PR 31969 runtime evidence Run the same driver against fresh homes and unchanged real server checkouts. No unit tests, mocked providers, fake forge CLI, accelerated clocks, database seed copies, or changes to the server implementation are used. Base: 70a7cfe5a20f6ea82f5a5b89fe08da13c468d80f Head: bb63fa7c260956ddda1f3bcd1204ff4ce694de17 Startup (in each checkout): T3CODE_DEV_SEED=0 T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0 \ node scripts/dev-runner.ts dev:server --home-dir --no-seed --port Seed (driver.mjs): 1. Create a synthetic git repository with a gitlab.com origin pointing to a missing project. 2. Use authenticated real WebSocket RPC to create a project and an idle thread. 3. Link missing merge request 1 through orchestration.dispatchCommand. 4. Leave each server running without a restart or a manual refresh for at least 16 minutes. 5. At 16 minutes 30 seconds request pullRequests.invalidate for that exact reference. 6. Keep watching until the next background read, including the 15-minute capped delay. observe.py records the live server trace monitor at approximately one frame per wall-clock second; the displayed times come from actual trace timestamps. Each side has the same synthetic repository, idle thread, reference, driver, monitor theme, and viewport. Only the server commit and its private fresh home differ. The two seeds are approximately 22 seconds apart. Elapsed times are relative to each link. Server sweeps start at server startup, not at link creation, so due reads are rounded up to a subsequent sweep. A link created between sweeps can wait almost two minutes for its first retry when a minimum one-minute delay is required. The JSONL extracts preserve actual span IDs, trace IDs, UTC timestamps, outcomes, and durations. They omit stack traces and filesystem attributes to avoid publishing local paths. A syncGroup span is one linked-MR synchronization attempt; glab process spans also include the GitLab adapter's supporting commands and interrupted siblings. No real model turn was needed for a server-only background retry bug. The thread's model selection is the free devReplay provider, but the thread stays idle. Limits: the missing project cannot return a successful read, so success resetting the backoff was not exercised. The explicit refresh is tested after the backoff has reached its 15-minute cap; growth following a refresh below that cap was not exercised. Recording recovery: the monitor process exited 143 before MP4 finalization. The recovered recording copies the existing H.264 frames, using matching x264 codec parameters from a short reference. One incomplete final NAL unit was discarded. No replacement frames or simulated timeline frames were generated. The full recovered recording is 1913 frames at one frame per second. The paired previews crop its two columns and play at 60x. recover-recording.py documents that conversion.