the NumPy scalar at
the call site in lang/field.py, which deserves its own change and its own
test; nanobind considers the old permissive behaviour a bug, so the bound
cannot stay forever. The other half of the nanobind change is latent too:
np.float32 arguments used to be silently truncated and are now rejected, so
there may be more call sites than this one.

Evidence it is the nanobind version and nothing else

#842 passed this
same test on 3.11/3.12/3.13 on Aug 6 with the same NumPy 2.4.6, built against
nanobind 2.13.0. On Aug 7 the same test on the same xdist worker in the
same position fails, built against nanobind 2.14.0. A dry-run resolution of
the new constraint selects 2.13.0.

Test plan

• CI green on Python 3.11, 3.12 and 3.13 across cpu, metal, vulkan, and the
  manylinux wheels, which is exactly the set that is currently failing.
• test_ast_refactor.py::test_single_compare passes.

[github-actions]

Diff coverage: 0% · 0 lines, 0 missing