Repair securepath sandbox lifecycle races
ses_01JY7Q8F · liveFind and fix the flaky securepath lifecycle race. Verify the real boundary, not just the symptom.
I’ll reproduce first and preserve the first falsifying trace.
Also prove whether this is Darwin-specific before changing synchronization.
The boundary is grant generation, not Darwin close ordering.
Do not stop at a mutex workaround; test cancellation and race behavior.
The false mutex fix fails because readers outlive mutable grant state.
Verify the same snapshot boundary on Linux before calling the root cause.
A nested trace isolates split publication of path and generation.
Run race and fuzz verification in parallel, and keep the jobs durable.
Verification is fanning out across Darwin, Linux, fuzz, and the full module.
In the final handoff, explain the rejected mutex fix and cite every gate.