Two flakes found by main going red over the weekend

.github/workflows/ci.yml

@@ -97,10 +97,14 @@ jobs:
         with:
           workspaces: crates/web
 
+      # Pinned, not `latest`: the resolver once handed one run trunk 0.19
+      # (whose wasm-opt predates the bulk-memory ops rustc now emits) and the
+      # next run 0.21, so the same commit failed and passed an hour apart.
+      # Matches the version developed against locally.
       - name: Install Trunk
         uses: jetli/trunk-action@v0.5.0
         with:
-          version: latest
+          version: v0.21.14
 
       - name: Check formatting
         working-directory: crates/web

crates/api/tests/experiments.rs

@@ -243,10 +243,17 @@ async fn the_results_series_buckets_by_hour_and_windows_a_week(pool: PgPool) {
 
     // Two events inside one hour, one in the next, and one far outside the
     // window — the first two must merge, the last must not appear at all.
+    //
+    // Anchored to a *truncated* hour, not to `now` minus two hours: the naive
+    // version put `hour_a + 20min` in a different bucket whenever the test ran
+    // past minute 40, which is a 33 % failure window that CI eventually hit.
+    use chrono::DurationRound;
     let now = chrono::Utc::now();
-    let hour_a = now - chrono::TimeDelta::hours(2);
+    let hour_a = (now - chrono::TimeDelta::hours(2))
+        .duration_trunc(chrono::TimeDelta::hours(1))
+        .expect("hour truncation");
     let hour_a_later = hour_a + chrono::TimeDelta::minutes(20);
-    let hour_b = now - chrono::TimeDelta::hours(1);
+    let hour_b = hour_a + chrono::TimeDelta::hours(1);
     let ancient = now - chrono::TimeDelta::days(30);
     let events = json!({
         "events": [