[codex] Follow up local-ai staging for unpackaged Electron runs

[joes021]

Summary

• support unpackaged Electron source runs by resolving staged build/local-ai/<platform>-<arch>/bin assets relative to the repo module path
• keep staged local runtime binaries out of git status by ignoring build/local-ai/* while keeping the checked-in README
• preserve the merged local inference packaging flow while making source smoke tests reproducible on Linux dev machines

Root cause

After PR #172 merged, source Electron smoke tests on Linux still fell back to GitHub release downloads. The app was resolving staged local-ai assets from app.getAppPath(), which can point at the smoke harness directory instead of the repository root during unpackaged runs. Separately, staged runtime binaries were local build artifacts, so clean clones needed a restaging step and those artifacts should not pollute git status.

Validation

• node --test tests/localInferenceAssets.test.js tests/localInferenceProgress.test.js
• Spark ARM64: source Electron smoke under xvfb after restaging linux-arm64 binaries from the installed app
• r5950x x64: source Electron smoke under xvfb after restaging linux-x64 binaries from the installed app