Controlled unit swap on mothership: single PID hosts all 4 workers,
shared repo + 30s reconcile pass confirmed working, no crashes. Live
delta beat the floor estimate. Batch group complete.
4 live batch workers = 509MB; consolidated startup floor = 118.5MB
(imports 102.5 + repo pool 15.7 once vs 4x + bus 0.3). Side-effect-free
measurement (no mutation tick). Floor is a lower bound; live adds modest
per-worker working set. Exact live number pending controlled unit swap.
forwarder/listener are role-split swarm singletons, not co-resident with
the herd — drop from grouping. Real master-resident batch group shares
the repo singleton (one DB pool when consolidated). Stage 1 = batch:
reconcile/enrich/orchestrate/mutate. webhook/canary deferred standalone.
HOW to consolidate: supervision-loop primitive (not TaskGroup, whose
all-or-nothing cancel breaks isolation); group by failure domain +
resource profile keeping per-group cgroup limits; every worker remains
config-extractable. Recommend process-groups now (~18->~9 units),
evaluate prefork+gc.freeze CoW on 3.14 as the higher-ceiling follow-on.
Measurement after C2: all 25 CLI modules transitively pull the SQLModel
ORM, and the table chain is a hub (one table loads the whole registry).
Most workers genuinely use the DB, so lazy imports only help the 2-3
DB-less workers. Consolidation (pay the 86MB floor once for the herd)
is the reliable ~600MB win — promoted from fallback to primary.
Fleet resident set ~2.57GB across 18 workers; ~1.5GB is the 86MB import
floor paid 18x. Pinned root cause: topology/__init__ eager re-export of
generate drags the full SQLModel ORM (26 tables, ~38MB) into every worker.
