PR #28374 real dev verification
main: 8b74a69019 (github/main)
PR: 33c7ea5405a8f0b852cde1167f332d716171a1e4
Linux; Node v24.22.1; Chrome via repo Playwright 1.60.0
One background dev server at a time. No mocks or seeded fixtures.
Project and initial files created through web UI and Codex gpt-6.1-sol.

main running-server Git trace:
2026-11-01T02:20:01.020388Z ["git","ls-files","--others","--exclude-standard","-z","--"]

pr running-server Git trace:
2026-11-01T02:21:43.835743Z ["git","ls-files","--others","--exclude-standard","-z","--",":(top,literal)bulk[1]/",":(top,literal)new [1].txt"]
2026-11-01T02:23:13.537024Z ["git","ls-files","--others","--exclude-standard","-z","--"]

Observed web client outcomes:
main: Changes +4 -0.
PR, same repo/data: Changes +4 -0; literal scan roots bulk[1]/ and new [1].txt.
PR Changes review: bulk[1]/a.txt +2, new [1].txt +1, README.md +1; ignored.txt absent.
Second real Codex turn on PR staged only README.md and set status.showUntrackedFiles=no.
git diff --cached --numstat: 1 0 README.md
git status --short: M  README.md
Refreshed web client: Changes +4 -0; Git trace confirms full-scan fallback.

Blacksmith focused regression command:
vp test run apps/server/src/vcs/GitVcsDriverCore.test.ts -t "counts many untracked files exactly|reads Changes totals with untracked files"
Result: 2 passed, 210 skipped; 1 test file passed.

Unverified: native Windows/macOS, Electron/native mobile, Node 26.12.1, latency/CPU improvement on large repositories, remote/tunnel. No performance claim from this small repository.