[Performance] Assess proposed CLI cold-start target

[nth-bailey]

Context

#62 added a reproducible Linux executable-cache startup benchmark and verified terminal diagnostics. On the 2026-09-22 WSL2/ext4 host, a release CLI measured a 7.61 ms median for --help and 7.71 ms for argument validation after requesting executable-page eviction; all 100 samples in each case had major page faults. Warm medians were 1.52 and 1.64 ms. This misses #50's proposed <5 ms cold-start target on that host. These measurements do not represent a first process after reboot, other shared-library cache states, or other operating systems.

Investigate

• Run benchmarks/cli/startup.py on representative native Linux hardware, and use an equivalent controlled methodology for macOS and Windows if a cross-platform target is desired. Record storage type, binary size, kernel/OS, major faults where available, and medians/p95s.
• Profile where the cold time goes (executable and shared-library page faults, loader work, or CLI initialization). Compare against a small executable baseline on the same host.
• If PolyXML-specific work or binary size dominates, optimize it without regressing warm startup or generated output; remeasure both cache states.
• Decide whether <5 ms is a realistic supported-host target or a host-specific goal. Document the decision and results.

Related: #50, #62.

[nth-bailey]

Closing after reassessing the premise. The <5 ms cold figure was an aspirational line in #50, not a supported-host requirement or user-reported bug. #62 measured a 7.61 ms median for executable-evicted help on one WSL2 host and found no CLI startup defect. We will keep the benchmark data in PR #64 without pursuing an arbitrary threshold.