[lweberru]

A live hosted Plan finished about 0.8 seconds after its first observation, before CF processed the matching warm Runner stop. The correct instance was selected, but this was a missed interruption window and is not passing recovery evidence.

Observe the bounded first native version record through a best-effort filesystem watcher instead of waiting for the normal 500 ms output publication interval. Keep that publication path as fallback; watcher failure never fails or delays the Plan. Only fixed identifier/timestamp observations leave the worker, never native contents.

Validate the immutable Dev organization/space/template boundary when arming, then recheck the exact app GUID and space immediately before mutation. Warm stops now need one app read followed by the stop request; cold tasks retain their unique-active-task check. A boundary that was never verified cannot authorize any stop. Existing confirmation, recency, completion, private durable intent and no-retry guards remain. Receipts additionally record intent/response timestamps. Complete English/German documentation is updated.

Validation: full Configurator check passed (1,056 tests passed, 71 gated skips); focused worker/observation suites passed (117 passed, 4 gated skips); 20 Python safety tests passed; diff checks passed. New tests cover partial version records, suppressed late/duplicate notifications, watcher unavailability, unverified boundaries and foreign app spaces. No new hosted interruption has been attempted with this source. The prior live missed window is retained as failed/inconclusive coverage, not retried to claim a pass. No State, lock, Plan-input, approval or Assistant-fact changes. Related: #90, #95.