Split the two self-drive sweeps into one case per corpus pair

[bbfrederick]

Tier 2 of the CI speed-up. Tier 1 (#25) removed a duplicate run and took a job from 1817 s to 1066 s, leaving Test as 97.4% of it. This addresses what Test is bounded by.

The problem

xdist distributes whole tests, so a test that loops the corpus pins the corpus to one worker. The two self-drive sweeps were the suite's slowest tests for exactly that reason — ~400 s and ~200 s on a single core while the others idled — and no -n could reach inside a test body.

The change

One case per archive/PDF pair: 21 for the self-drive invariant, 6 for the named exceptions.

	before	after
the two sweeps together	~600 s	133 s (583% CPU)
suite's slowest test	400 s (self-drive)	97 s (unrelated conversion sweep)
slowest self-drive case	—	82 s (CMRR_optionscan_P1.exar1)

The single-test floor drops ~400 s → ~97 s, and the binding constraint is no longer one test. Full suite green: 3943 passed, 26 skipped.

The cost, and the companion tests that pay it

A loop can assert something a parametrized case cannot. Both sweeps counted what they drove and asserted a floor — checked >= 10, available >= 4 — which is what stopped them passing on a corpus that had quietly stopped being discoverable.

One case per pair cannot make that claim. Demonstrated rather than assumed:

$ pytest -q test_vacuous.py        # a parametrized sweep whose list emptied
s                                                              [100%]
SKIPPED [1] ...: got empty parameter set for (pair)
1 skipped in 0.00s

Zero cases passing reads exactly like every case passing. So each sweep keeps a companion test asserting the length of the list its cases are generated from. Both falsified by truncating those lists to 2 and 1:

AssertionError: only 2 self-drivable pairs discovered; this proves little
assert 2 >= 10
AssertionError: only 1 of the named exceptions are present
assert 1 >= 4

The two guards overlap, which was not designed. An emptied parameter set is a skip in tests.test_exar_patch, so the CI report check added in #25 reports it too (::error::N .exar1 tests skipped). A guard phrased as coverage rather than outcome catches failure modes invented after it was written.

It also got stricter

The loop moved on silently when a printout stopped pairing with its archive (if not report.matched: continue), so a pair could leave the sweep without saying so. A case has to decide, so it asserts the pairing — and all 21 pass, meaning that branch had never once been taken in the sweep's lifetime.

Verification

• Full suite: 3943 passed, 26 skipped, 0 failed (11:37 with -n auto)
• The 27 new cases: 29 passed in 2:13
• Empty-parameter-set hazard demonstrated; both companions falsified; mutation reverted and confirmed gone
• black, isort, codespell clean

CLAUDE.md's "one test is the floor" paragraph predicted this fix and is rewritten with the measured numbers, the pattern (when the slowest test is a loop, parametrise rather than raise -n), and the non-vacuity caveat.

🤖 Generated with Claude Code