test: add Cortex test suite (77 tests, no LLM calls)

Tests cover:
  - Smoke: /health, /auth/status, /distill/status (test_health.py)
  - Persona validation: path traversal, bad names, list_personas (test_persona.py)
  - Chat API: persona routing, session persistence, error handling (test_api_chat.py)
  - Files API: ALLOWED set enforcement, read/write, missing files (test_api_files.py)
  - Webhooks: NC Talk HMAC accept/reject, Google Chat JWT (test_webhooks.py)
  - Tools: scratch read/write/append/clear, tasks CRUD, cron parser + tools (test_tools.py)
  - Security: path traversal, replay attack, known gaps documented (test_security.py)

All LLM calls mocked — suite runs in ~1.4s.
Run: cd cortex && .venv/bin/pytest

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cortex/tests/conftest.py

@@ -0,0 +1,100 @@
+"""
+Shared fixtures for Cortex test suite.
+
+Key design choices:
+  - All file I/O goes to a tmp_path, never touching personas/inara/ or real sessions.
+  - LLM calls are mocked by default — tests are fast and deterministic.
+  - The 'app' fixture patches settings before importing main, so all modules
+    see the temp directory.
+"""
+
+import json
+import pytest
+import pytest_asyncio
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+import httpx
+from httpx import ASGITransport
+
+
+# ---------------------------------------------------------------------------
+# Temp persona directory
+# ---------------------------------------------------------------------------
+
+@pytest.fixture(scope="session")
+def personas_root(tmp_path_factory) -> Path:
+    """A temp personas/ dir with a minimal 'inara' persona for testing."""
+    root = tmp_path_factory.mktemp("personas")
+    _make_persona(root, "inara", "Inara", "Scott")
+    return root
+
+
+def _make_persona(root: Path, name: str, agent: str, user: str) -> Path:
+    p = root / name
+    p.mkdir(parents=True, exist_ok=True)
+    (p / "IDENTITY.md").write_text(f"# {agent}\nTest identity for {agent}.")
+    (p / "SOUL.md").write_text(f"# Soul\nTest soul for {agent}.")
+    (p / "PROTOCOLS.md").write_text("# Protocols\nBe helpful.")
+    (p / "USER.md").write_text(f"# {user}\nTest user profile.")
+    (p / "HELP.md").write_text("# Help\nTest help content.")
+    (p / "MEMORY_LONG.md").write_text("Not yet populated.")
+    (p / "MEMORY_MID.md").write_text("Not yet populated.")
+    (p / "MEMORY_SHORT.md").write_text("Not yet populated.")
+    (p / "TASKS.json").write_text("[]")
+    (p / "CRONS.json").write_text("[]")
+    (p / "SCRATCH.md").write_text("")
+    (p / "REMINDERS.md").write_text("")
+    (p / "sessions").mkdir()
+    return p
+
+
+# ---------------------------------------------------------------------------
+# App fixture — patches settings before the ASGI app is started
+# ---------------------------------------------------------------------------
+
+@pytest_asyncio.fixture
+async def client(personas_root, tmp_path):
+    """HTTPX async test client against the live ASGI app with patched paths."""
+    import config
+    import persona as persona_mod
+
+    sessions_dir = tmp_path / "sessions"
+    sessions_dir.mkdir()
+
+    with (
+        patch.object(config.settings, "personas_dir", personas_root),
+        patch.object(config.settings, "sessions_dir", sessions_dir),
+        patch("scheduler.start"),   # don't run APScheduler in tests
+        patch("scheduler.stop"),
+    ):
+        # Force persona module to re-read patched settings
+        persona_mod._current.set("inara")
+
+        from main import app
+        async with httpx.AsyncClient(
+            transport=ASGITransport(app=app),
+            base_url="http://test",
+        ) as c:
+            yield c
+
+
+# ---------------------------------------------------------------------------
+# LLM mock
+# ---------------------------------------------------------------------------
+
+@pytest.fixture
+def mock_llm():
+    """
+    Patch complete() at every import site so no real LLM calls are made.
+    Each router does `from llm_client import complete`, creating a local reference.
+    Patching llm_client.complete alone won't intercept those — patch each site.
+    """
+    ret = ("Hello, I am a test response.", "claude")
+    with (
+        patch("routers.chat.complete", new_callable=AsyncMock, return_value=ret),
+        patch("routers.nextcloud_talk.complete", new_callable=AsyncMock, return_value=ret),
+        patch("routers.google_chat.complete", new_callable=AsyncMock, return_value=ret),
+        patch("llm_client.complete", new_callable=AsyncMock, return_value=ret),
+    ):
+        yield