[Performance] Add selected-tab tiered agent prewarming #842

Goal

Prewarm the shared master and a safely selected provider while limiting speculative helper materialization to the selected eligible tab.

Parent tracking issue: #786
Depends on: #841

Scope

• Acquire the process-level master lease when AI functionality is eligible.
• Prewarm a provider only when built-in identity, policy, availability, execution source, command/model semantics, and the exact pool key are known.
• Keep custom agents, WSL providers, and per-tab overrides lazy initially.
• Materialize at most one speculative helper for the selected eligible tab.
• Evict only unused speculative helpers when selection changes.
• Preserve visible, manually opened, restored, dragged, conversation-bearing, and user-triggered connecting helpers.
• Never run command recall, hooks, shell profiles, or other user code during speculative materialization.

Acceptance criteria

• Visiting five unused tabs leaves at most one speculative helper plus the shared master.
• Existing conversation helpers are never evicted by tab selection.
• Policy and settings changes cancel obsolete speculative work.
• Custom and WSL commands do not execute speculatively.
• Prewarm and real helper acquisition use the same exact provider pool key.

Non-goals

• Prewarming every tab.
• Speculatively executing custom commands.