No fork-safety — gunicorn / Celery prefork workers report nothing

[AndresL230]

Summary

recost/_init.py registers no os.register_at_fork hooks. After a fork:

• The patched method-object references are inherited (fine — class attrs).
• The module-level _handle global points at the parent's RecostHandle, but no timer thread is running in the child. Flushes never fire.
• Transport._local._loop references the parent's asyncio loop, which doesn't exist in the child. run_coroutine_threadsafe raises and is swallowed (_transport.py:195-197). All events queue silently and never send.

Symptom: every gunicorn pre-fork or Celery worker started after init() collects metrics into the void. Most production Python deployments use prefork servers — this is the single largest production gap.

Fix

Register os.register_at_fork(after_in_child=_reinit_after_fork) that re-creates the timer thread and transport thread in the child. Alternatively, document a "call init() in the worker's post_fork hook" pattern and refuse to instrument until init() has been called in the current PID.

Files

• recost/_init.py
• recost/_transport.py
• README.md
• tests/test_init.py

Priority

P0 — silent total failure on the most common Python production deployment shape.