feat: add verification queue for ambiguous AI-Radar signals

[Felix8686]

Why

Recent production alerts showed that page-level changes and catalog observations can still become noisy or misleading high-value pushes. At the same time, dropping ambiguous changes risks losing genuinely valuable free-credit, pricing, API, or release intelligence.

What changed

• Generic web pages now emit two observation types: page_change and exact linked_page candidates.
• Existing same-origin content links are baselined before they can be treated as new, preventing historical-link floods after upgrade.
• Ambiguous observations are routed into a D1 verification_queue instead of being directly classified/pushed.
• Verification fetches the exact changed/linked page, requires a concrete event kind, then reuses the normal classifier + AI quality gate.
• Queue retries every 12 hours, up to 3 attempts, and expires after 72 hours.
• Plain discovered_model catalog observations are retained only as low-priority internal data and can never directly trigger P1.
• Generic xxx changed observations can never directly reach Telegram.
• Migration 0007 also clears invalid expires_at values from model-event rows and forces one safe full generic-web fetch to establish link baselines.
• Added regression tests for page-change gating, linked-page verification candidates, and old Hugging Face catalog observations.

Safety / deployment order

1. Run npm run check on this branch.
2. Review migration 0007_verification_queue.sql.
3. Merge into ai-radar-mvp only if all checks pass.
4. Apply remote D1 migration 0007.
5. Deploy Worker.
6. Trigger one harvest and confirm no historical-link notification flood.

main is unchanged and is not the target branch.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	ai-radar	349cefb	Commit Preview URL Branch Preview URL	Sep 03 2026, 02:58 PM