Publish a bounded recommendation narration fact contract

## Problem

The narrator needs authoritative supporting facts, not arbitrary application state or an opportunity to recompute the pick.

## Proposed change

Define a versioned immutable, frontend-neutral narration input projected from the chosen recommendation and existing `PickRationale`/`PickReason` data. Include only relevant available metrics, color/commitment information, supported reasons, and a bounded set of alternatives. Distinguish missing data and uncertainty from zero or certainty. Preserve the existing deterministic explanation as fallback. Expected boundaries: `pickengine.py` data consumers, `session.py`, a focused fact-contract module, and corresponding contract/session/audit tests. Do not alter scoring or invent new confidence thresholds.

## Acceptance Criteria

- [ ] AC1: A projection check shows the selected card, available values, material reasons, and supported alternative comparisons faithfully reflect engine output.
- [ ] AC2: Missing-statistics and unsupported-synergy fixtures produce no fabricated fields or reasons; verified by fact-contract boundary tests.
- [ ] AC3: Input-size boundary tests show deterministic bounds on alternatives and text while retaining the recommended card and its essential reasons.
- [ ] AC4: Enabled-input projection versus baseline replay leaves ranking, scores, and audit evaluation identity identical; verified by session/audit comparison tests.
- [ ] AC5: A serialization/import check verifies immutable versioned inputs contain no frontend objects, external-service credentials, or arbitrary application state.

## Implementation classification

- **Estimated size:** M
- **Orchestration risk:** High
- **Reason:** This is the public factual boundary between authoritative decisions and untrusted narration.

## Dependencies

Direct prerequisites: None. Consumers: Define a fixed narration evaluation corpus and subsequent narration logic.