Verify the complete offline narration release journey

## Problem

Focused subsystem checks cannot prove that installed model management, live inference, cancellation, UI delivery and upgrades work together in a real application.

## Proposed change

Run and record the complete acceptance journey using the pinned artifact and actual packaged application plus terminal entrypoints. Reuse existing replay and bundle-smoke infrastructure and retain only regression cases that cover real gaps in earlier slices. Verify every epic acceptance criterion and the PRD traceability below. This issue is the final integration audit, not a bucket for missing product implementation.

## Acceptance Criteria

- [ ] AC1: A clean-user packaged journey verifies normal operation without AI, explicit size/progress download, enablement, restart, disable without deletion, removal and successful reinstall.
- [ ] AC2: With external networking blocked after installation, actual live-replay narration works without credentials across the supported frontend policy and no draft information leaves for inference.
- [ ] AC3: An enabled/disabled replay comparison confirms identical selected cards, ordering, scores and audit decision identities.
- [ ] AC4: The full race/failure matrix covers rapid picks/rescoring, disable/remove during generation, shutdown, interrupted/corrupt download, absent/incompatible runtime, and repeated inference failure without stale output or lost recommendations.
- [ ] AC5: The pinned corpus and supported-hardware measurements meet #454 factuality, repeatability, first-admitted-text, total-time, memory and package-footprint gates; missing evidence remains explicitly blocking.
- [ ] AC6: A merged-implementation epic audit maps every PRD success criterion and open decision to evidence; any missing requirement becomes an additional native child, and the epic is closed only after the final child implementation merges and the audit passes.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** Cross-process, offline and packaged lifecycle failures are not established by isolated tests.

## Dependencies

Direct prerequisites: #466, which transitively gates all delivery slices.