Use live host usage and running estimates for collection admission

docs/native-hosts.md

@@ -180,19 +180,21 @@ heartbeat ingestion, at most once per minute per host/fleet phase; new peaks
 raise the cache immediately. Launch admission reads only the cached model.
 
 Scheduling still leaves **5% host CPU and 500 MB RAM (500,000,000 bytes)** for the OS and bursts.
-These allowances are scheduling reservations, not Docker limits. Queued and
-unresolved creations consume the startup allowance; established instances keep
-the running allowance even while idle or stopped. Current container measurements
-are credited against their reservations because host usage already includes them;
-only the unused allowance is added. Missing or mismatched-generation measurements
-cannot release an unresolved launch's reservation. Observed host load also counts,
-so other processes and previously unseen load increases reduce available capacity.
+These allowances are scheduling estimates, not Docker limits. Collection admission
+budgets new and never-submitted queued servers at their running allowance, rather
+than assuming the entire collection starts simultaneously. Current host CPU and
+available memory account for existing containers and other processes. Observed
+containers are not topped up to historical startup or running peaks, even while
+awaiting health verification. Established instances are covered by host usage
+when individual container measurements are unavailable.
 
-Admission counts all queued reservations. Dispatch evaluates one launch against
-existing instances and submitted or uncertain work; other never-submitted queued
-launches wait their turn. They retain their port reservations. This lets the queue
-make progress even if a changed allowance means the whole backlog no longer fits
-at once. Capacity errors report projected usage, usable capacity and reservations.
+Dispatch checks the next launch's startup allowance against live host usage.
+Other never-submitted queued launches wait their turn and retain their port
+reservations. Submitted or uncertain launches missing a matching-generation
+measurement still consume a full startup allowance until telemetry accounts for
+them. This lets collections queue based on running demand and start as capacity
+becomes available, without stacking every queued server's startup peak.
+Capacity errors report projected usage, usable capacity and reservations.
 
 When startup capacity is available, provisioning checks each eligible host’s
 current CPU and available memory against reservations and headroom.