docs: updates for v0.21.0 release

[seratch]

This pull request updates the v0.21.0 documentation with a provider-neutral testing guide for Agent, Sandbox, Realtime, and Voice workflows.

The guide:

• shows fixed, tool-driven, request-aware, streaming, retry, and workflow-drift recipes for ScriptedModel;
• demonstrates a black-box SandboxAgent workflow with scripted_sandbox_session(), including FIFO matching, detached histories, structured errors, and caller-owned session cleanup;
• documents Python-native ScriptedRealtimeModel message and tool workflows instead of the TypeScript scenario API;
• adds full Voice pipeline coverage with scripted STT, TTS, transcription sessions, workflows, and PCM helpers;
• explains when tests should use the provider-neutral utilities and when they should keep the real provider adapter, transport, sandbox provider, or audio integration; and
• adds navigation and API reference coverage for agents.testing, agents.realtime.testing, agents.voice.testing, and the generated agents.testing submodules.

The organization follows the TypeScript testing guide's shared recipe and boundary concepts while preserving the Python SDK's distinct Realtime and Voice APIs.

Test plan

• make build-docs
• Execute representative Agent, Sandbox, Realtime, Realtime tool, retry-advice, and Voice examples without provider requests.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7c1d011f3f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Disable trace export in provider-neutral recipes

When OPENAI_API_KEY is set and the application uses the default trace processor, this Runner.run() call—and the other Agent, Sandbox, and Realtime recipes below—still enables tracing, so BatchTraceProcessor can make an OpenAI trace-export request despite the guide presenting these workflows as in-memory, provider-neutral tests. This can add network dependence and upload test activity; pass RunConfig(tracing_disabled=True) (and the Realtime equivalent) in the recipes or explicitly document a global tracing-disable setup.

Useful? React with 👍 / 👎.