PR #35631 real development-server verification main: ed6cb97f07df59c01467b919073550d310da3ef4 head: cbc6116b5494483b6249d29fcdc162f86e1997b6 Linux, Chrome /usr/bin/google-chrome, repo Playwright, Codex gpt-6.1-sol through local proxy Data created through the app: a local git project, real provider turns, 270 independent repositories with committed README files. No mocks, fixtures or demo data. Each run uses its own server/state. Main stopped before the PR server started. Browser-authenticated WebSocket /ws?orchestrationProtocol=2: Request subscribeVcsStatus {cwd, includeRemote:false} receive snapshot; Ack; Interrupt to release the subscription visit repo-000 through repo-269, holding repo-001 watched real Codex turn creates cache-marker.txt in repo-000, repo-001, repo-269 main oldest revisit: hasWorkingTreeChanges=false (original snapshot retained) main explicit vcs.refreshStatus localOnly=true: hasWorkingTreeChanges=true PR oldest before: false; after churn and marker edit: true (snapshot reloaded) PR newest before/after marker edit: false/false (recent snapshot reused) PR watched before/after churn and edit: false/false (watched snapshot retained) PR local-only refresh: true; live subscriber received localUpdated PR full refresh: true; no-upstream remote status unchanged PR two watchers: cancel first, Codex switches repo-001 to verification-neighbor; remaining watcher receives localUpdated with new ref and clears old remote PR cancel last watcher, Codex removes marker: cached dirty before churn=true; after another 268 repository visits=false (released snapshot can be evicted) PR provider turn-end checkpoint: ui-marker.txt appears as +1 in client diff All runtime assertions above passed against captured responses. Blacksmith: broadcaster suite 42/42; check:changed against parent 4f555c74 passed format, lint, knip, GraphQL, three package typechecks, 44/44 tests. Forge: 34 successful checks, zero pending/failing at inspected head. Not verified: heap/RSS reduction, latency, externally hosted PR metadata under churn, in-flight remote-operation race in a real network call, >256 simultaneous watchers, native desktop/iOS/Android, Windows/macOS, relay/tunnel. Existing focused tests cover remote-operation retention and release-time trimming; they are not runtime proof.