[codex] Add minimal OpenTelemetry instrumentation

[jooh]

Summary

• add an internal observability package with safe-by-default OpenTelemetry initialization
• instrument CLI, web/SSE, OpenResponses proxy calls, and the fake Responses provider
• add basic chat and upstream metrics plus local collector documentation

Impact

Telemetry remains disabled unless OTLP env vars are configured. When enabled, traces and metrics export over OTLP/gRPC and avoid prompt text, response content, auth tokens, cookies, user/session/turn/message IDs, and other high-cardinality or sensitive values.

Validation

• make tidy
• make fmt
• make imports
• make test
• make all-tests

[jooh]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 873668c498

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]

Don't mark completed web streams as cancelled

When a web chat turn reaches response.completed, turnJob.run emits the terminal done event and returns immediately, so the deferred TurnStream.Close() closes the OpenResponses stream before the trailing SSE [DONE] frame is read. In that normal success path s.done is still false here, so finish(context.Canceled) records the openresponses.stream.consume span as an error/cancellation for successful web turns, corrupting the telemetry this change adds. Track the completed event (or let Close distinguish a completed response from an early abort) before recording cancellation.

Useful? React with 👍 / 👎.