diff --git a/crates/fbuild-packages-fetch/src/downloader.rs b/crates/fbuild-packages-fetch/src/downloader.rs
index 64ecf373..c35bc589 100644
--- a/crates/fbuild-packages-fetch/src/downloader.rs
+++ b/crates/fbuild-packages-fetch/src/downloader.rs
@@ -619,15 +619,23 @@ async fn fetch_into_part(
     let mut last_report = Instant::now();
     let mut last_pct: u32 = 0;
 
+    // Set when the body stops arriving. Recorded rather than returned, so the
+    // flush below still runs — see the comment there.
+    let mut body_error: Option<DownloadAttemptError> = None;
+
     loop {
         let chunk = match tokio::time::timeout(timing.chunk_read_timeout, response.chunk()).await {
             Ok(Ok(Some(chunk))) => chunk,
             Ok(Ok(None)) => break,
-            Ok(Err(error)) => return Err(DownloadAttemptError::Body(error)),
+            Ok(Err(error)) => {
+                body_error = Some(DownloadAttemptError::Body(error));
+                break;
+            }
             Err(_) => {
-                return Err(DownloadAttemptError::BodyStalled {
+                body_error = Some(DownloadAttemptError::BodyStalled {
                     filename: filename.to_string(),
                 });
+                break;
             }
         };
         file.write_all(&chunk)
@@ -664,6 +672,14 @@ async fn fetch_into_part(
 
     // Flush before the caller stats the file to decide whether this attempt
     // made progress — buffered bytes would read as a stall.
+    //
+    // FastLED/fbuild#1370: this has to run on the *failure* path too, which
+    // is the whole point. A dropped connection used to return from inside the
+    // loop above, skipping this; `tokio::fs::File` makes no promise to flush
+    // on drop, so the bytes that did arrive could vanish. The retry loop then
+    // stats a short `.part`, resumes from a stale offset, and asks the server
+    // for a range it already has — losing exactly the progress this feature
+    // exists to keep, on the one path where keeping it matters.
     file.flush()
         .await
         .map_err(|error| DownloadAttemptError::PartFile {
@@ -671,6 +687,10 @@ async fn fetch_into_part(
             error: error.to_string(),
         })?;
 
+    if let Some(error) = body_error {
+        return Err(error);
+    }
+
     // A short body is a dropped connection that happened to end on a chunk
     // boundary. Treat it as a retryable failure so the resume loop continues
     // rather than renaming a truncated archive into place.
diff --git a/crates/fbuild-packages-fetch/src/downloader_tests.rs b/crates/fbuild-packages-fetch/src/downloader_tests.rs
index 3e903d92..c36111ed 100644
--- a/crates/fbuild-packages-fetch/src/downloader_tests.rs
+++ b/crates/fbuild-packages-fetch/src/downloader_tests.rs
@@ -457,6 +457,15 @@ async fn streaming_download_resumes_from_the_byte_offset_after_a_drop() {
         RESUME_BODY,
         "the resumed file must be byte-identical to the source"
     );
+    // Load-bearing, and deliberately exact. A third request means the resume
+    // asked for a range it already had — i.e. the bytes the first attempt
+    // delivered were lost and the retry started over. That is the failure
+    // FastLED/fbuild#1370 exists to prevent, and this count is the only thing
+    // that detects it: the file still ends up correct either way, so every
+    // other assertion here passes while the feature silently does nothing.
+    //
+    // It caught exactly that on macOS runners, where the unflushed `.part`
+    // read short.
     assert_eq!(
         request_count.load(Ordering::SeqCst),
         2,