Fix scenario runner event loop freeze and initial_unpark timeout

[deepmasq]

Summary

• Event loop freeze fix: In scenario mode (running_test_scenario), tool calls are now routed through scenario_generate_tool_result_via_model as fire-and-forget instead of going through local tool handlers. The local handlers use gql/aiohttp sessions whose cleanup (__aexit__) blocks the event loop, causing the scenario runner to hang indefinitely after the first tool call.

• Initial unpark timeout: _completed_initial_unpark now has a 30-second timeout. Previously, scheduler events (kanban sync, etc.) kept did_anything=True on every unpark cycle, so the flag never got set to True, blocking the scenario runner forever in the "WAIT for bot to complete initial unpark" loop.

• Background cloudtool_post_result: Moved cloudtool_post_result in _local_tool_call to an asyncio.create_task background task to prevent blocking the unpark loop.

• Fire-and-forget pattern in ckit_scenario: scenario_generate_tool_result_via_model now wraps the GQL mutation in asyncio.create_task and immediately raises AlreadyFakedResult, so the caller never awaits the HTTP response.

Test plan

• Ran 10 Karen scenarios on staging with grok-4-1-fast-reasoning -- 8/10 passed (2 failures were model quality issues, not infra)
• Verify no regressions in non-scenario bot execution (cloudtool_post_result background task change)
• Verify version bump 1.2.196 deploys cleanly

🤖 Generated with Claude Code