feat(studio): add guided scenario evaluation release gate

[shenchengtsi]

Summary

• add persisted scenario, dataset, evaluator, policy, evaluated-version, formal-run, badcase, and publish-audit lifecycle
• guide Studio users through a visible step-by-step evaluation flow with explicit publish feedback and evaluator calibration
• automatically pass immutable scene criteria, hard-failure conditions, and published case criteria into evaluator judgment without exposing expected answers to the Agent under test
• support deterministic required/forbidden regular-expression checks with bounded validation and execution timeout; keep LLM rubric text as optional supplemental guidance
• freeze evaluated code and deployment profiles, then enforce owner-scoped, audited normal, skip, and risk release gates for direct and GitHub delivery

Safety and compatibility

• require Agent management access for scenario data and Admin access for formal evaluation
• protect Agent assets with verified owner ACL, admin handoff, source-fingerprint gates, and server-checked Runtime identity
• hydrate previously persisted evaluator versions from their immutable scene version before formal evaluation
• require every structured LLM judgment to explicitly classify hard-failure status and preserve dynamic hard failures through calibration and release decisions
• limit regular expressions to 512 characters and 20 ms execution; return controlled 422 responses for invalid patterns
• execute only server-generated, safety-checked, signed project snapshots in formal evaluation; support shared rotatable attestation keys
• keep legacy deployment available when scenario storage is not configured
• mark GitHub source synchronization as submitted until Runtime delivery is actually Ready

Verification

• uv run --group dev pre-commit run --all-files
• uv run --frozen pytest -q — 2718 passed, 5 skipped, 36 warnings, 6 subtests passed
• scenario-evaluation focused tests — 135 passed
• npm test — 835 passed
• npm run build
• final blocking review — no Critical or Important findings

Validation boundary

• live Ark planner and real Pi runtime smoke tests remain opt-in and were skipped
• real cloud Runtime, TOS, and AgentKit acceptance still requires target Agent, cases, credentials, and environment data