fix(evals): swap cart classifier to claude-haiku-4.5 (gpt-5-mini drifted to 75%)

[BSalaeddin]

Why

openai/gpt-5-mini has drifted below the eval gate. No commit touched cartClassifier.ts, openrouter.ts, the eval fixtures, or the scorers since the last green run on 2026-07-14, so the drop from 97.5% -> 75.0% on the 2026-09-14 nightly (#242, and the recurring nightly RED since 2026-08-28) is provider-side model drift — exactly what the nightly gate exists to catch.

The swap

Single source of truth is env.OPENROUTER_MODEL (CI leaves it unset so the code default is what gets tested). Default changed openai/gpt-5-mini -> anthropic/claude-haiku-4.5.

Gated sweep (live CI workflow_dispatch, production classifyCart, temp 0, 40 cases)

Model	Primary-match	Schema-valid	p50	p95
anthropic/claude-haiku-4.5	100.0%	100%	848-878ms	1693-1957ms
openai/gpt-5.1	92.5%	100%	1493ms	3228ms
google/gemini-3.8-flash	90.0%	95%	2388ms	7000ms

Haiku 4.5 wins on every axis and is cheaper than gpt-5.1 ($1.00·$5.00 vs $1.25·$10.00 per M). It is already 100% on the sibling coupons classifier.

Eval-gate compliance

• Green x2 on the shipped model: runs 34889586347 and 34889877905, both exit 0, both 100%.
• Scoreboard row added (evals/SCOREBOARD.md) with the dated numbers and the sweep.
• Regression proof on record: the 75.0% nightly Nightly AI eval failed: cart classifier (2026-09-14) #242 is the gate biting the drifted model.

Prod picks this up on the next deploy. Closes the #242 series once merged and the nightly runs green.

🤖 Generated with Claude Code