[phil-scott-78]

Why

The ci run on main failed on FileWatcher_CoalescesRapidBurst_IntoSingleNotification — a flaky test, not a product regression. This PR deflakes it, sweeps the suite for the same class of defect (assertions on outcomes the OS/scheduler/clock control), and clears the one build warning that was riding along.

Changes

test: deflake timing-sensitive FileWatcher and AuditRunner tests

• FileWatcher_CoalescesRapidBurst drove a real OS FileSystemWatcher on the real clock and asserted a 6-write burst debounced to exactly one notification. Under load the watcher's event pump stalls >100 ms mid-burst, elapsing the 100 ms debounce window and splitting the burst into two notifications — the failing CI run. Now driven off a FakeTimeProvider: raw events still arrive in real time, but the debounce timer only fires on Advance, so load can no longer split the window. The production debounce is unchanged.
• AuditRunner tests — the headless-one-shot test polled the cache for up to 500 ms after ApplicationStarted (can expire under thread-pool starvation); the dev-mode test slept a flat 100 ms before asserting the rendered auditor never ran. Both now await the purpose-built WaitForInitialPassAsync() seam, removing the timing entirely.

test(docsite): fix CS8604 nullable-argument warning in area-home test

• IElement.GetAttribute("href") returns a nullable string; passing it straight to ShouldContain tripped CS8604. Assert the href is present first, then match — clears the warning and makes a missing href fail loudly.

Audit of the rest of the suite

Swept all four test projects for the same pattern. Everything else is either well-constructed or can't false-fail:

• BuildHtmlCacheTests concurrent coalesce — TaskCompletionSource-gated, deterministic.
• RenderedHtmlFetcherTests 50-way concurrent — stub-backed race stress test; a failure signals a real race, not flakiness.
• RedirectMiddlewareTests (real Kestrel socket) — ephemeral port :0, awaits StartAsync() (no raced sleep).
• The two other FileWatcherTests real-OS-watcher waits use a generous 5 s ceiling on a positive "did it fire" assertion.

Verification

• Both test projects build with 0 warnings.
• Reproduced the original flake 3/8 under CPU saturation on the old FileWatcher test; the rewritten test passes 12/12 and AuditRunner 6/6 under the same load.

[github-actions]

🛰️ Docs preview: https://pr-66.pennington-dev.pages.dev

Rebuilt on every push to this PR; torn down when it closes.