Summary

• enforce --trace-idle-gap-cap-seconds only as an observed runtime idle watchdog per complete trajectory tree
• preserve source dataset timestamps and end-to-start replay delays without load-time rewriting
• restore every surviving warmup stream to one flattened per-trajectory dataset clock so non-burst profiling preserves cross-stream order and spacing
• include the root, subagents, and flattened-agent streams under one runtime-root watchdog
• preserve deferred child and join state across accelerated warmup handoff
• release profiling joins only after both the recorded replay deadline and all required child completions
• keep both per-tree and whole-system watchdogs active when a cap-advanced timer is retained by a replay dependency

Runtime semantics

The trace watchdog starts for initial future profiling work and restarts whenever the final in-flight request across a root and all descendants completes. If no request from that tree reaches the wire before the configured cap, AIPerf uniformly advances only that tree's pending scheduler timers. If the first advanced request is still dependency-blocked, the watchdog continues within the same expired idle interval until an eligible request reaches the wire. Other roots are untouched, relative timer spacing is preserved, and dependency gates remain authoritative.

Accelerated warmup intentionally removes runtime delays. At handoff, timestamped roots, subagents, sidecars, and gated joins are restored to their next position on one shared tree clock. Profiling subtracts one global minimum, so the earliest request starts immediately while all remaining streams retain their original flattened order and relative start spacing. Timestamp-less datasets keep the existing per-stream delay fallback.

The global idle observer is registered when each phase becomes active, so warmup teardown cannot clear profiling's observer. The dataset loader does not rewrite timestamps for the trace cap: recorded timestamps, api_time, and derived end-to-start delays remain unchanged. Actual server completion behavior determines when a runtime tree becomes idle.

Validation

• warmup handoff tested with 1, 10, and 100 additional requests per lane
• prior pending-child turn-0 thundering herd reproduced before the fix and eliminated after it
• exact root/subagent ordering and SPAWN_JOIN completion gates verified through the HTTP mock server
• dependency-blocked watchdog candidate reproduced; the next eligible timer issued within the same cap interval
• issue agentic_replay drops post-child delay on parent resume ai-dynamo/aiperf#1231 root-child-root timing shape verified with and without the separate global idle guard
• independent sibling-stream drift verified against the runtime whole-tree idle cap
• active-phase observer ownership and initial profiling idle verified with real scheduler timers
• broad unit, component, adversarial, and real HTTP mock-server suites passed independently by multiple reviewers
• all pre-commit hooks passed