Greptile Summary

This PR adds per-turn latency tracing to talk mode, threading timing anchors through the full pipeline from final transcript to speech output (ElevenLabs streaming and system TTS), and also introduces soft-boundary splitting in IncrementalSpeechBuffer to start speaking sooner on long stream chunks.

Key changes:

• TalkTurnLatencyTrace struct captures timestamps for each stage (transcript → process start → chat send/ack → assistant stream → first TTS request/chunk → playback start → system speech start).
• instrumentTTSStreamFirstChunk wraps AsyncThrowingStream<Data, Error> to record the first-chunk timing anchor before forwarding data downstream.
• isSoftBoundary triggers a segment split at any whitespace once the buffer reaches 72 characters, enabling earlier speech start on long chunks.
• TalkSystemSpeechSynthesizer.speak(text:language:onStart:) gains an onStart callback wired to AVSpeechSynthesizerDelegate.didStart.

Issues found:

• The didStart delegate in TalkSystemSpeechSynthesizer does not validate that the firing utterance is the currently-active one. When speech is interrupted and restarted rapidly, a queued Task { @MainActor in } for the old utterance's didStart event can consume the new onStart callback before the new speech actually starts, causing firstSystemSpeechStartAt to be recorded prematurely or not at all.

Confidence Score: 2/5

• Safe to merge for user-facing behavior, but latency data from the system TTS path may be inaccurate due to the didStart callback race condition.
• One real functional correctness issue exists: the didStart handler in TalkSystemSpeechSynthesizer lacks an utterance-identity guard, which can cause the onStart callback to fire for the wrong speech instance when speech is interrupted and restarted quickly. This does not affect playback, only latency measurement accuracy for the system TTS fallback path. The ElevenLabs and incremental-speech tracing paths are correct.
• apps/shared/OpenClawKit/Sources/OpenClawKit/TalkSystemSpeechSynthesizer.swift — the new speechSynthesizer(_:didStart:) delegate method needs an utterance identity check before firing the callback.

_{Last reviewed commit: eb90e00}

Thanks for the solid work here, @ngutman.

Closing this original PR in favor of smaller, reviewable slices that preserve the reliability/UX improvements:

• fix(ios): guard talk TTS callbacks to active utterance #33304 fix(ios): guard talk TTS callbacks to active utterance
• fix(ios): start incremental speech at soft boundaries #33305 fix(ios): start incremental speech at soft boundaries

This keeps risk isolated and makes it easier to validate each behavior independently.