--- a/gitnexus/src/core/ingestion/workers/worker-pool.ts
+++ b/gitnexus/src/core/ingestion/workers/worker-pool.ts
@@ -1461,6 +1461,7 @@
       // no exclusion attribution, requeueRemainder quarantines items[0]
       // as a best-guess culprit to break the death loop.
       const unattributedJobDeaths: Map<number, number> = new Map();
+      const fileRetries = new Map<string, number>();
       let completedFiles = 0;
       let activeWorkers = 0;
       let stopped = false;
@@ -2045,10 +2046,20 @@
         const recoverAndResume = async (reason: string, excludePaths: readonly string[]) => {
           busySlots.delete(workerIndex);
           inFlightProgress[workerIndex] = 0;
-          requeueRemainder(job, excludePaths);
-          // Keep recovery in flight so another slot finishing cannot settle
-          // this dispatch before the replacement is ready for the next one.
-          await handleWorkerDeath(workerIndex, reason, excludePaths);
+          const failedPath = excludePaths[0];
+          const transient = !/\bSyntaxError\b/.test(reason) &&
+            (/\bEAGAIN\b/.test(reason) || /exited with code 75\b/.test(reason));
+          const retries = failedPath ? fileRetries.get(failedPath) ?? 0 : 0;
+          const retry = !!failedPath && transient && retries < 2;
+          if (retry) fileRetries.set(failedPath, retries + 1);
+          else requeueRemainder(job, excludePaths);
+          // Do not quarantine a classified transient file until its budget is
+          // exhausted. Keep recovery active until replacement and backoff end.
+          await handleWorkerDeath(workerIndex, reason, retry ? [] : excludePaths);
+          if (retry && !stopped) {
+            await new Promise<void>((resolve) => setTimeout(resolve, 10 * 2 ** retries));
+            if (!stopped) jobs.unshift(job);
+          }
           activeWorkers--;
           if (stopped) return;
           // Slot may have been dropped or respawned. Kick the current slot