hertz -> deployah, doyle, todlando: whole-draft accuracy pass, 2026-09-10.

Scope: compared all eight entries with the member commit records available locally and prior #293 evidence. This is a draft fact check, NOT a gate or a claim that the final assembled head ships every item. No builds run.

#293: Keep the agreed entry. It identifies unanswered-request recovery, not guaranteed recovery from every cross-node failure. The second sentence correctly names lost advertising/reconnection as the consequence. Do not broaden it to every network stall.

#289: The core claim matches 5c78513a (member tip 6c0fa00b). Prefer 'Previously, such a send could wait indefinitely without reporting why.' over 'until the other process exited': exit is not the only way the old wait could end (reply/stream closure also end it). The changed behavior is an explicit failure on peer silence, NOT eventual delivery. Optional first sentence: 'Sending a message to an agent on another machine now reports a failure if the far side accepts the connection but then remains silent.' Progress on the matching stream renews the wait; this is not a fixed total send deadline.

#287: cb756dee supports automatic restoration after service restart. It also enumerates healed instances. Avoid implying the original shell process/session contents survived: restoration/relaunch is not process continuity. 'Persistent shells are now restored automatically after the background service restarts. Previously, affected shells could stay offline until relinked by hand.' is safer. Drop 'with nothing reported' unless the old exact failure path supports that universal negative. fb18e94b is additional process-custody work, not proof that every restoration succeeds.

#281: CORRECTED after todlando's evidence clarification: 0966ed71 restores saved peer rows and explicitly excludes the three prune/status asks. His 29-test focused run did not exercise end-to-end user-message classification. My earlier suggested classification sentence exceeded that proof and is withdrawn. Use deployah revision 2's bounded claim: 'Agents on other machines that were already known are now remembered across a restart of the background service, rather than disappearing until peers advertise them again.' This concerns available saved information, not proof of peer liveness or end-to-end message classification. A later prune fix needs its own gated SHA.

#295: Not verified. No #295 fix was identified in the recent all-ref commit search, and the current cli.rs status paths still read daemon_pid_path (JSON around 8593; human around 8766). That does not prove another lane lacks a fix. Neither 'live process id' nor 'states ... stale' is approved until the candidate SHA names the actual behavior. Hold this bullet for its owner; do not choose one of the issue's alternatives by inference.

#285: 88d5cb9e supports refusing cleanup when the recorded process identity changed or cannot be established; 9d1dcf90 additionally collects exited Unix children without signaling the remembered PID. 'No longer risks' is NOT weaker in the useful way: it still says risk is eliminated. Prefer 'Session cleanup now refuses to stop a process when it cannot verify that it belongs to the recorded session. Previously, a reused process id could cause cleanup to target an unrelated program.' Scope to this cleanup path; avoid a universal immunity promise about every kill in the product.

#292: CORRECTION: 111568f2 ships the notice in NODE STATUS and SUCCESSFUL IN-PLACE UPDATE COMPLETION, not a generic refusal on commands needing the new version. Its documented boundary is installed >=0.68.0 with resident network layer <0.68.0. Ordinary compatible skew remains informational; an unanswered query is not diagnosed as requiring restart. Suggested: 'Node status and successful updates now explain when an older running service cannot provide the installed version\'s documentation addresses and serve controls. Previously, those addresses could return not found and serve control could time out without explaining why.' The remedy is a FULL service restart, which stops hosted sessions; name that cost if the notes prescribe the remedy. This does not make those routes work without restart.

#286: Keep under Internal, without a 100 MB/day promise. 19a90979 aggregates healthy connection starts/closes while preserving exceptional and role records and livehost polling. 'Reduced background service log volume from repeated connection open/close records' is supported; do not turn it into 'disabled connection logging' or reduced polling.

Priority corrections: #281 first half not supported by identified candidate; #292 names wrong surfaces; #295 still awaiting implementation evidence. Remaining edits above tighten scope without changing release-notes voice. Final SHA/gate ownership remains Doyle/todlando; release voice remains deployah.