Opus 5 migration QA evidence

What was tested

• bun test script/opus5-model-recommendation-audit.test.ts
  • Proves user-facing README, guide, JSONC example, model requirement, and website locale recommendations contain no Opus 4.6 or 4.8 recommendation.
• bun test packages/model-core/src/model-requirements-agents.test.ts packages/model-core/src/model-requirements-categories.test.ts
  • Proves Metis resolves to Opus 5 high then Kimi K3 low and Prometheus resolves to Fable 5 xhigh then Kimi K3 max.
• bun run typecheck
  • Verifies TypeScript across root, scripts, and packages.
• cd packages/web && bun run type-check && bun run build
  • Verifies the website build surface.
• Headless Playwright against the built site at /en
  • Observed Claude Opus 5 High and Claude Fable 5 XHigh in rendered content.
• bun test --reporter=dot
  • Final full repository gate.

What was observed

• The stale-recommendation audit was captured RED before the migration, then GREEN after it: 1 pass, 0 fail.
• Focused role and capability contracts passed after migration.
• Final full gate: 12,151 pass, 3 skip, 0 fail across 1,555 files.
• The built English website page exposed both migrated model labels.

Why this is enough

The audit protects the user-facing recommendation surface. The model-core tests protect exact Metis and Prometheus selection order and variants. Typecheck, website build, browserless render verification, and the full suite cover integration and packaged behavior.

What was omitted

Raw environment dumps, provider credentials, and generated logs are omitted. Temporary local web servers used for browserless QA were terminated after verification.