[masarray]

Architecture specification only — NO firmware changes, no operator build, and NO main merge. Based exactly on checkpoint merge e82040716c549af887579b369965ba96a890a4db (merged PR #288). Root cause: stop_sample_clock() revokes smpSynch and deletes GPTimer when injection stops, while PTP SOURCE continues; baseline already has a dry preflight but it can still rearm/reset. Split always-on ClockAuthority / SampleGrid / SV-TX gates, with configurable INTERNAL_GM_ONLY, EXTERNAL_ONLY, AUTO_PREFER_EXTERNAL, MONITOR_ONLY and safe handover/HOLDOVER. First implementation PR must fix STOP lifecycle without touching working SV/PTP frame codecs, and prove 100 STOP/START cycles with no loss of qualified phase. Coordinate with existing M3 issue #162, M2 issue #154 and Draft PR #166; DO NOT blindly merge #166 into PCAP-proven checkpoint. Full acceptance in docs/architecture/ARSTACK_PTP_SV_ALWAYS_ON_CLOCK_M3_20261023.md.

[masarray]

M3.H integration audit (2026-10-24): this is a DESIGN-only divergent branch, not an ancestor of final tested firmware a922e9bd... (comparison: 1 unique commit on this branch). Integration PR #299 preserves the verified runtime in a separate two-parent merge. Do not merge this standalone design PR into main blindly or mark its optional AUTO/external-GM architecture fully implemented. Keep as a separate architecture review until its unique document is reconciled.

[masarray]

Triage closure (24 Oct 2026) — superseded, NOT merged. This is an architecture-only draft with no firmware change. Its successor M3.A→M3.H implementation is now on main through #299; broader external GM/AUTO policy remains tracked in issue #162. The single divergent design commit is not claimed merged and its branch is retained for historical reference.

Verified main merge #299 passed 14/14 workflows; release and firmware assets unchanged. Closing this old competing PR is a review-queue cleanup, not a deletion or certification.