tests/test_repl.py::test_a_held_scope_is_never_collected fails on aarch64-linux, every run:
> assert _collected_after_churn(state) > before, "the control never fired"
E assert 19 > 19
The control half fails: after made.clear() the scope made on the exited thread is still not finalized. The held half passes.
Reproduces: (import ./. { system = "aarch64-linux"; }).checks.suite (lane nix_2_35), on a libkrun aarch64-linux builder. Three runs at 6747c03, one at d782044, one at b4d564a. Same count, 19, each time.
Guess, not proven: the collector scans stacks and registers conservatively, and on aarch64 a stale copy of the environment pointer survives gc_release_thread() or the churn. The test then reports a real leak where there is only a conservative root.
Activity