Fix runtime embedder injection for offline tests

[taditya1997]

What breaks

AgentRuntime() always constructs OllamaNomicEmbedder, so runtime-level tests that only need graph/capability wiring can still make a live request to localhost:11434 while writing the inbound memory record. On a clean machine without Ollama running, tests/test_capability_contracts.py::test_every_registered_capability_has_a_worker_and_every_worker_is_registered fails before it reaches the patched executor probe.

Proof before the fix

From a fresh checkout with no Ollama service on localhost:11434:

FAILED tests/test_capability_contracts.py::test_every_registered_capability_has_a_worker_and_every_worker_is_registered
urllib.error.URLError: <urlopen error [Errno 61] Connection refused>

I also added test_agent_runtime_accepts_an_injected_memory_embedder, which initially failed with:

TypeError: AgentRuntime.__init__() got an unexpected keyword argument 'memory_embedder'

Fix

Add an explicit memory_embedder constructor argument to AgentRuntime, defaulting to the existing OllamaNomicEmbedder() for production behavior. The capability-contract probe now uses the public injection seam with DeterministicEmbedder, so it remains offline and verifies the intended worker/registry contract.

Verification after the fix

uv run ruff check .
All checks passed!

uv run pytest -q
356 passed, 1 warning in 38.58s

No .env, credentials, tokens, real messages, or personal data are included.