Make strided AxisArray data contiguous before serializing

[cboulay]

Restacked: now based on #271 (v3.10.0b4) rather than dev, as the first PR after the b4 release in the stack for the next beta. The only conflict was in tests/messages/test_axisarray.py (both sides appended tests).

Draft, same as #268 — parking it for review rather than merge. Independent of #268; they touch different files and can land in either order.

The problem

numpy hands a C- or F-contiguous array to a protocol-5 pickler out-of-band, and ezmsg's marshal writes those buffers straight into shared memory without copying. An array that is neither falls back to an in-band copy through the pickle stream.

That matters because it is the one size-dependent cost in serialization. Everything else is a flat ~7 us regardless of payload. Any transformer emitting a decimated or channel-sliced view (data[:, ::2], data[::2, :]) hits this silently, and it scales with the array.

The fix

ArrayWithNamedDims.__getstate__ makes data contiguous when it is neither C- nor F-contiguous, which puts the payload back on the out-of-band path.

layout	before	after	out-of-band buffers
C-contiguous (4 MB)	6.5 us	7.1 us	1 → 1
F-order transposed	6.6 us	6.8 us	1 → 1
strided [:, ::2] (2 MB)	311.9 us	141.5 us	0 → 1
strided [::2, :] (2 MB)	253.5 us	41.1 us	0 → 1

In-band bytes for the strided cases drop from 2,097,501 to 307.

Why this is transparent, not a behaviour change

Unpickling a strided array already yields a C-contiguous one — numpy's reduce does tobytes() on the way out. So the receiver sees exactly the same array either way; it just gets there for a lot less. Verified explicitly in the tests.

Why F-order is excluded

This is the part worth a second pair of eyes. F-contiguous arrays are already on the out-of-band path, and forcing C order on them measured ~80x worse (7 us → 582 us on a 4 MB transposed array). A blanket np.ascontiguousarray would have been a serious pessimization on a common case — a transposed AxisArray is not unusual. The guard is not (c_contiguous or f_contiguous), costing two flag reads on the common path.

Scope

• Only numpy arrays are touched. torch/cupy have their own serialization and no flags attribute to consult.
• Lives on ArrayWithNamedDims, the shared base, so CoordinateAxis gets it too — a ch axis with a strided data hit the same cliff.
• The caller's array is never mutated; the copy goes into the returned state dict only.

Testing

7 new tests: all four layouts round-trip out-of-band with correct values, F-order is confirmed not reordered, the source message is confirmed unmutated, and CoordinateAxis is covered.

Full suite green: 431 passed, 1 skipped. The 2 remaining ruff errors in these files (E731 at axisarray.py, E402 in the test) are pre-existing on dev — verified against the pristine versions.

Measurements are single-machine (Darwin, arm64, Python 3.13).

🤖 Generated with Claude Code