fix(rate-limit): fail open on non-RedisError connection failures

[arena-ai-coding-agent]

Problem

The docker-integration CI job ("Docker build and integration tests") failed with:

FAILED tests/test_integration.py::test_complete_batch_lifecycle - RuntimeError: Event loop is closed
FAILED tests/test_integration.py::test_organization_isolation - RuntimeError: Event loop is closed

Root cause

Both failing tests POST /api/batches (no trailing slash). FastAPI 307-redirects that to /api/batches/, and Starlette's TestClient follows the redirect as a second internal ASGI call that runs on a fresh event loop.

The rate-limit middleware (backend/app/core/rate_limit.py) holds a single long-lived async Redis client created once at import time. When the second call's Redis operation reuses a pooled connection from the first event loop, redis-py's asyncio transport detects the mismatch and raises a bare RuntimeError("Event loop is closed") instead of a redis.RedisError subclass.

The middleware only caught redis.RedisError to implement its documented "fail open if Redis is unreachable" behavior, so this RuntimeError escaped as an unhandled exception, turning into a 500 and failing both integration tests.

Fix

• Broadened the fail-open exception handling in create_rate_limit_middleware to also catch OSError/RuntimeError (transport-level failures and event-loop mismatches), matching the module's stated contract that Redis problems must never break a request.
• Added a regression test (test_middleware_fails_open_when_redis_is_unreachable) covering all three failure shapes: redis.ConnectionError, RuntimeError, and OSError.

Verification

• Reproduced the exact CI traceback locally with the pinned requirements.lock versions and a real Redis server (byte-for-byte match of the reported error).
• Confirmed test_complete_batch_lifecycle and test_organization_isolation now pass.
• New regression test fails on the old code and passes with the fix.
• flake8, mypy app, and the broader pytest suite show no regressions from this change.