Follow-up startup profile for e94c13bd (release NAPI, /home/firmclaw/fraunces.shift, five warm-filesystem samples):

The working-document open phase improves 8.1x; inspect + open + complete atlas improves 41.6%. The resumed atlas includes lazy SQLite layer acquisition. This measures a complete atlas, not placeholders; renderer transfer/upload is outside this harness. In the app, prepareOpen() now starts the utility while the launcher/dialog is active, so the observed 1.36 s cold utility boot can overlap source selection. The next app-log run should confirm the process is already ready before the dialog returns and that only one inspect-package line occurs.

Background atlas preparation profile for 6c48c586 (Fraunces, release NAPI):

The native profiler independently measured resumed authored compilation at 245.2 ms and aligned preparation at 1.72 ms. The compilation request is now sent from main only after the open response is received, then runs in utility while the workspace window and WebGPU start. The previously observed renderer/window interval was about 500 ms, so this work should be hidden rather than added to open latency. The grid still waits for the complete atlas; no placeholder or partial geometry path was introduced. Output remains 699 roots, 1,290 atlas glyphs, 12,713 curves, 998 components, and 2,190,648 bytes.

Follow-up 09d52877 applies the performance-fixture pattern used by Fontations/Zed/Vello/VS Code: correctness remains in ordinary fixture tests, timings remain in an external repeated benchmark, and no profiling DTOs enter the product APIs.

Native changes:

• read each SQLite layer directory fact once and reuse it for bounded payload/component batches;
• retain unchanged layer-owner/source indexes during payload replacement instead of removing and reinserting them;
• make profile-slug-atlas.mjs run independent clean-resume samples and report p50/p95 (<source> [warm-iterations] [resume-iterations]).

Fraunces release samples:

• clean resume: ~51.0 ms p50;
• resumed authored preparation: 227.760 ms p50, 230.235 ms p95 (previous projection-reuse baseline 242.474 ms);
• aligned preparation: ~1.18 ms p50;
• output unchanged: 699 roots, 1,290 atlas glyphs, 12,713 curves, 998 components, 2,190,648 bytes.

All commit hooks, Rust tests, native bridge tests, and Clippy passed. The existing documentation drift checker still reports the same 13 unrelated main-branch errors.