[berkgeveci]

These are the Tack fixes from the 0.7 rounds on the Mac (CPU and Metal) and on jura (CUDA). Tack itself had no other failures: every other failure was in a test script or an example port.

The commits

1. test_arrays_protocol: the permutation's ids first. This was found on both machines.
   • test_implicit_arrays_index_from_the_host_as_arrays_do (from f312f1c) built Permutation(stored, ids), but the signature is (ids, base). So float values became ids, read past the three-element id array.
   • In plain mode the test passed, because the host and the kernel read the same garbage. The suite under TACK_DEBUG=1 caught it. CI never runs debug mode.
   • The fix puts the ids first and checks the permutation's values. Permutation takes float ids without complaint; that's unchanged here.
2. Examples 51 and 52 fetch the OpenFOAM case's directory, not its .foam. This is jura's 87b3a01.
   • Asking _vtk_data.path for OpenFOAM/<case>/<case>.foam fetched only that empty file. On a machine without VTK's data, the reader then had no case.
   • Example 53 already asked for the directory. On jura, both examples pass from an empty cache with this fix.
3. Device arrays go to VTK only where VTK holds that device's memory. This is jura's 0a040d9, with a changelog entry added.
   • VTK's nightly wheels have DLPack, but their Viskores runs on the CPU alone. They still list CUDA in dlpack_support.supported_device_types(), and refuse CUDA memory only when asked.
   • So on VTK master, every Tack VTK algorithm on a GPU failed handing its output back.
   • The interop now tries a four-value array once per type of device (_vtk_holds), and copies to the host where VTK can't hold the memory. The two interop tests that assumed device memory skip on such a VTK.
   • On jura, VTK master went from 2 failed test_vtk_*.py tests and failures of examples 50-53 on CUDA to all passing. The local CUDA VTK with Viskores' device arrays still shares without a copy (212 passed).

The rounds

The Mac (main at dc4ca7a):

CPU	Result
Suite	6768 passed
64-bit ids	3547 passed
L2 cell-fastest	3548 passed
TACK_DEBUG=1	commit 1's failure only
timing	3 passed
Examples	39 passed
Docs' 163 Markdown examples, mkdocs build --strict, lint	clean

Metal	Result
Suite, one process	9390 passed
Examples	33 passed
round07.py on all 11 real datasets, CPU against Metal	113 rows agree, with the script's own fixes

jura: CUDA, every dataset. The details are in the vault's jura ticket.

On this branch: test_vtk_interop.py and test_arrays_protocol.py pass (44 passed), examples 51 and 52 pass, and lint and the changelog check are clean.

🤖 Generated with Claude Code