Measured on Windows with Node 24.19.0, October 14, 2026. Original source baseline:
107f6024ea4786a58a83c921728a113b849de192. Source fixes were subsequently split onto
current upstream main at 8b460a3c851d629df896f7420aa32eb32df1daca for submission.
The files contain benchmark measurements and profiling harness source. Fixture databases, credentials, application content, and raw trace logs are excluded.
listener-probe.json: three runs per command/count; about 50 pre-existing listener rows plus 0/30/200 real Node TCP listeners. Timings include process startup. At 200 added listeners, median original query 3,706.46 ms, netstat 35.90 ms. This is a listener-command comparison, not a whole-app speedup.git-status.json: 20 interleaved before/after samples per state in a warm 1,000-file temporary repository. Clean: 108.06 → 71.08 ms; untracked-only: 99.94 → 64.68 ms. Median uses the mean of the two middle samples. Both remove one subprocess (3 → 2). The unchanged modified-checkout path is not claimed faster.trace-flush.json: three interleaved runs per version, 7,680 records in 30 batches of 256, about 8 MB, 5 MB rotation and three backups. Producer-call median 56.24 → 26.16 ms. Total wall-time median 60.14 → 65.06 ms; asynchronous scheduling is not a disk-throughput improvement. Warm filesystem.profile-comparison.json: two 120-second, 1 ms CPU captures of an isolated bundled server with roughly 16,234 retained events and ten projects. Synchronous append stack samples 5,048.6 → 0 ms. These include startup and different client activity on a shared machine, so they do not establish a controlled whole-app percentage. Event-loop histograms reset each second and are not a complete stall inventory.ui-final.json: current-state web development client in the T3 Browser panel, 1280×800, seeded conversation, no messages sent. A 2,000-character draft followed by 23 individual keys, then alternating virtualized-timeline scrolling. Individual input to next animation frame median 11.4 ms, p95 14.7 ms. Bulk input about 96 ms. Frame callback timing is not guaranteed physical presentation. Browser measurements do not validate native Electron IPC/FPS or establish a before/after UI improvement.
Held-write, failure, rotation, retry, and topology tests provide deterministic behavioral evidence separately from these timing samples. The PR descriptions name the exact focused test commands and current-branch validation results.
Implemented and measured using GPT-6.1 Sol in T3 Code through the Codex harness.
Electron 44.4.5 / Chrome 152.0.7977.130, Windows x64. Built preloads from current upstream main (before) and the desktop topology branch (after) ran in actual sandboxed, context-isolated BrowserWindows with hidden fixture pages and separate task-owned profile/session directories. No full T3 app, live database, provider, or account was used.
Each sync/async topology response was deliberately delayed by 30 ms at the main-process IPC seam. After one bootstrap/enabled warmup read each, thirty read pairs ran while a five-millisecond renderer heartbeat was active. The after workload started one async refresh before those reads and waited for it afterward. Three interleaved runs per version.
| Metric | Before | After |
|---|---|---|
| Read-pair loop median | 2,342.2 ms | 0.4 ms |
| Longest heartbeat gap, median of runs | 2,343.3 ms | 6.2 ms |
| Synchronous calls including two startup warmups, per run | 62 | 2 |
| Asynchronous refresh calls, per run | 0 | 1 |
This deliberately delayed-reply benchmark proves the native renderer avoids recurring sync IPC while a refresh is pending. It is not a natural-workload typing/FPS measurement, and its ratios should not be presented as overall desktop speedups. The two startup sync reads are intentionally retained. CPU profiles also include these warmup reads.
The native harness and build/launch script are attached along with numeric results and
one CPU profile per version. The helper expects the two named PR worktrees beside it and
uses an isolated D-drive directory; adjust those paths when reproducing elsewhere.
backgroundThrottling: false keeps the hidden fixture's heartbeat active.
sandbox: true and contextIsolation: true exercise the production preload boundary.
main-response delay = 30 ms is a synthetic fixture, not a measured real main-process
delay. There were no screenshots or recordings of a changed visual layout.