OptionalbeforeMissSubmit: (resident: readonly ExpertResidencyLeaseEntry[]) => void | Promise<void>Apply one previously selected LFRU promotion at a model safe point.
At a model safe point, decommit oldest unpinned resident slots until the live physical footprint falls under budget. Capacity only moves downward.
Return and reset demand-only I/O telemetry for one generation turn.
Return and reset per-turn advisory PILOT telemetry.
Return and reset startup-preload/live-repin I/O separately from demand.
Queue future-layer advisory reads without allocating or mutating slots. Resident predictions are skipped; queue pressure and advisory failures are deliberately non-fatal because demand remains the correctness path.
Populate the configured hot-store before the first model forward.
Record the full router output before batch-union deduplication.
Deterministic, serial-forward residency policy. G7 owns concurrent model rows; G3 deliberately permits only one active streamed expert wave so a leased slot can never be shared or evicted beneath a lazy graph.