fix(world-lab): make swarm runs observable and recover planner failures

[christopherjnelson]

Summary

• Expose the execution mode in World Setup and the World Lab header, with visible swarm strategy, action, fallback, pressure, and provider diagnostics.
• Constrain Agent Zero's OpenRouter response to opaque worker and target choices; the server materializes authoritative IDs and directive lifetimes. Forward Zero's selected reasoning profile and show specific safe planner rejection reasons.
• If Zero planning fails before a worker has an unexpired directive, choose an engine-legal deterministic local action without spending a Jev call.
• Distinguish provider-reported cost from admission exposure. TypeSafe Jev monetary cost remains unknown; its configured attempt reserve is a budget hold, not measured spend.
• Distinguish Jev unknown-choice failures from inconsistent probability distributions, preserve TypeSafe token usage on rejected decisions, and show safe worker failure details in the UI and read-only diagnostic.
• Make the fixed-height bottom Swarm activity panel scrollable, including keyboard focus, with browser regression coverage.

Verification

The repository owner reported pnpm validate and pnpm test:e2e green after the final local commit. No paid provider test was run by the coding agent.

Scope

Legacy mode and deterministic world rules are unchanged. Raw provider requests, responses, and credentials are not added to safe telemetry.