[santoshkumarradha]

Stopping a conversation now ends its delegated run, background jobs and automatic helpers. --max-hours ends owned work at the deadline and explains how to continue in a new conversation. A pending run start cannot revive after Stop.

Page Stop prevents a pending isolated landing from continuing into checking or narration. History now shows the live task once instead of re-adding its stale saved record, so opening the task reaches its working room and Stop control. A home merge already in progress retains its actual result; usage from requests already sent remains billable. Window detach behavior is unchanged.

Actual terminal evidence

All scenarios used isolated fictional counter projects and drove the real terminal controls.

Scenario	Observed result
Page Stop	History showed one task; entered its live room and confirmed Stop. Child output stopped, no checker/acceptance/resume. Request starts stayed at 2 and heartbeat at 15 bytes for 10 seconds.
16-second conversation limit	Owned child stopped before the terminal limit message. Disclosed stopped state and exact stored time-limit reason verified. Request starts stayed at 2 and heartbeat at 30 bytes for 10 seconds.
No-limit control	Child kept writing beyond 18 seconds, then stopped through the real task page. Request starts stayed at 2 and heartbeat at 105 bytes for 10 seconds after Stop.
Live multi-turn workflow	Normal chat, delegated 120-second bash counter, actual task-page Stop. Partial output retained and unchanged for 10 seconds; no checking or acceptance. All 16 recorded calls used OpenRouter deepseek/deepseek-v4.1-flash.

Scripted scenarios passed in 95.02 seconds; the live workflow passed in 28.58 seconds. Each scripted run also verified unchanged nonzero prompt/output/cache/cost totals after Stop.

Regression checks

Reproduced before fixing: conversation Stop left an owned run active, auxiliary retries continued, late landing overwrote stopped state, and history duplicated the live task. Focused cancellation, admission, landing, helper and history regressions pass, including race checks.

make pr-ready completed successfully. All selected session tests passed across eight shards. One unrelated credit-selection assertion (TestLowCreditReadSwapsOnlyAnUntouchedImplicitConversation) fails identically on the unchanged dev base; the gate independently reproduced and classified that baseline failure. No test assertion or skip was weakened to hide it.

Closes #1350.
Closes #1352.