Test and evaluate

Guardrails, consistency, refusals, and latency budgets for OpenBot.

• Bridge unit tests under bridge/test/*.test.mjs (connectors, providers, computer fakes).
• Voice bench: python3 ops/voice/bench.py with warm budget.
• Approval deny paths must fail closed.
• Do not weaken safety-gate allowlists to make evals pass.
• Measure chat latency per provider timeout table; Ollama has a separate budget.