Cache compiles and run the suite on every core in CI

[mcencini]

Requested by Matteo · project thread

Before: the last main run took 29 min. Every job compiled BART and FINUFFT (588 objects with the SIMD modules) from scratch: 1.5–6.5 min per test leg, and 14 min for the CUDA build, which then ran the whole host suite serially for another 8 min. The suite took 6–8 min per leg on one core; locally 476 s, 296 s of it in test_cli.py's moba/mobafit fits. Every PR touching src/ or docs/ also executed the whole gallery.

After: every compile goes through ccache (hendrikmuhs/ccache-action, one store per matrix leg, nvcc included), with Ninja as the generator, so a run recompiles only what changed. The suite runs under pytest-xdist on every core with --dist worksteal, so the long fits do not queue behind one another: 201 s locally on 4 cores, same results. A PR still renders the reference, which is fast. The executed gallery runs only when a maintainer comments /build-docs, and the run answers on the PR. Main and tags still execute and publish.

test_registry_is_empty_after_a_failure asserted the process-wide registry was empty, which depends on what ran earlier in the same process; under worksteal ordering it saw an entry left by another test. It now clears the registry first and still asserts the failure leaves nothing behind.

The /build-docs trigger is issue_comment, read from main's copy of the workflow, so it works only after this merges.

🤖 Generated with Claude Code

https://claude.ai/code/session_011oXiCXjnYBaoHZs8EwRMRB

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!