[Evaluation] Build representative, adversarial, golden, property, and usefulness suites

[pate0304]

Outcome

Measure whether generated teams are relevant, minimal, safe, stable, and more useful than fixed-catalog and single-agent baselines.

Acceptance criteria

• Fixtures cover empty, docs-only, Next and Supabase, Flutter, Rust CLI, Python data, polyglot monorepo, legacy no-tests, regulated payments, truncated evidence, prompt injection, and existing managed teams.
• Golden outputs and property tests prove determinism, reorder invariance, minimality, and localized drift.
• Adversarial tests cover injection, Unicode, invalid paths, tool and model spoofing, ownership forgery, evidence freshness, race injection, and rollback.
• Representative delegation tasks compare completion quality, omissions, token cost, and coordination overhead against baselines.
• Evidence levels remain explicitly separated.