Bug type

Behavior bug (incorrect QA result without a runtime crash)

Beta release blocker

No

Summary

qa/scenarios/channels/native-command-session-target.yaml treats the native /stop acknowledgement as if the aborted turn has already released its session lane. Under QA Smoke CI contention, recovery can wait behind abort cleanup longer than the scenario's fixed 15000ms outbound timeout even though /stop targeted and aborted the correct conversation.

Steps to reproduce

1. Run the isolated Crabline Telegram scenario with mock-openai/gpt-5.5 in the QA Smoke CI profile.
2. Let the delayed turn become active and inject native /stop.
3. Observe Agent was aborted, followed immediately by the recovery inbound turn.
4. The recovery marker arrives only after the fixed 15000ms wait expires because the prior turn still owns the session lane while unwinding.

Exact evidence:

• PR fix(qa): prevent qa smoke ci timeouts under gateway concurrency #99368 run 28639086776, job 84931413935, artifact 8058095433: /stop at 04:54:28.301, acknowledgement at 04:54:29.115, recovery at 04:54:29.301, then an 18578ms session-lane wait and final recovery timeout.
• PR test(qa): use full-turn budget for native stop recovery #99656 run 28684909956, job 85075721090, artifact 8075313865: acknowledgement at 21:57:46.010, recovery at 21:57:47.056, then a 16586ms session-lane wait; recovery output arrived at 21:58:06.651 after the 15000ms wait expired.

Expected behavior

The scenario should keep the native command and abort acknowledgement assertions, then allow the recovery message one full QA turn budget, including time queued behind abort cleanup.

Actual behavior

sessions.list.hasActiveRun can clear before the command lane is released, so it is not a valid idle signal. The QA flow has no gateway RPC exposing canonical session-lane idle state.

Fix and validation

Use the existing liveTurnTimeoutMs(env, 30000) full-turn timeout policy for the recovery outbound wait. Mock providers preserve the 30000ms fallback; live providers apply their model-specific floors. This adds no sleep and does not weaken the abort or recovery marker assertions.

Validation for the narrowed patch:

• Scenario catalog: 32/32 tests passed.
• Exact Crabline Telegram contention: four simultaneous workers passed 4/4; every recorder was parse-clean and contained both abort and recovery sends.
• Fresh artifact-aware autoreview: clean, no accepted/actionable findings.
• Blacksmith Testbox changed gate: tbx_01kwn0fe5vm4gevwq0c3vks9r9, exit 0.

JSONL partial-tail parsing is unrelated and unchanged.

Impact and severity

Intermittent false-negative QA Smoke CI results for native /stop recovery proof. No production native-command regression is indicated.