fix: clarify single-variant recommendation semantics

[salim4n]

Summary

• Treat one-variant recommendations as baseline measurements instead of comparative wins.
• Add comparisonAvailable and recommendationKind metadata to trainer recommendations/tradeoff explanations.
• Normalize exporter recommendations so legacy single-variant report inputs render as baseline with low confidence.
• Update trainer/exporter docs and IgnitionRAG checklist wording.

Test Plan

• bun test packages/trainer/src/recommendation.test.ts packages/exporters/src/index.test.ts
• bun run lint
• bun run typecheck
• bun run test
• bun run build
• bun run pack:check

Notes

• Multi-variant Markdown still renders Winner: and keeps the existing comparative summary.
• Existing winner fields are kept for compatibility; consumers should use comparisonAvailable before showing winner/best language.