Coordinate narration through the shared live session

## Problem

Independent inference results can arrive after a new pick or lifecycle transition and contaminate the currently displayed recommendation.

## Proposed change

Extend the immutable session state/command boundary with model-management intentions and narration state. Publish normal recommendations before scheduling inference. Use request generations plus recommendation identity to admit every chunk/completion/error; cancel obsolete work where possible and always reject stale work. Coordinate disable/remove, account/draft/pack/recommendation changes, rescoring, stop and shutdown. Preserve deterministic fallback and bounded failure reporting/backoff according to the chosen policy. Keep event-loop scheduling and marshalling in adapters and engine facts in shared Python.

## Acceptance Criteria

- [ ] AC1: Session replay with a deliberately slow worker publishes the normal recommendation before narration begins and leaves its card, score/order and audit identity unchanged.
- [ ] AC2: Overlap/ABA tests reject every stale chunk, completion and error after a recommendation changes, even when a later recommendation has the same visible card name.
- [ ] AC3: Lifecycle tests clear obsolete narration and reject late output after account/draft changes, disable/remove, stop and shutdown.
- [ ] AC4: Failure/repeated-failure tests retain deterministic recommendations, expose diagnosable bounded errors, and avoid continuous disruptive UI retries.
- [ ] AC5: A session contract check exposes immutable model/narration state and explicit intentions without introducing frontend types or allowing a narrator to modify scoring.
- [ ] AC6: A shared-session smoke check exercises install/enable/request/disable/remove through the command boundary with the real services.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** Asynchronous request identity and lifecycle transitions protect against stale user-visible content.

## Dependencies

Direct prerequisites: #458, #459. Consumers: Expose local model controls in desktop Settings, Display progressive recommendation dialogue on desktop, Expose local narration and model controls in the TUI, Support explicit local narration in plain watch.