- harden isolation ModelPatcher proxy/registry behavior for DynamicVRAM-backed patchers
- improve serializer/adapter boundaries (device/dtype/model refs) to reduce pre-inference lockups
- add structured ISO registry/modelsampling telemetry and explicit RPC timeout surfacing
- preserve isolation-first lifecycle handling and boundary cleanup sequencing
- validate isolated workflows: most targeted runs now complete under
--use-sage-attention --use-process-isolation --disable-cuda-malloc
Known issue (reproducible):
- isolation_99_full_iso_stack still times out at SamplerCustom_ISO path
- failure is explicit RPC timeout:
ModelPatcherProxy.process_latent_in(instance_id=model_0, timeout_ms=120000)
- this indicates the remaining stall is on process_latent_in RPC path, not generic startup/manager fetch
