Generate and publish reviewed set-specific synergy enrichment

## Problem

Draft Omen's deterministic semantic roles recognize generic Limited concepts, but profiles do not contain reviewed, set-specific relationships derived from the complete Oracle card pool and a format guide.

A raw model-generated JSON artifact is not a usable review interface. Generating set enrichment also needs observable progress, cost accounting, explicit human approval, reproducible profile publication, and a clear runtime distinction between enhanced and unenhanced sets.

## Goal

Provide a CLI that accepts a set code and Draftsim guide URL, analyzes the guide and complete set card pool through a configured OpenRouter model, summarizes the identified mechanics and principal synergies, and requires Confirm or Cancel before generating an enhanced set profile.

Enhanced profiles expose their status and provenance. Draft Omen offers AI-enhanced contextual suggestions only when the active profile contains validated enhancement data, and every frontend makes that availability clear.

## Scope

- Versioned semantic-enrichment artifact and deterministic validation.
- Secure Draftsim guide acquisition and OpenRouter access.
- Guide, card, and bounded relationship-analysis stages.
- Resumable processing with progress, token, and cost reporting.
- Human-readable findings summary and Confirm/Cancel workflow.
- Set-profile schema and publication support for confirmed enrichment.
- Pool-aware use of validated relationships in draft scoring.
- Enhancement availability in shared application state.
- Availability and mode presentation in terminal and desktop frontends.

## Out of scope

- Calling a model during a live draft.
- Publishing unreviewed model output.
- Letting the model choose numerical Draft Omen score adjustments.
- Exhaustively evaluating every possible card pair.
- Automatically committing, pushing, or opening a pull request.
- Scheduled unattended enrichment or publication.
- Publishing raw guide content or unvalidated model responses.
- Replacing empirical 17Lands evidence with model judgments.

## Child issues

- [ ] Define the semantic-enrichment artifact
- [ ] Add secure guide and OpenRouter clients
- [ ] Extract set mechanics and candidate relationships
- [ ] Add semantic enhancement to set-profile metadata
- [ ] Compile confirmed enrichment into local profile artifacts
- [ ] Add the interactive set-enrichment CLI
- [ ] Apply enhanced relationships to draft scoring
- [ ] Expose enhancement availability through the live session
- [ ] Show enhancement availability in terminal and desktop interfaces

## Acceptance Criteria

- [ ] AC1: The CLI accepts a set code and Draftsim guide URL and analyzes the corresponding frozen card-data artifact without requiring the user to supply individual cards.
- [ ] AC2: The CLI reports guide, card, and relationship-analysis progress, including card counts, percentages, token usage, and running cost.
- [ ] AC3: The review screen summarizes identified mechanics, format findings, principal synergies, uncertainty, validation failures, and total cost without requiring inspection of raw JSON.
- [ ] AC4: Cancel leaves every set profile unchanged, while Confirm generates a deterministic enhanced profile and publication report suitable for a normal pull request.
- [ ] AC5: Enhanced profiles contain validated relationship data and explicit source, model, prompt, input, review, and artifact provenance.
- [ ] AC6: Unenhanced profiles remain loadable and are explicitly treated as not enhanced rather than receiving inferred enhancement.
- [ ] AC7: AI-enhanced contextual suggestions affect ranking only for profiles carrying compatible confirmed enhancement, with deterministic bounded scoring and auditable explanations.
- [ ] AC8: The terminal and desktop interfaces clearly show whether AI-enhanced suggestions are available for the active set.

## Epic classification

- **Overall size:** XL
- **Overall orchestration risk:** High
- **Reason:** This spans external networking, paid resumable processing, versioned profile publication, recommendation behavior, and multiple UI surfaces.