Minimal fix for issue #614 — Bad file descriptor

2-line change. Adds if self.is_closed: return at the top of LibevConnection.handle_read().

Root Cause

Thread race between close() and handle_read():

• close() calls self._socket.close() synchronously (from any thread)
• Watcher stop is asynchronous (batched in _loop_will_run(), next loop iteration)
• If the reactor has already dispatched handle_read() for that connection before the watcher is stopped, recv() hits a closed FD → EBADF

Thread A (app):    close() → _socket.close()     [immediate]
Thread B (reactor): handle_read() → recv()       [EBADF - socket already closed]
                    ...
                    _loop_will_run() → watcher.stop()  [too late]

The Fix

 def handle_read(self, watcher, revents, errno=None):
+    if self.is_closed:
+        return
     if revents & libev.EV_ERROR:

Why this is safe

• is_closed is set under self.lock in close() before _socket.close() is called
• Reading a boolean attribute is atomic in CPython (GIL guarantees)
• If is_closed is True, the socket is either already closed or about to be — no useful work can be done

Reproducer

100% reproduction rate: https://github.com/scylladb/scylla-dtest/pull/7079

Note on existing fix branches

• fix/factory-close-race-614 — fixes factory returning dead connections (different race)
• fix/asyncio-close-race-614 — fixes AsyncioConnection race (different reactor)

Neither adds the handle_read guard for LibevConnection. This PR is complementary.

How to test with SCT/Hydra

# In SCT requirements, point python-driver to this branch:
pip install git+https://github.com/scylladb/python-driver.git@fruch/fix-614-handle-read-guard

# Or in hydra, override the driver version in the test config

@vponomaryov — could you try building hydra with this branch and check if it fixes the EBADF reports in longevity runs? This is the minimal possible fix (2 lines). If it works, we can merge this independently of the other #614 fix branches.

Closes #614