feat(mcp): assemble the 8-tool toolkit with capability gating, proven via real dispatch

Also fixes async-test DB isolation: test_mcp_toolkit.py drives real MCP
dispatch, which runs tool functions' ORM calls on a sync_to_async worker
thread outside pytest-django's per-test atomic transaction, so writes
committed there weren't rolled back and leaked into later tests. Sets
DJANGO_ALLOW_ASYNC_UNSAFE for the async ORM setup calls and flushes the
Printer table after this file's tests via conftest.py.

tests/conftest.py

@@ -0,0 +1,30 @@
+"""Suite-wide test fixtures.
+
+test_mcp_toolkit.py drives real MCP dispatch through `mcp.client.Client`,
+which routes each tool call's synchronous Django ORM work through
+`sync_to_async` onto its own dedicated worker thread rather than the test's
+own thread. pytest-django's `db` fixture isolates each test by wrapping it in
+an outer atomic transaction on *its* thread's connection and rolling that
+back afterwards -- but the worker thread opens a separate, thread-local
+connection that was never part of that transaction, so writes made through
+it commit for real and survive into later tests (e.g. Printer rows from
+test_mcp_toolkit.py leaking into test_resolve_printer_device.py's counts).
+
+This fixture flushes bambu_run's Printer table after each test in that file
+specifically, restoring isolation without touching the test file itself
+(which mirrors mcp-gate's proven end-to-end test pattern verbatim).
+"""
+from __future__ import annotations
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def _isolate_async_mcp_dispatch_writes(request, django_db_blocker):
+    yield
+    if request.node.module.__name__ != "tests.test_mcp_toolkit":
+        return
+    with django_db_blocker.unblock():
+        from bambu_run.models import Printer
+
+        Printer.objects.all().delete()