Baseline: ade473018d32003658b7d38693e083ee2e314bd8.
Candidate: 37aa1e1fedfe794155571d498f30cc894e5b91d1.
Ubuntu 24.04 x64 Boat VM, 4 vCPU / 8 GB. Unsigned Linux directory packages, Electron 44.10.1, embedded Node 24.22.1, V8 15.2.124.45-electron.0. Build Node 26.12.1 / pnpm 12.14.0. Original source and candidate are built independently with their own commit metadata and the same flags:
pnpm install --frozen-lockfile
T3CODE_COMMIT_HASH="$COMMIT" node scripts/build-desktop-artifact.ts --platform linux --target dir --arch x64 --output-dir "$OUTPUT" --verboseBoth packages run on the same VM with warm OS pages and matched cold runtime code caches. Each launch restores the same frozen synthetic data (12 projects / 48 stopped threads), removes only runtime code cache from the copied fixture, and gets its own HOME, XDG, T3 and Chromium state. No real account data or provider turns. Original fixtures and packages are unchanged.
Two warmup pairs are discarded before 20 alternating AB/BA pairs. All 40 measured launches are valid; no failed or slow samples were removed. Server ports start at 12000 and CDP ports at 14000, incrementing by pair and side. Timing runs have no inspector, tracing, NODE_DEBUG_NATIVE, or NODE_DISABLE_COMPILE_CACHE. Fixed launch flags: --no-sandbox --password-store=basic --remote-debugging-port=<per-run> --user-data-dir=<isolated>.
xvfb-run -a -s '-screen 0 1280x900x24 -noreset' python3 "$PAIRS" \
--node "$NODE26" --harness "$HARNESS" --source "$CONTROLLER_SOURCE" \
--base "$BASE_BINARY" --head "$HEAD_BINARY" --output "$NEW_OUTPUT" \
--fixture returning --returning-home "$FROZEN_COLD_FIXTURE" --rounds 20Milestones remain distinct: PID-owned X11-visible native window; successful environment descriptor behind the command-readiness gate; visible editable composer that accepts synthetic probe text with the synthetic sidebar present. The timing controller then attempts .fill('') but does not assert clearing. Clearing is proven separately by the integrated client check, not by this timestamp. Screenshots occur after timestamps. The controller polls at 20 ms, equally for both builds. Cleanup signals only its own captured process groups.
The JSON contains raw paired timing values in milliseconds, launch timestamps, alternating order, and independently recomputable statistics. p95 is nearest rank. Bootstrap intervals use 20,000 paired median resamples with deterministic seed 6417151. Negative candidate-minus-baseline deltas mean faster. The command-ready paired median is -141.38 ms, 95% interval [-246.75, -83.13]. Window and usable-composer intervals cross zero; no improvement is claimed for either. Three separately restored, normally warmed control pairs are descriptive only, with no gain claim.
Baseline ASAR SHA256: e33ab1057f536ed6b501d357edeb3afc20e813f0b03bad277e6f653f79830aec.
Candidate ASAR SHA256: 62d350cc951a7e2c0a9d5a5b3a1ec4ab395d7c76700c7124d5f43f35d05ffc8c.
Both Electron executables: 7722b3b95b0a31e24906f8046c1bf29cb03c1b99f67058b0b9a4c0e8681c0d87.
Candidate AppImage: ed03b093e6f4af5420cebdff9f0f200ff5b9dd2a337bae12b4b4d00694bd41f8.
Frozen synthetic fixture archive: b04725503fd8e5fa8a4744342185b35bf67907ece8ab65cce8d66b8362223d34.
Separate correctness diagnostics prove actual read-only FUSE AppImage relocation, accepted server/runtime bytecode, crash reclamation preserving a concurrent live cache, normal-exit cleanup, and unchanged package/fixture hashes. Normal quit restores accepted main-process caches in the warm controls; SIGTERM is not treated as proof of that. No AppImage timing, fresh-onboarding performance, native macOS/Windows/arm64 timing or purity, or combined-ticket gain is claimed. Linux arm64 generation remains skipped.
Earlier experiments against 6417151c09, including CLI-prepared/full-prepared proxy trials, are historical. They are not these results and are not relabeled as current-main evidence.
Independent final-client correctness check on the exact candidate navigated to a stopped synthetic thread, entered text, cleared it with keyboard input and a whitespace-empty assertion, visually inspected the empty editor, and quit Electron normally (exit 0, no signal). Source/package hashes remained unchanged. Earlier controller-only failed attempts were retained; they were not timing samples or product changes.