CLI transport overhead

Tracked by pm-yse5dt, with RSS admission reliability owned by pm-pz49xc.

Each result starts from a fresh isolated workspace containing exactly one item. The command runs in a fresh Node v26.5.0 process on linux/x64; setup and fixture generation are outside the timed interval. The default gate measures ten post-warmup fresh processes per command and uses the best observed latency, which keeps the immutable ratchets meaningful under transient host contention; the report retains p50 and p95 evidence. RSS admission uses the measured median so one page-level outlier cannot false-fail the gate; the maximum remains in the report as diagnostic evidence. Every post-warmup RSS sample must be a finite measurement; an unavailable sample makes the admission median unavailable and fails closed when a budget exists. Admission adds a fixed 512 KiB noise margin without changing the committed budget; a majority persistent increase beyond that bounded margin still fails.

Current attribution

• Node process and ESM loader floor is measured independently in the SDK entrypoint table.
• Static CLI bootstrap loads the shared error, output, telemetry, extension discovery, Commander, and SDK-client kernels before command registration.
• Command registration is already family-selective; only the family owning the requested command is loaded, while bare/root help intentionally registers the complete discoverable surface.
• Settings/schema reads and extension discovery happen after the module floor and remain observable through --profile; the one-item fixture keeps their data-dependent work negligible.
• Focused SDK entrypoints remove 40-98% of aggregate import overhead for governance, graph, merge, query, authoring, contracts, and testing consumers. The compatibility aggregate and core client remain intentionally broad.

The committed budget is a ratchet. Existing absolute scale budgets are not relaxed, and scale reports additionally gate the CLI-minus-SDK delta for every operation whenever both transports are measured. The generated scale workspace has no installed extensions and its paired SDK client uses extension-free mode, so the comparison excludes project package execution. It also leaves author attribution to the CLI resolver instead of injecting PM_AUTHOR.