[rajeeja]

Two things made a green tree look broken on a developer machine. Neither was a
defect in the server — and investigating them turned up two more problems in
the same places.

1. SETUP.sh had drifted from CI

It ran a bare uv sync and an unfiltered pytest tests/ --ignore=tests/test_remote_agent.py, from before the suite moved to
marker-based selection. Following it today:

19 failed, 1494 passed

Every failure a ModuleNotFoundError — 16 hpc-marked tests wanting
globus_compute_sdk, 3 yac-marked tests wanting a YAC build. The first thing
a new contributor runs told them the project was broken.

Now syncs --extra hpc --dev and runs CI's own partition: -m "not hpc and not live and not yac", then -m "hpc and not yac" as a separate step, so a
broken hpc extra is reported as itself rather than vanishing into the main run.
The hpc tests are mocked and need no cluster. yac and live are printed as
deliberately not run rather than being silently absent.

2. The --ignore was hiding passing tests

pytest.ini already says CI selects by marker and never by filename. This was
the counter-example, and it was hiding 16 tests in that file that pass.
Dropping it takes the local run from 1493 to 1509 selected.

3. The prerequisite check refused the wrong interpreter

[ERROR] Python 3.11 or newer is required.

.python-version pins 3.12 and uv provisions that interpreter itself, so the
python3 on PATH is not the one the project runs on. On a machine with 3.9 on
PATH the script exited 1 and refused to install — while uv sync on that same
machine builds a working 3.12 environment. It rejected the one interpreter that
does not matter. This is why I could not run SETUP.sh end-to-end until I fixed
it.

The 3.12 Globus-submitter warning is kept, moved after the sync and asked of
uv run python, which is the interpreter that would actually submit the task.

4. The oversized-SCRIP fixture OOMs below 8 GiB

test_subset_bbox_lazy.py built each dask array as a single 2000 × 3,500,000 uint8 chunk — 6.5 GiB per array, 13 GiB for the two — and
to_netcdf materializes a chunk at a time. Measured peak RSS for that one
test: 7.45 GiB. Fits GitHub's 16 GiB runners, does not fit a laptop, so the
runners masked it.

Chunking along grid_corners to (2000, 100_000) holds each write to ~191
MiB. Intent is unchanged — the guard reads corner_var.size, the declared
shape, and never touches the data, so it still refuses a grid whose footprint
is ~104 GiB against a 100 GiB threshold.

	before	after
test_subset_bbox_lazy.py (3 tests)	7.45 GiB	1.67 GiB
full selected suite	—	1.36 GiB

Guard

TestSetupScriptMatchesCI holds SETUP.sh to the hpc extra, to marker-based
selection, and to not gating on the ambient interpreter. Reverting any of the
three fails it (verified). It parses only the pytest command lines, so the
comment explaining why --ignore is wrong does not trip its own check.

Verification

Ran the edited SETUP.sh end to end in a clean clone, on a box whose python3
is 3.9.18:

[SETUP] Running automated tests (no external data required)...
1509 passed, 44 deselected, 1 xfailed
[SETUP] Running HPC tests (mocked; no cluster required)...
35 passed, 2 skipped, 1517 deselected

No failures. Full selected suite on the branch: 1511 passed, 1 skipped, 44
deselected, 1 xfailed.