test(bench): ACME ERP enterprise ambiguity + workload recommendation harness

[venkateshsakamuri-lab]

Summary

Adds scripts/enterprise-bench/ — a reproducible multi-schema Postgres ERP (acme_erp) that plants schema ambiguity, seeds business context before questions, creates multi-user access policies, generates pg_stat_statements traffic, then scores DeepSQL on:

1. Schema ambiguity / business-rule adherence (agent answers vs ground truth)
2. Workload analysis + index/performance recommendations
3. Multi-user table/PII access enforcement

Dataset

• Schemas: crm, sales, finance, inventory, hr
• Homonyms: customers/accounts, orders/payment_orders/order_header, products/items/product_master
• ~15k customers, 80k orders, 240k lines, 60k invoices + synthetic workload

Users

analyst@, finance@, hr@, intern@ acme.example — CHAT_EDITOR + English chat policies.

Run

bash scripts/enterprise-bench/setup_and_run.sh
# → /opt/cursor/artifacts/enterprise-bench/VERDICT.md

Sample verdict from this environment is checked in as scripts/enterprise-bench/VERDICT.sample.md.

Headline findings (this run)

• Agent: 4/6 ambiguity/rule checks (revenue + open orders + payment_orders + salary refuse strong; active-only filters weak)
• Access: 6/12 — deny/redact failed; policies saved with empty deniedTables/deniedColumns
• Recommendations: cold-start DROP INDEX …_pkey false positives after pg_stat_statements_reset
• Ambiguity API nearly blind (only pg_stat_statements name similarity)
• Business-rule learner polarity bugs (Exclude … is_deleted=true → require deleted=true)