Package the optional local inference component

## Problem

Source execution does not establish that native inference dependencies work in installed or frozen applications, or that the default download remains approximately unchanged.

## Proposed change

Implement #454's optional-runtime delivery decision in packaging/release configuration. Keep weights out of the normal package and bundle, preserve no-AI startup with absent optional components, explicitly declare required Python/native/runtime resources, and resolve models through application-managed storage. Audit `pyproject.toml`, `pysidedeploy.*.spec`, and existing package/bundle smoke paths. Frontend tickets own their new QML registration; this slice owns native component packaging, not dialogue implementation.

## Acceptance Criteria

- [ ] AC1: A package/bundle content and size comparison confirms no model weights and default-download growth within the #454 threshold.
- [ ] AC2: Clean installed/frozen application launches work without optional runtime/model files and retain normal recommendations.
- [ ] AC3: A packaged real-model smoke check demonstrates runtime loading and local generation from the managed installation, not a developer-only path.
- [ ] AC4: Supported-platform checks verify native dependency/resource inclusion and licensing; unavailable platform checks are recorded as outstanding release gates rather than passed.
- [ ] AC5: An upgrade/compatibility smoke check preserves user model storage and fails gracefully when runtime/artifact versions are incompatible.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** Optional native dependencies and frozen builds introduce platform-specific failure and footprint risks.

## Dependencies

Direct prerequisites: #456, #459. Documentation consumer: Document local narration installation and troubleshooting.