Too fragile for CI — CPU contention on shared runners makes
process_time() measurements unreliable.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

The pragma: no cover was on the inner `if memory://` check but the
uncovered line was the `return` in the outer `if not streams` block.
Move the pragma to the outer if so the whole branch is excluded — on
real Redis, xread blocks for 60s so this branch is never taken in tests.

Also add tzdata as a conditional Windows dependency so zoneinfo can
find IANA timezone names (Windows doesn't have /usr/share/zoneinfo).

Relates to #339

…servers

On Windows, select.select() only works with actual sockets — if the
context manager exits and server_close() invalidates the socket while
serve_forever is still selecting on it, you get WinError 10038. Calling
shutdown() first signals serve_forever to exit its loop cleanly.

Relates to #339

- Calibrate test_task_duration_is_measured against actual platform sleep
  behavior instead of hardcoding 0.1-0.2s bounds (asyncio.sleep can
  return early on Windows)
- Widen redelivery_timeout in test_workers_with_same_redelivery_timeout
  from 200ms to 1s — gives lease renewal more margin on Windows where
  event loop timer resolution is ~15ms, without changing test duration
- Add test_worker_drains_active_tasks_on_shutdown to reliably exercise
  the finally block drain path (worker.py:608) using event handshake
- Set Windows 3.11 coverage threshold to 98% matching non-Windows 3.11
  (known coverage.py reporting instability on 3.11)

Relates to #339

Co-Authored-By: Claude Opus 4.6 <[email protected]>

The redelivery counter test used 50ms timeout with 75ms sleep (25ms margin),
and the abandoned worker test used 100ms/125ms. Windows ~15ms timer resolution
makes these too tight, causing worker crashes in CI. Widened both to 200ms/250ms.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

redis-py 7.2.0 (released Feb 16) broke async cluster initialization —
their internal NodesManager passes `lib_name` to ClusterNode.__init__
which now raises a deprecation error via the new DriverInfo API (#3880).
We'll unpin once they fix it.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

The cluster include entries need to match the base matrix's
redis-py-version exactly, otherwise GHA treats them as new combinations
and they lose cov-threshold from the python-version includes.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>