Fix Agent delegation and structured mentions

[spytensor]

Summary

• Make resolved Agent mentions the single source of truth from the composer through delegation, persistence, history restore, and regeneration.
• Render one removable Agent chip instead of retaining duplicate raw @agent text and derived UI state.
• Make delegated-run failures and worker progress truthful: classify truncation and repeated failed tool calls, omit diagnostic transcript paths from parent-model evidence, and keep loop counters out of product status.

Root cause

Agent identity was represented both as prose and as inferred state, while persistence and replay did not consistently carry the resolved identity. That split ownership allowed duplicate presentation and made regeneration re-parse user text.

Validation

• Focused backend tests: 93 passed
• Focused chat UI tests: 146 passed
• pnpm typecheck
• pnpm --filter mozi-ui build
• pnpm verify:public-export