Deferring unrelated CLI registration did not establish a useful packaged startup improvement. The predeclared replication remains inconclusive; the experiment stops here. This does not prove the underlying effect is zero.
| Combined 40 pairs | Paired median change | 95% bootstrap interval |
|---|---|---|
| Command-ready descriptor | −4.04% | −11.17% to +3.99% |
| Actual app usability | −7.35% | −13.43% to +0.21% |
Negative values favor the candidate. Twenty original pairs plus one predeclared independent cohort of twenty, alternating AB/BA. Twenty relocated same-bytes A/A controls were noisy and are shown below. No arbitrary minimum gain was applied.
Boat bx_6a8qz8n5, Ubuntu 24.04.4 x64 VM, 4 vCPU / 8 GB. Production Linux unpacked directory builds, Electron 44.10.1. Fixed synthetic 12-project / 48-stopped-thread returning home, isolated per launch. Compile cache and Code Cache reset identically; warm OS pages. No profiling during timing. Node 26.12.1 / pnpm 12.14.0. Native macOS, Windows and WSL timing is unmeasured.
Base 6417151c093e54dbba0e6c824c69d03983793b9a; candidate 5378317f942ddc36a2f4ce4cb7e06cf841a90480. Identical packaging flags: --platform linux --target dir --arch x64. Matched-pair median bootstrap: seed 20261108, 100,000 resamples, percentile 95% interval. Whole pairs were resampled. The interval assumes exchangeable pairs and cannot remove all VM temporal effects.
Both cohorts span two display epochs. Failed incomplete pairs were retained and repeated on fresh endpoints; an original A/A screenshot-only exclusion was analyzed in sensitivity checks with no changed conclusion. Identity inventory was captured late and cannot prove prelaunch forensic parity. Ratios of marginal medians are not claimed as causal gains.
87 focused tests and changed-file checks passed. Packaged fd0/fd3 startup, SIGTERM shutdown, same-home restart on different ports, and eight CLI fallback checks passed. Integrated abrupt launcher loss, same-port restart and native platform proof remain incomplete. No PR, push, bot watch or merge was performed. Candidate stays committed for reference.
The shared root and runtime policy have one owner. Followups if revisited: keep the exact bootstrap argv allowlist aligned with desktop launchers; prove abrupt launcher loss and closed streams at the packaged boundary; keep packaging compile-cache coverage separate.
GPT 6.1 Sol through Codex in T3 Code. PostPlan uses fallback black colors because the active client theme could not be read.