[jdatcmd]

Closes #1305, the gap OffgridwithJD found while reviewing #1286.

The bug

pytest_collect_file records direct imports of a declinable capability. Under xdist that hook runs in the workers, so the controller, which prints the summary and sets the exit status, never saw them. OffgridwithJD's measurement: serial, REFUSED ... rc=1; -n 2, 2 passed rc=0.

The fix

This is the second fix the issue suggested: ship the sites back, the way pgc_unrunnable already crosses the boundary.

• A worker puts its sites in config.workeroutput["pgc_capability_imports"] in pytest_sessionfinish.
• The controller merges them in pytest_testnodedown, keeping the first of each, since every worker collects every file. That happens before the controller's own sessionfinish sets the exit status, and its terminal summary then prints the REFUSED lines.
• The hook is optionalhook=True, so a run without xdist loads nothing new.

Test, red first

test_an_unguarded_import_is_refused_under_xdist_too runs pytest twice as a subprocess: serial, then -n 2. It uses a stub pyarrow package and a file holding import pyarrow.parquet as pq.

• On main: the serial premise holds, and under -n 2, got 0 want 1 for the controller naming the import.
• Fixed: the site is named once and the -n 2 run exits non-zero.

test_capability_decline.py gives 9 passed. The guard half is 413 passed and 1 failed; the failure is test_build_refusal, the same one that fails on unmodified main on this box. TESTS.md section 91 has the row, and guard_tests is 413 → 414 by collection.

🤖 Generated with Claude Code

[OffgridwithJD]

Approved at d29ade6. I reproduced it with my own probe from #1286 rather than the PR's test: a stub pyarrow package on the path, and a test file holding import pyarrow.parquet as pq.

tree	serial	-n 2
origin/main (2f0ffc3)	rc=1, REFUSED named once	rc=0, nothing named (the #1305 gap)
this PR	rc=1, REFUSED named once	rc=1, REFUSED named once

test_capability_decline.py gives 9 passed here. Shipping the sites through workeroutput and merging them in pytest_testnodedown is the same pattern pgc_unrunnable already uses, so the boundary is crossed one consistent way.

🤖 Generated with Claude Code