Closes #77.

The polling thread that refreshes the local evaluation environment document is created in the parent process. Threads do not survive os.fork(), so pre-fork servers (gunicorn, uwsgi, multiprocessing) end up with worker processes whose Python Thread object is intact but whose underlying OS thread is dead. is_alive() correctly reports False, no polls happen, and workers serve a frozen environment-document snapshot from the moment of fork for their entire lifetime.

EnvironmentDataPollingManager now composes a threading.Thread instead of inheriting from it, and registers an os.register_at_fork(after_in_child=...) hook so a fresh polling thread is started in each child. The hook also closes the parent's requests.Session because connection-pool sockets are inherited as shared FDs across fork; reusing them would interleave bytes between processes.

Notes

• The same problem applies to the SSE event stream thread and the analytics processor thread. Perhaps we need a better approach than simply multiplying threads.

Test plan

• New regression test test_polling_manager_keeps_polling_after_fork spawns a multiprocessing.get_context("fork").Process, asserts is_alive=True in the child, and asserts the child's ident differs from the parent's (i.e. it is a fresh thread, not the dead inherited one).
• All existing tests pass (95/95).
• mypy clean.

Notes

• pytest-httpserver added as a dev dependency for the new test fixture.