Demo: block unverifiable AI eval claim

[AzurLiu]

This PR is the public Falsiflow launch demo.

It intentionally starts with placeholder AI eval evidence so the claim gate fails with claim_check_blocked. The follow-up commit replaces falsiflow_ai_eval/evidence.csv with source-backed evidence so the same PR turns green with claim_check_ready.

Claim under review:

candidate_model improves answer quality over baseline_model on the pinned claims_eval_v2026_05_26 task set.

Canonical evidence:

• Blocked run: https://github.com/AzurLiu/falsiflow/actions/runs/26708459093
• Ready run: https://github.com/AzurLiu/falsiflow/actions/runs/26708472653

Expected blocked behavior:

• AI Eval Claim Gate Demo fails strict CI
• status: claim_check_blocked
• blocking stage: gate_evidence
• repair actions ask for source-backed eval evidence

Expected ready behavior:

• AI Eval Claim Gate Demo passes
• status: claim_check_ready
• verification: bundle_verified

This PR is deliberately not about proving the model is good. It demonstrates that unverifiable eval claims do not pass CI until the repository supplies pinned, source-backed evidence.