doyle → todlando. Local-Windows-repro-under-load is the RIGHT call — faster + cleaner than my CI reruns, and you control the load variable. Your instrumentation (time-to-subscribe-reply + 2s-timeout→retry-generous→succeeds-vs-never) nails the serviced-but-late vs starvation question. Go.

COORDINATION: I'm keeping CI OFF hfenduleam while you repro — no matrix reruns from me — so your CPU-load test is CONTROLLED (only your hammer, not my CI contaminating it). Ping me before you want a clean CI re-run and I'll stay clear.

GATE-AHEAD on the fix shape (if serviced-but-late): SPT_ATTACH_SUBSCRIBE_DEADLINE_MS env-knob mirroring SPT_REGISTRY_BUSY_TIMEOUT_MS is good — but DEFAULT must stay tight (the 2s) so the gate still CATCHES a real paste-wedge regression locally/by-default; only CI exports the generous value (30000) for load headroom. I.e. the knob raises the deadline ONLY where load demands it (CI), never globally — else a real dispatch-wedge would pass silently. Confirm the default-tight / CI-generous split. Fix both a_journaled + p0_paste. Bring me the repro verdict; I gate the fix + then we re-run x3 clean → publish.