Retain bounded invalid ASR output and atomic failure evidence

tools/speech-eval/README.md

@@ -115,5 +115,15 @@ SHA at completion while the on-disk harness had changed; that field cannot prove
 the exact executing harness bytes. Its original raw token failures remain valid
 diagnostic observations, and its excluded window already prohibits an accepted
 quality score. Preserve that historical limitation instead of rewriting output.
+The first DTW run stopped on strict UTF-8 parsing in the second piano window
+(offset 30000 ms): native JSON bytes at 1137–1138 were invalid. Its temporary
+raw output was not retained by that harness, so exact malformed bytes/native
+cause cannot be established. The partial checkpoint retains all three NASA
+windows (21 accepted DTW sentence spans, including one independently recorded
+ordinary-interval failure) and the first piano window. Missing remaining windows
+and sources prohibit any accepted retrieval score. It was not rerun or decoded
+lossily. Subsequent harnesses retain bounded exact failing output bytes under
+the owned scratch parent with SHA/size/parse position, and atomically record a
+failed incomplete checkpoint before terminating; they do not skip competitors.
 References: pinned [token data/API](https://github.com/ggml-org/whisper.cpp/blob/d1be6fde11ac6e0407606b4e42fe72d34add8037/include/whisper.h)
 and [DTW implementation](https://github.com/ggml-org/whisper.cpp/blob/d1be6fde11ac6e0407606b4e42fe72d34add8037/src/whisper.cpp).