Establish local narration feasibility and delivery decisions

## Problem

The PRD intentionally leaves model, runtime, quantization, delivery, lifecycle, hardware limits, and streaming safety unresolved. Implementation must be based on measured narration performance, not unrelated extraction benchmarks.

## Proposed change

Run a bounded comparative narration experiment using #453. Evaluate the PRD candidate set without prescribing a winner; consider alternatives only with demonstrated advantage. Record the tested artifact/runtime revisions, prompt and generation settings, quality results, load time, first admitted text latency, completion time, throughput, peak memory, and disk/download size. Compare upstream, controlled mirror, and release-artifact distribution. Resolve every PRD open decision A–M, including supported OS/architecture policy, optional runtime delivery and dependencies, unloading, cache location, compatible updates, and typewriter behavior. This is an investigation/report, not a production downloader or inference service.

## Acceptance Criteria

- [ ] AC1: A reproducible corpus experiment reports factual adherence, naturalness, hallucination failures, and output repeatability for every tested configuration.
- [ ] AC2: The benchmark report records all six PRD performance metrics and first admitted text latency on representative supported Macs, including limited-memory Apple Silicon; unavailable hardware is explicitly an unresolved gate, not a claimed pass.
- [ ] AC3: A decision record resolves PRD A–M and sets numeric acceptance thresholds for interactive latency, generation bounds, memory, and default-download growth before downstream delivery starts.
- [ ] AC4: A streaming experiment demonstrates an admission strategy that catches the agreed malformed-output cases without requiring every response to finish before any text appears; the report distinguishes checks from guarantees against semantic hallucination.
- [ ] AC5: Distribution/licensing and optional-runtime feasibility checks establish a viable pinned artifact source, reproducible identity, notices, supported hardware policy, and disabled-startup behavior.
- [ ] AC6: A recorded go/no-go decision names the evidence and unresolved blockers; downstream implementation is not marked unblocked until the decision supports the PRD requirements.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** Native compatibility, factual quality, streaming safety, and package footprint can invalidate implementation assumptions.

## Dependencies

Direct prerequisites: #453. Consumers: Version the narration prompt and output admission policy, Publish a pinned and licensed narration model artifact.