a restart that no longer costs the day

Every restart used to cost most of a working day during which the system was running and not allowed to decide, because its rolling state had to be rebuilt from live data before the first decision was trusted. The fix seeds that state from the platform's own history at boot, from the same source the live stream comes from.

That only helps if the seeded state is identical to the state live accumulation would have produced. Otherwise the system comes back holding a slightly different opinion than it held when it stopped, and the seam is invisible: no error, no gap in the record, just a decision that would have gone the other way the day before.

So the obligation was set as identity rather than closeness, and tested two ways. A unit-level comparison of seeded state against live accumulation over a synthetic series came out identical in every field. A platform-level comparison over real history then matched on every value that is read, with one field differing, that field never being read by anything.

Two disciplines came with it. Fresh subscriptions were established only after the backfill, so the boot window's backlog could not arrive looking like current input. And any fetch failure falls back to the slow, honest path of waiting, behind a minimum-state guard that refuses to evaluate empty state. What it proves is that restarts are cheap here because equality was demonstrated, not because nobody checked.