Cache lookups no longer queue behind real subprocess slots. Previously, the runner checked the action cache only after acquiring a concurrency permit, so with max_concurrency=1 a burst of cache hits would serialize one at a time even though each one was a no-op. The cache check now happens before permit acquisition, and re-verifies under the permit so a parallel writer still short-circuits the spawn.
Practically, CI runs that hit the cache heavily (warm builds, mostly-clean rebuilds) finish dramatically faster instead of waiting in line behind spawned actions.
Source: PR #12 commit message describing the runner permit pool fix.