Persist descriptive Agent thought labels with task history

[marksftw]

Summary

Build on #646 by persisting descriptive Agent thought labels with task history instead of relying only on browser localStorage.

Scope

• First identify the smallest existing persistence path that can store presentation metadata with an Agent task or timeline. Add backend work only if it is required.
• Persist the generated label with a stable reasoning-phase identity so it survives history reloads and follows the task wherever that history is available.
• Keep the generated label separate from raw reasoning and from the Agent's model context. It must not be sent back to the Agent as conversation content.
• Prefer a persisted label when loading history and do not regenerate it.
• Do not generate labels for old, unlabeled history or require a bulk migration of existing local labels.
• Remove persisted labels when the related task or Agent history is deleted.

Done when

• Labels created by Give completed Agent thoughts descriptive labels #646 load with their matching reasoning phases from persisted task history.
• The implementation documents where the labels are stored and why that is the smallest appropriate persistence path.
• Missing or corrupt label data safely falls back to Thought.

Out of scope

• Changing label generation or prompts
• Redesigning tool-call rows
• Grouping reasoning and tools into a shared container