Run bounded local narration in a reusable worker

## Problem

Native inference can block the interface, reload expensive state on every pick, or fail independently of the Python application.

## Proposed change

Implement the #454 local runtime boundary consuming #452 facts through #455 prompt/admission logic. Use the selected worker/isolation strategy, lazy load and reuse, deterministic-leaning settings, bounded output/time, cooperative cancellation where supported, and the selected unload/shutdown policy. Accept a verified artifact location; do not download from the inference path. Keep frontend imports and scheduling policy outside the service. Native faults must degrade through the selected isolation boundary rather than terminate recommendation delivery.

## Acceptance Criteria

- [ ] AC1: A real-runtime offline smoke check produces admitted narration using the pinned local artifact without API keys or external connections.
- [ ] AC2: A repeated-request instrumented check shows lazy first load and reuse for later picks; disabled application startup neither imports/initializes heavy runtime behavior nor loads weights as defined by #454.
- [ ] AC3: Cancellation, timeout, malformed-output, and worker/native-failure checks return bounded failure outcomes without bringing down the recommendation process.
- [ ] AC4: A repeated-corpus run meets the #454 quality, variability, generation and resource limits.
- [ ] AC5: Disable/shutdown/unload checks release worker resources according to policy without hanging the application.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** Native dependency failures, resource ownership, streaming, and cancellation need isolation and real-runtime proof.

## Dependencies

Direct prerequisites: #455. Consumer: Coordinate narration through the shared live session. Packaging consumer: Package the optional local inference component. Artifact identity is already fixed by #454; production publication is owned by #456.