Tests: a fast tier by default, a reliable 4-minute slow tier behind --slow

[jries]

The suite had grown to 1322 tests and took 11:22 on four cores. Running it in parallel hung or failed in about half of the runs.

	before	after
after each change: pytest tests -n auto	11:22	~1:25
before a PR: pytest tests -n auto --slow	11:22, and in parallel it failed about half the time	~3:45, 8 of 8 green

No test was removed or weakened.

Changes

Two tiers. About 60 tests are marked slow: the tutorials, the docs-figure tests and the 11 heaviest numerical checks. They took two thirds of the time and are skipped unless --slow is given. They are skipped rather than deselected, so the summary still counts them.

The parallel hang, fixed. A GUI test that left a Qt widget in a reference cycle left it for Python's cyclic garbage collector. The collector runs in whichever thread allocates next; here that was a later test's TIFF reader thread. The widget's destructor calls QWindow::close, which waits for the GUI thread to flush window events. The GUI thread was the test, itself waiting for that fit, so the worker deadlocked. py-spy showed QWindowSystemInterface::flushWindowSystemEvents under tifffile in the reader thread. This caused:

• test_live failing in about half of the parallel --slow runs;
• the 300 s timeouts in test_gui_view3d and test_gui_render_axes.

The bug predates this PR: main's setup failed the same way. The fix is pyqtgraph's GarbageCollector remedy: conftest.py turns off automatic collection and collects on the main thread every 100 tests. Before the fix, about half of the parallel runs failed; after it, 8 of 8 passed. The fast tier's time is unchanged.

Balanced parallel runs. Three changes, all in conftest.py:

• Work stealing by default, through xdist's pytest_xdist_make_scheduler hook. xdist's own scheduler gave one worker 372 s of tutorials while the other three idled after about 100 s.
• Slow tests first when --slow is given.
• One BLAS thread per worker. Otherwise four workers each start a thread per core and are barely faster than one.

Draft tutorials in the test. SMAPPY_TUTORIAL_DRAFT=1 takes screenshots at 1× instead of 2× and saves them as JPEG (120 ms per shot) instead of WebP (550 ms). The published tutorials are unchanged. Most of a tutorial's remaining time is the GUI really working (fits, renders, simulating its data), which this test is there to exercise.

Smaller changes:

• pytest-xdist added to the test extra.
• CLAUDE.md: describes the two tiers and when the slow one must run (before a PR, and before major work is called done); asks a PR to state its --slow result; and notes that automatic garbage collection is off in the tests.

Test results

--slow -n auto: 1321 passed, 4 skipped (3:39, 3:44, 3:44 in the last three runs)
-n auto: 1260 passed, 65 skipped (1:24)

🤖 Generated with Claude Code

https://claude.ai/code/session_01NjvC6By9NAqXj9KRfwWdLM