[chris-beckman]

Stopping an active response can finish the Stop barrier before the provider unwinds. The host now classifies cancellation against the captured session generation before publishing turn failures, so an intentional Stop does not produce a failure banner. Unexpected cancellation, deadlines, and joined cleanup/provider failures remain reportable. Late submission completion also recognizes a persisted interruption only when it belongs to the same nonempty claim token, preserving stale-claim rejection and real receipt errors.

Depends on #1735 and #1737 → #1736. This work lands after the other open work. Keep draft at the user's request. No database or wire shape changes.

Verification covers active and completed Stop, wrapped/joined cancellation, unrelated-session Stop, unexpected cancellation, real joined failures, and prompt-lock/drain behavior. Earlier memory/SQL receipt and prompt API fixtures passed 53 results. The new four-package admission/lifecycle/submission/execution run passed 36 results with no skips; final admission-only recheck passed 14 results on tree 627e262e80141636151132780e7536316599669d. Budgets passed before a three-line simplification of the cancellation predicate (no added surface or growth). Actual packaged M1 recheck on composed source 3f6cad5f0e1df178f4f1ecef773e39eaba988b69 passed: a running dispatched response settled interrupted and the UI showed Stopped without a failure banner before recovery. The subsequent UI request answered ready in 4.473 seconds (11,991 prompt tokens; 11,402 reused), with the same host/generation/decision processes. Both Bialy decisions were nonabstaining. This is scoped lifecycle qualification, not final-model quality or signed-release qualification. All owned remote app processes were stopped after evidence export.

[chris-beckman]

Delivered on main through #1987 (ad462d8629) from this exact source head e51c395ec692. The source branch is retained.