A race condition at process startup could let two threads be
handed the same per-thread record, because records were allocated
from an Apache memory pool shared by every thread without a lock.
Threads sharing a record also shared the mod_wsgi.thread_id value
and, for named application groups, the same Python thread state,
so one request's context (such as session data) could leak into
another concurrent request. On Python 3.14 and later the shared
thread state could instead crash the worker process.