fix(swift): anchor realtime answer-generation span to response.created

[cornelcroi]

Closes #578

Problem

In LangSmith, every realtime response / presenter LLM-call span shows Latency 0.00s, while the non-realtime llm.completion spans show real durations (e.g. 2.60s).

Root cause: the realtime answer generation is created and ended in the same synchronous block at response.done (OpenAIVoiceAssistant.swift:275, OpenAIGroundedVoiceAssistant.swift:265), so endedAt − startedAt ≈ 0. The non-realtime Agent.callModel (Agent.swift:157) opens its llm.completion span around the actual model.complete() call, which is why it measures correctly — this fix makes the realtime path do the same.

The OpenAI Realtime API provides no server-side timing (no timestamps on response.created/response.done, nothing in response.usage/status_details), so latency can only be measured client-side from the response.created → response.done receive-times.

Fix

Backdate the generation's start to the response's created receive-time:

• Tracing — add a defaulted SpanHandle.generation(_:model:input:startedAt:) overload. The default ignores startedAt and stamps now, so every existing conformer (OSLogTracer, test doubles, any consumer custom SpanHandle) keeps compiling unchanged. ProcessedSpan overrides it and threads startedAt through child(...).
• OpenAIVoiceAssistant — record responseStartedAt[id] at .responseCreated, pass it to the response generation, clear in resetTurn.
• OpenAIGroundedVoiceAssistant — record presenterStartedAt when the presenter/direct response is created, pass it to the presenter generation, clear on finish/fail.

The generation now measures the true synthesis latency; tool time still shows as sibling tool.* spans, and the voice.turn root keeps its full duration.

Tests

• RecordingTracer captures per-span start/end and implements the new overload.
• ProcessingTracerTests.generationBackdatesStartedAtToTheGivenTime — the backdate survives finalize/export.
• One test per assistant path (response, presenter, direct): a real gap between response.created and response.done is reflected in the generation's duration. Verified stable across repeated runs.

Full suite green (394 tests). No public breaking change (additive, defaulted).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9dd2b42193

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]

Clear presenter start before awaiting persistence

When a grounded voice session has a store and persist suspends, this defer does not run until after the await persist(...) below. Because Swift actors are reentrant at that await, a new turn can process its presenter/direct response.created and set presenterStartedAt; when the old persist returns, this defer then clears the new turn's timestamp, causing the next presenter generation to fall back to Date() and record ~0 latency again.

Useful? React with 👍 / 👎.

[cornelcroi]

Good catch — fixed in 0fce739. The defer ran at scope exit, i.e. after await persist(...); actor reentrancy at that await let the next turn's presenter response.created set presenterStartedAt during the suspension, and the deferred clear then wiped the new turn's timestamp (→ ~0 latency again). Now cleared synchronously before the await, matching the existing pendingUserText/replyText cleanup that already guards the same reentrancy window.