feat(telemetry): trace the end-of-turn wait and the turn handoff

[davidzhao]

What

eou_detection only covered the turn-detector inference, so the endpointing delay itself (often seconds) showed up as an empty gap between user_turn and agent_turn. This makes the wait, the user hook, and the speech queue visible.

How

eou_wait span, one per user turn, child of user_turn:

• back-dated to last_speaking_time, so the bar starts where the user stopped talking;
• ends on the turn decision with lk.eou.outcome (committed / user_resumed / dropped), lk.eou.wait_duration, lk.eou.rearm_count, the final lk.eou.endpointing_delay, and the last lk.eou.source;
• a later trigger for the same turn (late STT final, another VAD end of speech) re-arms the wait and records a rearmed event instead of starting a new span, so the bar stays whole and late transcripts are readable off it;
• a rejected commit (min_words, realtime backchannel) records a not_committed event and keeps waiting; resumed speech ends the span at the VAD/STT speech start; teardown drops it.
• resumed speech ends the span at the VAD/STT speech start, kept as a user_resumed event; when that start is reported after the fact and a detection or re-arm was recorded in between, the bar runs to the last of those so it always contains its children (an eou_detection still running is closed with a superseded event). A turn-detection mode change to or from manual drops the pending wait.

eou_detection keeps its name and now nests under eou_wait; the span names and the lk.eou.* attribute keys stay on the same vocabulary, so nothing on the cloud side needs to change.

on_user_turn_completed span around the user hook that gates the reply, nested under the user_turn it completes: recognition hands the open user_turn span to the activity when it schedules the hook, and the activity ends it after the hook (or when the turn is skipped), so the turn's duration covers the wait for the hook. Exceptions are recorded redaction-aware, honouring record={"redaction": True} on the session as well as the job's flag; StopResponse is an event.

lk.speech.queue_wait on agent_turn: seconds between scheduling the speech and its first generation authorization.

user_turn
├─ user_speaking
└─ eou_wait                 ← new, 2.5s, rearmed ×1
   └─ eou_detection         ← now nested under the wait
on_user_turn_completed      ← new
agent_turn  (lk.speech.queue_wait=0.01)
├─ llm_node
└─ tts_node

Tests

tests/test_eou_wait_span.py drives _run_eou_detection with real spans: commit (parent, back-dated start, attributes), re-arm keeps one span, resumed speech ends at speech start and leaves user_turn open, teardown drops, not_committed keeps waiting, eou_detection nests under the wait and stays inside it when the user resumes mid-inference, a mode change drops the wait; plus a full fake session asserting eou_wait, the hook span, the queue-wait attribute, and a session-only redaction keeping a hook exception's message out of the trace.

Stacked on #7128.

🤖 Generated with Claude Code