Add sandbox AgentAssertion auth for Responses

[neil-oai]

Summary

This adds opt-in sandbox runtime auth for Python Agents SDK OpenAI Responses calls. When SandboxRunConfig(agent_runtime_auth=OpenAIAgentRuntimeAuthConfig(...)) is provided, or OPENAI_AGENT_RUNTIME_AUTH_ENABLED=true is set, sandbox-backed runs register an agent runtime/task with AuthAPI and attach a verified Authorization: AgentAssertion <token> header to Responses traffic.

The goal is to populate verified agent_runtime_id/task_id on the server-side auth context so Responses analytics can attribute sandbox Agents SDK token/model usage to registered runtimes and tasks. The implementation deliberately does not use arbitrary client attribution headers, rejects explicit user-supplied Authorization headers in ModelSettings.extra_headers, and leaves non-sandbox agents and Chat Completions unchanged.

This covers both Responses transports: HTTP requests receive the AgentAssertion in request extra_headers, and Responses websocket sessions receive it in the websocket handshake headers. This also adds PyNaCl to the encrypt extra/dev dependencies because AuthAPI returns task IDs encrypted with libsodium sealed boxes.

Test plan

• uv run pytest tests/models/test_openai_agent_runtime_auth.py tests/sandbox/test_compatibility_guards.py -q
• make format
• make lint
• make typecheck
• make tests
• uv run pre-commit run --all-files was attempted, but this repo does not have a .pre-commit-config.yaml; the repo-required Makefile verification stack above is clean.

Issue number

N/A

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format
• I've made sure tests pass

Docs note: this is an opt-in attribution/auth plumbing change for an internal rollout path; I left public docs for a follow-up once the rollout surface is agreed.