[Aaronontheweb]

Deflakes CanDrawWithTouch, which intermittently fails its first navigation with net::ERR_NETWORK_CHANGED.

Root cause

Chromium aborts in-flight requests when it observes a host network-state change. Docker creating/removing container bridge interfaces while Aspire starts resources produces exactly that. This is a browser/transport behavior, not an app error.

Evidence

From run 40782982599, job Test-ubuntu-latest:

[xUnit.net] DrawingTouchInteractionTests.CanDrawWithTouch(viewPortX: 768, viewPortY: 1024) [FAIL]
Microsoft.Playwright.PlaywrightException : net::ERR_NETWORK_CHANGED at http://localhost:34677/
  - navigating to "http://localhost:34677/", waiting until "load"

• The failing case died in 114 ms versus 1-3 s for the three passing viewports - it aborted before the page rendered anything.
• Rerunning the failed jobs passed all four viewports.
• Same signature documented in microsoft/playwright#13062 (container, retry clears it) and Grizzly-Endeavors/residuum#592 (Docker bridge churn named as the cause).
• Kestrel logged no connection abort server-side. A dedicated Akka.NET/Aspire review found no mechanism in this test configuration - single-node cluster, local-only actor on the /paint path, no management/bootstrap/discovery, no restart policy - that can abort an in-flight request.

Fix

GotoWithNetworkChangeRetryAsync retries only when the exception message contains ERR_NETWORK_CHANGED, up to 3 attempts with 250 ms/500 ms backoff. Every other PlaywrightException propagates unchanged, so real failures stay real. The abort happens before the page has rendered, so retrying that specific navigation is side-effect free.

Applied to the first navigation in both interaction tests - DrawingMouseInteractionTests has the identical unguarded pattern and the same exposure.

Deliberately out of scope

• The separate 5s locator timeout for a[href='NewPaint'] seen on dev in August (run 33213636917) - a Blazor render race after form submit, not a network change. Two independent timing assumptions.
• The unguarded _localActorHandle.Tell at Paint.razor:268 - an IActorRef? field, assigned at line 162, no null-guard, while lines 240/246 do use ?.. It logs an NRE on every prerendered /paint/{id} teardown, but it is causally independent of this flake (Kestrel swallows it after the response completes). Tracking separately.

Verification

Build: 0 errors. Full verification is CI-only - Playwright browsers and a warm SQL Server image are not available on this host, and the flake is intermittent (~2 known failures over months), so a green run is necessary but not sufficient signal.