Nightly AI eval failed: cart classifier (2026-09-09)

[github-actions]

The nightly cart-classifier eval run failed on 2026-09-09 13:41:10 UTC. This can mean the OpenRouter model degraded/drifted, the OPENROUTER_API_KEY secret expired or was revoked, or a real regression landed in cartClassifier/openrouter. See the failed run: https://github.com/DevinoSolutions/caramel/actions/runs/34358185668

[BSalaeddin]

Fixed by #243: root cause was provider-side drift of openai/gpt-5-mini (97.5% on 2026-07-14 -> 75.0% now, no code change to the classifier/prompt/fixtures/scorers since). Swapped the default model to anthropic/claude-haiku-4.5, which scored 100% primary-match / 100% schema-valid in a gated CI sweep (beating gpt-5.1 92.5% and gemini-3.8-flash 90%), green x2 + a fresh 100% run on merged main (run 34892596564). Scoreboard row added in evals/SCOREBOARD.md.