# Azure DevOps Issues loading comparison

The same real web app, Windows Server 2025, Node 26.13.1, Azure CLI 2.93.0 and azure-devops extension 1.0.11. Fresh isolated T3 homes; 13 real temporary Azure DevOps projects, each with one synthetic Task; no artificial command delays. Chromium headless, 1440x900, dark theme.

The app uses the actual Azure DevOps CLI and the actual Issues page. The driver observes the page's own WebSocket requests, awaits successful list responses, verifies all 13 synthetic titles and a hash of their IDs/titles/states, and waits until Refresh issues is enabled again. It does not invoke the tracker directly.

`azure-seed.py seed` creates the temporary remote fixtures in the test organization; `cleanup` deletes them. Its `create` mode first creates project 01. The scripts read credentials from environment variables, never from literals. `fixtures.json` is deliberately excluded because it contains the test organization/repository addresses.

Check out base and head as sibling `base` and `head` folders beside `measure.mjs`, install dependencies with `vp i`, and start each real app with `start.ps1 <base|head>`. The startup script reads the local test.env file and starts each side with an empty, separate T3 home and automatic seeding disabled. On Windows, launch long-running commands using Win32_Process.Create as described in the task environment; an ordinary SSH session may end before the benchmark completes.

Run `node measure.mjs <base|head> seed` once per app. Then run `node measure.mjs <base|head>` one side at a time. The focused comparison uses `node measure.mjs <base|head> focus`, also one side at a time. Every invocation creates a timestamped output directory; no previous results are deleted. The run navigates to Issues, refreshes seven times, then refreshes once more and dispatches pointermove plus window focus 10.5 seconds into the pending load.

In `focus` mode, after one completed priming load, each of five trials reloads Issues and dispatches pointermove plus window focus 10.5 seconds into the pending load. Reloading resets the unrelated five-minute timer and matches the reported entry action. Both commits use exactly the same mode and steps.

`initial-navigation` is kept separate from the seven repeat samples. It includes route navigation and rendering; `rpcMs` separately records the Issues RPC. The head's first uncached load was measured in a preliminary run, before an SSH disconnection, and is reported separately rather than mislabeled as part of the completed run.

The recordings and screenshots hide the issue metadata line (account identity and test organization), using identical visibility-only CSS on both sides; titles, item counts, loading controls and list layout are preserved. No app source is changed. Every final still waits for two identical screenshots 500ms apart.

The original seven-refresh phase is retained in full. Base refresh 5 was interrupted by the real five-minute timer and took 74.797 seconds; no outlier is dropped. This widens the base full range, so the five reopen-and-focus trials supply the primary stable median comparison. The PNG is generated from those raw JSON outputs by render-evidence.py.
