test(eval): add SimilarWeb budget stop-loss cases

[eddiearc]

Summary

• add a deterministic high-cost SimilarSites hard-budget regression: the plan must remain over_budget rather than become runnable
• add the matching natural-language user request to the evaluation suite
• document that the scenarios are sanitized feedback-derived regressions; no customer identifiers or raw prompts are stored

Validation

• npm run build
• npm test (169 tests)
• npm run eval (24/24 scripted scenarios)

The harness remains local-only and poisons AISA_API_KEY, so it cannot make a paid AIsa gateway request.