fix: a background record could be killed at exit; add flush()

[hugo8xx]

From the outside review, which described the failure exactly.

The bug

threading.Thread(target=_send, daemon=True, name="khwan-record").start()

record(background=True) sends on a daemon thread, and the interpreter does not wait for daemons. A CLI, a serverless handler, or any script that ends soon after its last turn has that write killed mid-flight — no exception anywhere, the turn simply never learned.

It reads as unreliable memory rather than as a missing call, which is the worst way for it to fail.

flush()

kw.record(turn, answer, background=True)
kw.flush()          # → how many were in flight

The timeout is a total budget, and running out returns rather than raises: a record that did not land costs one turn of learning, and turning that into an exception at exit would be worse than the thing it reports.

An atexit hook calls it anyway, bounded to five seconds, so a developer who never read the docstring loses latency instead of the turn. Threads stay daemons on purpose — a hung request must not be able to wedge the process shut.

The test that matters

A real subprocess that exits immediately after a background record, with no flush() and no sleep, exactly as a CLI would. Without the hook:

AssertionError: the process exited before the background write landed —
the failure flush() exists to prevent

✓ flush: returns only once the in-flight write has landed
✓ flush: a hung write is given up on, not raised and not waited out
✓ flush: nothing in flight → nothing to wait for
✓ exit: a script that ends immediately still lands its last write

All three existing suites still pass.

0.3.1

AsyncKhwan already had this — aclose() awaits pending tasks, since asyncio holds only a weak reference to a task. The sync client is now level with it.