Flaky test: the failure-facts reader always reads once before checking the writers (#4312)

[erikdarlingdata]

Closes #4312.

Why

TheNewestFailuresFactsAreNeverABlendOfTwo polls the writer tasks with a while (!writers.All(w => w.IsCompleted)) loop and counts each iteration in observations, then asserts observations > 0 after the writers finish. A while loop checks the exit condition before the body ever runs. On a fast run, or a slow test thread, all five writer tasks can finish (WritesPerWriter = 200_000 each) before the test thread's first check of writers.All(...), so the loop body never executes, observations stays 0, and the "the reader observed nothing" assert fails. That is the flake filed as #4312.

The same shape exists in UnderConcurrency_TheInstanceTotalNeverTrailsItsParts a few hundred lines above: a while loop over its own two writer tasks, counting observations, with the same Assert.True(observations > 0, ...) after the loop. It has the identical race and was fixed in this PR too, per the brief's instruction to fix every loop of this shape in one pass. A full grep of both test projects (git grep -n "while (!.*IsCompleted" -- Darling/Darling.Tests Lite.Tests, plus a broader while (! sweep) found only these two occurrences; no other writer-polling loop of this kind exists in either project.

What changes

In Darling/Darling.Tests/AlertReadFailureSurfaceTests.cs:

• Both loops changed from while (condition) { body } to do { body } while (condition);, so the reader body runs at least once regardless of how fast the writers finish.
• The comment above each Assert.True(observations > 0, ...) was reworded to say the observation count is now guaranteed by the loop shape rather than hoped for from scheduling luck. The first test's comment also notes the pre-seeded bucket makes that guaranteed first observation a complete trio, matching what the assert right below it already checks.
• No behavior of the counter under test changed, and the body of each loop is untouched.
• UnderConcurrency_TheInstanceTotalNeverTrailsItsParts also now seeds both parts before its writers start, the same way TheNewestFailuresFactsAreNeverABlendOfTwo seeds its bucket: the do guarantees a first observation, but without this seed that first observation (and any others before the scheduler resumes the writers) can still see both parts at zero, so the observedBothParts > 0 liveness assert could still fail under starvation. The settled counts (2 * WritesPerWriter + 1) and the comment above the liveness asserts were updated for the extra seeded failure.

Test plan

• dotnet build Darling/Darling.Tests/Darling.Tests.csproj — 0 Warning(s), 0 Error(s).
• Darling.Tests.exe -class "*AlertReadFailureSurfaceTests*" run 5 times in a row: 27/27 passed each time, 0 failed.
• Darling.Tests.exe -class "*DocCommentHygiene*": 77/77 passed.
• Full suite: not run here, CI runs it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3