Buffer should flush() on __del__

Since it doesn't, data loss can occur, especially if you're using partd from multiprocessing workers where the workers aren't responsible for dropping the partd buffer.

Alternatively, the documentation should really mention that users MUST call flush() or they can lose data.

Start by locating the Buffer implementation and its existing flush() path; no file or test is named in the issue. Reproduce the multiprocessing-worker cleanup case, then verify either that destruction flushes pending data or that the documentation clearly requires callers to flush.

Written by the indexing model from the issue text.