Add live LLM media-matrix runner and aggregate evaluator

[terisuke]

Problem

cmd/scenario/media_matrix defines the planned cross-media cases, but there is not yet a live runner that executes those cases one by one against Evo X2 Tailnet and aggregates comparable metrics. Running every case manually is error-prone, expensive, and hard to resume after a long failure.

This is a child implementation issue for #40.

Scope

• Consume cmd/scenario/media_matrix outputs.
• Run the six live cases one by one against Evo X2 Tailnet OpenAI-compatible API.
• Reuse source selectors proven by [Phase B2] Generalize SourceFetcher beyond note.com (Zenn, Qiita, RSS, HTML) #22/Implement persona and format aware question templates #56.
• Record:
  • endpoint and runtime path,
  • model per phase,
  • elapsed seconds,
  • first-token or first-chunk latency when streaming is available,
  • style score,
  • runes,
  • verification result,
  • failed metrics,
  • output path,
  • scenario case id.
• Support resume/skip so a failed long run does not restart the whole matrix.
• Write one aggregate JSON report and one Markdown report for Stabilize Tailnet Evo X2 draft quality and runtime metrics #40.

Acceptance criteria

• Offline mode remains the default.
• Live mode requires explicit env vars and refuses to run against workstation-local fallback unless fallback mode is explicitly requested.
• All six cases produce comparable rows.
• Note, Qiita, Zenn, and company-blog rows are clearly marked as the final publishing-target comparison.
• Failures are grouped by source, persona, format, target length, verifier, and runtime path.
• The aggregate report links each generated draft and verification artifact.

Dependencies

• Stabilize Tailnet Evo X2 draft quality and runtime metrics #40 owns the runtime-quality target.
• [Phase C1] Replace JSON store with SQLite-backed schema (extends #14) #26 should land before the full expensive run if we want the results to become product history instead of loose artifacts.

[terisuke]

Closed by PR #60 merged into develop. cmd/scenario/live_media_matrix now reads cmd/scenario/media_matrix outputs, writes planned aggregate JSON/Markdown by default, and only performs live Evo X2 runs when RUN_LIVE_MEDIA_MATRIX=1 or make scenario-media-matrix-live is used. #40 remains open for actual live Evo X2 quality results.