feat: make OpenTag fast and restore rich UI

[tylerslaton]

Summary

This PR is stacked on #11 and contains only the follow-up agent-speed, rich-UI, and cleanup work.

• Make OpenTag triage-first with low reasoning and verbosity by default.
• Restore strict rich rendering across Slack and Microsoft Teams.
• Remove dead local integrations while retaining optional remote Notion support.

It depends on CopilotKit/Intelligence#612 for deeply nested Channel render content.

Why

The deep-research workflow planned and delegated routine requests, which made ordinary @kite replies slow. OpenTag's render tools could also produce valid Block Kit and Adaptive Card UI that the current Intelligence depth validator rejected. Several local-only paths were no longer part of the managed launch architecture.

How

• Validate OPENAI_REASONING_EFFORT and OPENAI_VERBOSITY, default both to low, disable the automatic general-purpose subagent and execute, cap recursion at 25, and replace nested research with direct Tavily search.
• Add a strict render-tool mapping with no duplicate prose, register durable incident actions, and cover Slack Block Kit, Teams Adaptive Cards, image uploads, render failures, and postFile behavior.
• Remove the Notion sidecar, browser token grabber, Telegram placeholder, nested-research modules, and redundant exports. Require both NOTION_MCP_URL and NOTION_MCP_AUTH_TOKEN for optional remote Notion.
• Preserve the single managed open-tag Channel, Python AG-UI identifier, two-service Railway topology, historical documents, and exact Channels/Runtime canaries.

Validation: frozen Node and Python installs, TypeScript checks, 139 Vitest tests, 39 Python tests, and Railway graph evaluation all pass. No live Slack or production smoke test was run.

[linear]

OSS-613