(fix) tests: fix flaky TestTwistedConnection.test_connection_initialization

[mykaul]

Motivation

TestTwistedConnection::test_connection_initialization is flaky in CI. It was observed failing on PyPy 3.11 + macOS x86 (Rosetta 2) in this CI run, but it can fail on any platform where the Twisted reactor's global running state leaks between tests.

The root cause: TwistedLoop.maybe_start() checks reactor.running to decide whether to spawn a new reactor thread. If a prior test leaves reactor.running == True, maybe_start() becomes a no-op, the reactor.run mock is never called, and the assertion at line 116 fails:

AssertionError: expected call not found.
Expected: run(installSignalHandlers=False)
  Actual: not called.

Change

Patch twisted.internet.reactor.running to False in TestTwistedConnection.setUp(), ensuring maybe_start() always enters the branch that spawns the reactor thread regardless of leaked global state. The patcher is properly stopped in tearDown().

Testing

• All 10 tests in test_twistedreactor.py pass.
• 1000 runs × 6 tests per run = 6000 individual test executions with zero failures (completed in ~16 seconds), confirming the flakiness is resolved.

Batch 1/10: 100 runs OK (600 tests total, 1.8s)
Batch 2/10: 100 runs OK (1200 tests total, 3.1s)
Batch 3/10: 100 runs OK (1800 tests total, 4.6s)
Batch 4/10: 100 runs OK (2400 tests total, 6.2s)
Batch 5/10: 100 runs OK (3000 tests total, 7.6s)
Batch 6/10: 100 runs OK (3600 tests total, 9.5s)
Batch 7/10: 100 runs OK (4200 tests total, 11.6s)
Batch 8/10: 100 runs OK (4800 tests total, 13.0s)
Batch 9/10: 100 runs OK (5400 tests total, 14.5s)
Batch 10/10: 100 runs OK (6000 tests total, 16.2s)
All 10 batches x 100 runs = 1000 runs (6000 individual tests) passed in 16.2s

[mykaul]

reviewed with Opus 4.6 and GPT-5.4

[mykaul]

Failure seen @ #748