Keep replay tests off the Docker boundary

[webdevtodayjason]

Problem

Startup replay tests could reach the real get_backend factory when a test exercised the append core without replacing append_solo_instance. On a host with Docker, that path could create ainode-vllm-node-solo and collide with a real engine.

Decision

Patch ainode.engine.backends.get_backend in the replay module's autouse fixture so every replay test stays behind a fake backend boundary. Keep the backend command-construction tests meaningful by continuing to instantiate those backends directly and patch their subprocess seam. Add a regression test that calls the real append core and proves the resulting engine is the fixture fake with the requested model config.

Proof

• .venv/bin/pytest -q tests/test_replay_bind_window.py tests/test_browser_federation.py: 50 passed in 7.39s
• .venv/bin/ruff check: All checks passed
• Docker container census after the focused run: no ainode-vllm* containers

Changelog

• Keep startup replay unit tests from creating real vLLM Docker containers.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}