Hosts clusterer/campaign-clusterer/attribution/reuse-correlate in one
process. The two O(n^2) connected-components kernels (cluster_observations,
cluster_identities) offload to ONE shared forkserver pool via decnet.offload
.run_kernel, so they run in parallel instead of serialising under the GIL.
- offload.run_kernel: pool when installed + offload_if holds, else inline.
Standalone workers and all tests run inline => behaviour unchanged
(424 clustering/correlation tests green).
- offload_if gates on input size (>=256) to skip pickle cost on small passes.
- forkserver (not fork): supervisor is multithreaded via bus clients.
- attribution/reuse co-located but not offloaded yet (lighter; same run_kernel
path extends to them if profiling shows contention).
- systemd unit Conflicts= the 4 units it replaces; no docker/raw-socket priv.
Hosts reconcile/enrich/orchestrate/mutate in one process via the
supervision primitive: one import floor, one shared repo/DB pool instead
of 4. Static group registry (membership is architectural, not a knob);
factories lazy-import only the hosted workers. systemd unit Conflicts=
the individual units it replaces and documents the union-of-privileges
cost. Worker code unchanged — any member is extractable by editing _build_specs.
