Close the latent LLM-booking loophole in research crews

[atc964]

Summary

Every channel crew (branding, mobile, CTV, performance) and the social crew placed an execution agent in the crew's agents=[] list, wired with live OpenDirect order-writing tools (CreateOrderTool, CreateLineTool, ReserveLineTool, BookLineTool), even though no task was ever assigned to that agent. All real booking already goes through the deterministic DealBookingFlow / MultiSellerOrchestrator path.

In crewai's hierarchical process, the manager LLM may delegate a task to any agent present in agents=[], whether or not it was ever assigned one. So an idle, live-tooled execution agent sitting in a research crew was one delegation away from writing a real order outside the deterministic booking path. The existing LLM-free CI guard (tests/unit/test_booking_path_llm_free.py) only proved the booking path can't reach an LLM; it said nothing about a crew's own manager LLM reaching booking, since it never inspected crew tool inventories.

Changes

1. src/ad_buyer/crews/channel_crews.py: removed execution_agent (and its _create_execution_tools wiring) from _build_channel_crew and from create_social_crew. Research/recommendation crews now contain only agents that actually have a task assigned. create_execution_agent itself is untouched — it's still exercised directly by its own unit tests (test_agents.py, test_agent_hierarchy.py, test_llm_pricing_guardrails.py), so it wasn't left dangling and didn't need deletion.

2. tests/unit/test_booking_path_llm_free.py: added TestNoCrewAgentCarriesAnOrderWritingTool, which builds every production crew (branding, mobile, CTV, performance, social, portfolio) the way the app does and asserts no agent — manager included — carries CreateOrderTool/CreateLineTool/ReserveLineTool/BookLineTool. A second control test cross-checks the tool classes against the crew module's own _create_execution_tools bundle so the guard can't silently go stale if an execution tool is renamed or added. I verified locally that reverting the crew fix makes this test fail with a clear per-agent message.

3. tests/unit/test_crew_factories.py: updated the branding/mobile/CTV/performance agent-count assertions from 2 to 1, reflecting the removed execution agent.

Test plan

• uv run ruff check clean on all changed files (3 pre-existing, unrelated ruff findings in demo/ and patches/ confirmed present on origin/main before this change)
• Full tests/unit suite: 3502 passed, 1 skipped
• Manually reverted the crew fix and confirmed the new guard test fails with a clear message identifying the offending crew, agent role, and tool class; restored the fix and reran green

[Sirajmx]

Verified: 3612 passed, 83 skipped, 0 failed (+2 from the new guard tests). Confirmed the real
risk this closes: every research crew carried a live, order-writing execution agent
(CreateOrderTool/CreateLineTool/ReserveLineTool/BookLineTool) with no task ever assigned
to it — one manager-LLM delegation away from a real booking outside DealBookingFlow. The new
guard test (TestNoCrewAgentCarriesAnOrderWritingTool) is a real regression check, not just a
snapshot — it inspects actual crew tool inventories rather than asserting agent counts.