feat(agent): track terminal turn outcomes

[zhoum-hub]

Summary

• capture the exact workspace UI mode (os or agent) on Agent create/send requests
• persist submit metadata with the canonical Turn submission envelope
• emit agent.turn_completed, agent.turn_failed, and agent.turn_cancelled from canonical Host root-turn settlements
• exclude child sessions, Goal/provider-initiated turns, imported history, and legacy submissions without a valid mode
• keep raw prompts, responses, paths, workspace IDs, and error messages out of terminal analytics

Why

The local desktop did not have a stable, canonical event for whether a user-submitted Agent turn completed, failed, or was cancelled. The renderer knows the UI mode, while the daemon owns the authoritative terminal outcome; without durable submission provenance, choosing either producer alone would lose settlements or create mode=null data.

This change joins those facts by persisting submission-time mode and reporting only after the canonical Turn settlement commits.

Product impact

• every newly reported terminal event has mode=os or mode=agent; invalid or missing modes are skipped rather than guessed
• completed measures technical Turn completion, not whether the user's goal was semantically satisfied
• existing DataFinder dashboards are unchanged until they explicitly adopt the new event family
• startup-reconciled interruptions remain distinguishable through startup_reconciled=true

Validation

• pnpm check:changed — 28 lanes passed after implementation
• pre-push pnpm check:changed -- --push-ready — 32 lanes passed
• focused desktop adapter tests for create/send mode propagation
• focused Host, store-sqlite, tuttid Agent/API, OpenAPI generation, and TypeScript type checks

Post-release DataFinder checks

• verify mode contains only os and agent, with zero null
• verify a turn_id appears in at most one event across the three terminal events
• verify all rows have turn_origin=user_prompt and is_child_session=false