[iLevyTate]

Summary

Two changes, both prompted by the startup follow-ups:

1. Ship only each build's own better-sqlite3 prebuild (431b4c1).

• better-sqlite3 13 bundles N-API prebuilds for eight platform/arch combinations (about 16.9 MB). Only Windows excluded other OSes, and it still kept both win32 arches; macOS and Linux shipped all eight.
• lib/binding.js getPrebuildPath() loads prebuilds/<platform>-<arch>.node, or linuxmusl-<arch>.node on musl Linux. Each platform now keeps only that file through an arch-expanded extglob (app-builder-lib expands ${arch} per packed arch, as Close the startup follow-ups' known limitations: driver-version reset, embedding worker, x64 payload, Linux raster #53 already relies on for the node-llama-cpp prebuilts):
  • win.files: !**/node_modules/better-sqlite3/prebuilds/!(win32-${arch}).node
  • mac.files: !**/node_modules/better-sqlite3/prebuilds/!(darwin-${arch}).node
  • linux.files: !**/node_modules/better-sqlite3/prebuilds/!(linux-${arch}|linuxmusl-${arch}).node
• Estimated savings: about 15 MB / 7 files per macOS build, about 12 MB / 6 files per Linux build, one more 2 MB file per Windows build. Linux keeps the same-arch musl prebuild because binding.js selects it when glibc is absent (2.4 MB; cannot be tested here, so kept).
• Without its own prebuild binding.js falls back to build/Release, which is not packaged, so SQLite would fail to load. verify-packaged-artifacts.js now requires each build's prebuild on Windows, macOS and Linux and fails if one for another platform or arch is packaged.
• docs/RELEASING.md (payload-trim table) and the CHANGELOG [Unreleased] describe the change.

2. Fix the e2e first-run wizard race for every spec (c6d263a, c2ec12e).

• In a model-less E2E profile the first-run model setup wizard opens about 2 s after ready, after waitForAppReady's single dismissModals() pass; its aria-modal overlay then swallowed the "Relate" click in Search Modal - Cluster Context Menu on this PR's first e2e-smoke (ubuntu-latest) run (1 failed, 2 flaky). 32 of the 40 specs that launch the app (six in the CI smoke list) had the same race; only eight called dismissModelSetupWizard().
• waitForAppReady() now takes dismissModelWizard (default: the value of dismissModals) and, when set, waits for and skips the wizard last. Callers that keep modals open on purpose (dismissModals: false, e.g. the accessibility audit of the wizard) still get it untouched.
• dismissModelSetupWizard() remembers windows whose wizard it skipped (skipping latches it closed for that renderer session), so the explicit calls already in eight specs return at once instead of waiting 7 s.
• c6d263a first added explicit calls to search-modal.spec.js; c2ec12e moves the fix into the helper and returns that spec to plain waitForAppReady(). Net change to the spec: none.
• Cost: about the 2 s the wizard takes to open per launch, or up to 7 s when no wizard appears.

Scope

• Area(s): packaging (electron-builder.json), release verification script, E2E helper, docs
• Risk Level: low to medium (the prebuild rules are only exercised by packaged builds, guarded by verify-packaged-artifacts.js in every release job; the E2E helper change affects every spec's startup)
• Breaking Changes: none

Type of Change

• Bug fix
• Enhancement
• New feature
• Refactor
• Documentation update
• Breaking change

Test Plan

Automated

• npm run format:check
• npm run lint
• npm test (as part of npm run ci on 431b4c1: 774 suites, 11,231 passed, 26 skipped, coverage thresholds met; audit:production 0 vulnerabilities; build OK)
• Additional tests added/updated when behavior changed: electron-builder-config.test.js checks all six platform/arch builds with app-builder-lib's own node_modules filter; verifyPackagedArtifacts.test.js covers the prebuild policy, a missing own prebuild, a foreign-arch prebuild and the Linux musl allowance.

Also: app-builder-lib's validateConfiguration accepts all three builder configs; the two unit test files pass with the starting platform faked to darwin and win32 (80/80 each). E2E locally under xvfb-run: with explicit per-spec dismissal the two describes that failed passed 4/4 and the first 28 search-modal tests passed, with the wizard detected and skipped in each; the full CI smoke list is running locally against the helper version.

Manual QA

• Manual checklist completed (see docs/TESTING.md)
• Not applicable (explain below)

Manual QA notes:

• Packaged builds are not produced in CI for PRs; the release workflows' verify-packaged-artifacts.js step now fails if a build lacks its own SQLite prebuild or carries another platform's.

Checklist

• I scoped this PR to one logical change
• I self-reviewed code and naming for clarity
• I updated docs/config references where needed
• I called out follow-up work or known limitations

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q5Uo5hC58DRSh1qB489Sos

[iLevyTate]

e2e-smoke (ubuntu-latest) failed on 431b4c1: test/e2e/search-modal.spec.js:1093 › Search Modal - Cluster Context Menu › should have cluster visualization in graph view (plus two flaky tests in the same file), 88 passed. Each timed out clicking the "Relate" button because an element inside a role="dialog" overlay intercepted the click; that element is the step list of ModelSetupWizard.jsx ("Set up on-device AI").

This PR does not cause it: it changes only electron-builder.json files patterns, scripts/verify-packaged-artifacts.js, their unit tests and docs. The e2e job runs the app from the source build, where electron-builder's file filters are never applied. e2e-smoke (windows-latest) passed on the same commit, and the push runs on main (#50 to #53) passed e2e.

Cause: the Cluster Context Menu and Quick Actions on Hover describes call only waitForAppReady(), whose single dismissModals() pass can run before the model setup wizard mounts. test/e2e/helpers/electronApp.js documents exactly this race at dismissModelSetupWizard(): a late wizard's aria-modal overlay swallows later clicks. Other specs call that helper after waitForAppReady(); these two describes do not.

Fix (not pushed here, to keep this PR to one change): add await dismissModelSetupWizard(window); after waitForAppReady(window) in those two beforeEach blocks. I will re-run the failed job once when this workflow run finishes.

---

Generated by Claude Code