Fresh-process checkpoints
Before adapter import, after import, connected, post-warm-up, post-run, before cleanup GC, and after close plus GC.
Speed is not the whole cost
Resources and cleanliness
The resource lane records memory, GC, event-loop delay, package footprint, and teardown separately from uninstrumented speed timing.
Current profile resource state
Peak and post-cleanup memory, GC, installed footprint, and teardown are reported per accepted environment and adapter.
Loading compact benchmark data…
| Profile | Peak RSS | Post-cleanup RSS | Installed | Cleanup |
|---|---|---|---|---|
| Raw | Pending | Pending | Pending | Pending |
| Fast | Pending | Pending | Pending | Pending |
| Guarded | Pending | Pending | Pending | Pending |
| Assured | Pending mutual-TLS record | Pending | Pending | Pending |
The compact dashboard is unavailable. Missing evidence stays pending; it is never converted into a zero.
Memory and garbage collection
Process memory
RSS, heap total/used, external, ArrayBuffer memory, maximum RSS, CPU, and event-loop utilization.
GC observation
Workers run with --expose-gc. Three forced collections after warm-up occur outside timing; the resource pass separately observes GC work.
Package footprint
A full recorded run packs and installs each subject in an isolated temporary directory. It reports compressed and unpacked tarball bytes, package file count, complete production-install bytes/files, and direct, installed, and unique dependency counts.
Comparator packages stay isolated.
They are benchmark subjects, never node-ipc dependencies. Package-size comparisons must use the same package manager and clean-install procedure.
Cleanliness gate
- Every worker and reflector exits naturally; a forced termination fails the sample.
- Sent, reflected, and received messages and bytes match exactly.
- No socket remains open and the endpoint is immediately reusable.
- No unexpected active resources remain after close and GC.
- No temporary socket, file, or directory is left in the private sample root.
- Oracle saturation, timeout, protocol error, or cleanup mismatch makes the record nonpublishable.
Cleanup is a gate, not an average. One failed sample prevents the run from becoming official evidence.