test: allow for EG v1.9.2 shutdown time in TestRunExtprocStartFailure

[abdulaziz-dev-lab]

Description

TestRunExtprocStartFailure has been flaky on both ubuntu and macOS since the Envoy Gateway v1.9.2 bump (#2787). It fails with "expected extproc start to fail promptly" after its 10s timeout, on main and on PRs alike.

When the extproc fails, run cancels the Envoy Gateway server's context and returns once server.ExecuteContext returns. Since v1.9.2, EG's shutdown waits for every runner goroutine to exit [1][2]. That includes the file provider's Start, which runs the initial load of the resources even when the context was canceled before it began. With v1.9.0 nothing waited for it.

With the CI flags (-race, atomic coverage), pinned to 4 cores, the time from the cancel to ExecuteContext returning in this test went from 0.15s to 3.12s (median of 10 runs). The wait is mostly single-threaded, and on CI runners, which run several test packages at once, it often goes past 10s. Running the test on one core shared with two busy loops reproduces the failure on main (10.01s). With this change it passes (12.01s).

The timeout only needs to catch run hanging after the extproc fails, which is what #1134 fixed. This raises it to 60s and adds a comment explaining why.

Related Issues/PRs (if applicable)

Fixes #2793

Special notes for reviewers (if applicable)

Making run itself return faster would need EG to skip the initial load once the context is canceled, which is an upstream change.

1: https://github.com/envoyproxy/gateway/blob/v1.9.2/release-notes/v1.9.2.yaml
2: envoyproxy/gateway#10128

[netlify]

✅ Deploy Preview for theagentrouter canceled.

Name	Link
🔨 Latest commit	3c05e65
🔍 Latest deploy log	https://app.netlify.com/projects/theagentrouter/deploys/6ac4590a2d14710008eb0dc3

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[abdulaziz-dev-lab]

/retest

[abdulaziz-dev-lab]

A note on this PR's CI history.

The first Build and Test run passed everything except Data Plane Test (macos-latest). Both Unit Test jobs, where TestRunExtprocStartFailure runs, passed. The data-plane failure was in TestOtelOpenAIEmbeddings_metrics_modelNameOverride. Before any request was made, the extproc couldn't bind its admin port (listen tcp :49906: bind: address already in use), so the test environment never came up and the test failed after 60s. That's a port collision in the data-plane test setup, unrelated to this change, which only touches cmd/aigw/run_test.go.

https://github.com/theagentrouter/agent-router/actions/runs/37402998623/job/112074199353

I ran /retest once. It re-ran only that job, which passed, and ci-required is now green.

[siddharth1036]

LGTM
minor nit, after timeout the t.Fatal still logs expected to fail promptly, worth aligning the log with 60s window

[nacx]

Great catch. Thank you!

[abdulaziz-dev-lab]

@siddharth1036 Thanks for the review. Good point about "promptly". The message also blames the extproc startup even though the extproc fails right away. The timeout actually catches run not returning afterwards. The comment above the select explains this, and the FAIL line shows the elapsed time. So I'll leave it as merged.

@nacx Happy to send a one-line follow-up if you'd like one.