huskies: merge 1175 bug Flaky tests intermittently fail merge gates

server/src/agents/pool/pipeline/merge/runner.rs

@@ -32,7 +32,19 @@ impl AgentPool {
     /// Called at the top of [`start_merge_agent_work`] to unblock retries,
     /// and also by the periodic background reaper in the tick loop so stale
     /// entries are cleaned up even when no new merge is triggered.
+    ///
+    /// A job's `server_start` round-trips through JSON text (see
+    /// [`encode_server_start_time`]/[`decode_server_start_time`]), and
+    /// `serde_json`'s float parser is not guaranteed bit-exact for
+    /// high-precision Unix timestamps — it can decode a value a couple of
+    /// ULPs below the original. Comparing with a bare `<` against a
+    /// freshly-read `current_boot` would then occasionally treat a job
+    /// written by *this very server instance* as belonging to a previous
+    /// one. Real server restarts are always seconds apart at minimum, so a
+    /// generous tolerance absorbs that noise without weakening genuine
+    /// stale-boot detection.
     pub(crate) fn reap_stale_merge_jobs(&self) {
+        const STALE_TOLERANCE_SECS: f64 = 1.0;
         if let Some(jobs) = crate::crdt_state::read_all_merge_jobs() {
             let current_boot = server_start_time();
             for job in jobs {
@@ -40,7 +52,7 @@ impl AgentPool {
                     continue;
                 }
                 let stale = match decode_server_start_time(job.error.as_deref()) {
-                    Some(t) => t < current_boot,
+                    Some(t) => t < current_boot - STALE_TOLERANCE_SECS,
                     None => true, // Legacy (pid-encoded) or malformed: stale
                 };
                 if stale {