Claude settlement observations

VUH-1527: this bounded change removes false completion claims from the existing Claude worker channel. It does not establish exact message-to-turn completion.

A channel transcript receipt can be a queued-command attachment. The forwarded Stop and StopFailure payloads have a session ID and bounded native text/error, but no native turn or message identity. A later observed Stop, even after an idle dispatch, therefore cannot identify which message finished. Hook sequence, wall-clock time, equal text and counting Stops are not substitutes for that missing native correlation.

Scope decision, 2026-10-04: every accepted Claude channel receipt reports state: "queued", including an idle pane. Neither that state nor the retained deliveryStage: "consumed" means a model saw the message or began its turn. The initial adapter start result still describes the native TUI starting; it is separate from a brief starting a model turn. The mailbox and wire APIs stay the same.

The existing bounded hook log now writes a dispatch boundary before handing a brief or follow-up to the mailbox. A save failure prevents the handoff. The boundary is retained through newly attached controls and service restarts; idle, Stop, SessionStart and owner prompts cannot clear it. The log records its own observation order only. It adds no queue, transcript mirror, delivery retry, transport, tool authority or account configuration.

Claude settlement returns an explicit unconfirmed observation, with authentic eligible Stop/StopFailure data when available. It never synthesizes a successful turn from idle or uses a pre-dispatch Stop as a new message's completion. The watcher still wakes the original conversation once and quotes uncorrelated Stop text as data. That wake is not acceptance of queued work; the lead must inspect the worker's thread and evidence. The ordinary blocked, released and missing-seat outcomes remain distinct, and caller cancellation aborts a held status wait.

Missing, unreadable and legacy log data cannot establish a known-empty dispatch history. In particular, the current native hook payload does not distinguish a freshly born manual session from a resumed session with older queued work. A first-seen manual watch therefore also reports an unconfirmed observation while preserving its authentic Stop/StopFailure. There is no fixture-only or inferred known-empty state. A sequence watermark rejects older observations; it is not native turn correlation.

The focused causal checks cover stale and absent Stops, a busy turn stopping after a queued follow-up, a fast Stop during initial delivery, reattachment and restart, unknown state, persistence failure before either handoff, overlapping dispatches, owner prompts, native failures, cancellation, lost status and changed session/pane/terminal bindings. They use the production adapter/log/watch with fake native dependencies, not model calls. Live remote delivery, native turn correlation and owner PC acceptance remain unverified; the issue stays open.