Classify DEVICE_LOST generate failures in wrapperHEADmaster

A failed /generate now keeps the server's JSON error body instead of
discarding it (dropped curl -f). A level_zero DEVICE_LOST body prints a
clear "GPU wedged ... reboot needed" message; other failures surface the
actual server error. Root cause of the wedge is a GuC firmware / xe
driver version mismatch (GuC 70.65.0 vs driver-wanted 70.49.4), a
host-side fault the wrapper cannot fix, only report clearly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

1 files changed, 12 insertions, 3 deletions

diff --git a/imggen b/imggen
index b385712..193b10d 100755
--- a/imggen
+++ b/imggen
@@ -115,12 +115,21 @@ if daemon_up; then
     [ -n "$H" ]     && req="$req,\"height\":$H"
     req="$req}"
 
-    if curl -sf -X POST "${BASE}/generate" -H 'Content-Type: application/json' \
-        -d "$req" -o "$OUT_PATH"; then
+    # No -f: keep the error body so we can classify it (e.g. GPU wedged).
+    code="$(curl -s -w '%{http_code}' -X POST "${BASE}/generate" \
+        -H 'Content-Type: application/json' -d "$req" -o "$OUT_PATH")"
+    if [ "$code" = 200 ]; then
         echo "Saved ${OUT_PATH}"
     else
-        echo "generate failed (see daemon logs: docker compose logs)" >&2
+        # On failure the body is the server's JSON error, not a PNG.
+        err="$(cat "$OUT_PATH" 2>/dev/null)"
         rm -f "$OUT_PATH"
+        # level_zero DEVICE_LOST = GuC/xe driver hang; only a reboot recovers.
+        if printf '%s' "$err" | grep -qi "device.lost\|DEVICE_LOST\|error 20"; then
+            echo "GPU wedged (level_zero DEVICE_LOST): GuC/xe driver fault, reboot needed" >&2
+        else
+            echo "generate failed: ${err:-see daemon logs: docker compose logs}" >&2
+        fi
         exit 1
     fi
 else