Deliver Slack research sources and complete tool continuations

[jerelvelarde]

Slack research now posts native source buttons directly from validated Exa URLs, makes search failures visible, and retains the evidence for the agent's answer. A channel-only wrapper creates a fresh inner agent for each tool-continuation run, preserving transcript, state and cancellation while avoiding the pinned runtime's already-running guard.

The app uses managed Slack's default hidden tool progress. With the optional progress stream enabled, live runs delivered cards but dropped final prose and left the working indicator visible. The default configuration passed the full live flow. The search tool also tells the agent that returned sources have already been posted as native cards.

Validation at bb4e0dc:

• Root typecheck and verification pass: 50 root tests, 22 channel tests, 34 web tests. Independent source and regression reviews have no remaining findings. Tests reproduce the real BuiltInAgent continuation guard and cover transcript/state, clone, cancellation and stream cleanup; removing cleanup makes its regression fail.
• Fresh managed Slack/OpenAI/Exa run passed: thread-aware incident card, native source cards, an AWS source button opening its PDF, final explanation with two next checks, and cleared working indicator. No closed-path or handler-failure events were captured.
• Current main's three-app layout, web persistence, React Native, credits guidance and CI/Auth0 cleanup are preserved. No removed directories are restored; shared web/mobile runtime behavior and dependency pins are unchanged.

Live OpenRouter calls remain unverified. Temporary diagnostic logging used during investigation is excluded from this PR.