fix: accept numpy integers as chunk sizes

[d-v-b]

🤖 AI text below 🤖

Fixes zarr-developers#4255.

Problem

Since 3.3.0, a per-dimension chunk size given as a numpy integer raises a TypeError:

zarr.zeros((4, 4), dtype="f4", chunks=(np.int64(2), np.int64(2)))
# TypeError: 'numpy.int64' object is not iterable

Two functions in src/zarr/core/chunk_grids.py disagreed about what counts as an integer:

• normalize_chunks_nd dispatches the scalar convenience form on isinstance(chunks, numbers.Integral)
• normalize_chunks_1d narrowed on isinstance(chunks, int)

np.int64 is numbers.Integral but is not an int subclass, so a numpy integer passed the outer dispatch, failed the inner narrowing, and fell into the else branch meant for explicit per-dimension chunk sequences — where list(chunks) was called on a scalar.

This matters because numpy integers arise naturally whenever a chunk shape is computed rather than written as a literal; any numpy reduction or elementwise op yields numpy scalars. The reporter hit it through ultrack, which sizes chunks from the array shape and is currently pinned to zarr<3.3.

Fix

• Narrow on numbers.Integral and coerce with int(), matching the caller. This is consistent with the rest of the module: guess_chunks already ends with tuple(int(x) for x in chunks), and the sequence branch already accepts Integral elements and coerces them.
• Move the -1 sentinel check inside that branch. It previously ran on the raw input, so a numpy array chunk specification made chunks == -1 evaluate to an array and raise an ambiguous-truth-value error. Rectilinear specs given as numpy arrays now work.
• Name the offending value when the spec is neither an integer nor iterable, instead of surfacing an opaque 'float' object is not iterable from list(chunks). The issue called out that the old error named neither chunks nor the bad value.

Verification

All of these previously raised and now work:

chunks=(np.int64(2), np.int64(2))         → (2, 2)
chunks=(1, 3, np.int64(16), np.int64(16)) → (1, 3, 16, 16)
chunks=np.int32(10)                       → (10,)
chunks=(np.int64(-1), 5)                  → (10, 5)
chunks=np.array([5, 5])                   → (5, 5)
shards=(np.int64(32), np.int64(32))       → (32, 32)

Regression cases added to test_normalize_chunks (scalar, per-dimension, mixed python/numpy, -1 sentinel, numpy arrays, rectilinear numpy arrays) and error cases for np.int64(0) and the non-iterable scalar.

One existing assertion in tests/test_api.py asserted the old opaque 'float' object is not iterable message and was updated to the new one.

Full test suite: 6357 passed, 373 skipped, 4 xfailed. mypy and ruff clean.