upload per-fixture iso render for the leaderboard gallery

After uploading the report bundle, also push renders/<id>/<fixture>.png
(the iso view) per fixture so the gallery has standalone, addressable
thumbnails. Skips fixtures with no iso render (invalid/missing), which
the gallery draws as the dashed cell. The full multi-view renders stay
base64-embedded in reports/<id>.html for the self-contained report.

Co-authored-by: Cursor <cursoragent@cursor.com>

eval_job.py

@@ -54,6 +54,14 @@ EVAL_TIMEOUT_SECONDS = 30 * 60
 REPORT_TIMEOUT_SECONDS = 5 * 60
 
 REPORTS_DIR_IN_REPO = "reports"
+RENDERS_DIR_IN_REPO = "renders"
+
+# Single canonical view uploaded per fixture for the leaderboard
+# gallery thumbnail. "iso" matches the GT render the gallery pairs it
+# with, so the gallery columns stay a comparable matrix at one fixed
+# camera angle. The evaluator writes this at
+# ``<run_dir>/<fixture>/renders/iso.png`` (cadgenbench DEFAULT_VIEWS).
+GALLERY_THUMB_VIEW = "iso"
 
 
 def main() -> int:
@@ -94,6 +102,7 @@ def main() -> int:
     _upload_reports(
         submission_id, html_path, report_json, submissions_repo, token,
     )
+    _upload_gallery_renders(submission_id, RUN_DIR, submissions_repo, token)
     print(f"[eval_job] done: {submission_id}", flush=True)
     return 0
 
@@ -241,5 +250,56 @@ def _upload_reports(
     )
 
 
+def _upload_gallery_renders(
+    submission_id: str,
+    run_dir: Path,
+    submissions_repo: str,
+    token: str,
+) -> None:
+    """Upload one ``iso`` thumbnail per fixture for the leaderboard gallery.
+
+    Stages every ``<run_dir>/<fixture>/renders/iso.png`` as
+    ``renders/<id>/<fixture>.png`` in the submissions dataset (one
+    commit). These are the standalone PNGs the gallery's
+    ``renderFor()`` points at; the full multi-view renders stay
+    base64-embedded in ``reports/<id>.html`` for the self-contained
+    report. A fixture with no ``iso.png`` (missing output, or a render
+    that never ran) is simply skipped, the gallery reads the
+    per-fixture status from the row and draws the dashed "invalid
+    generation" cell, so a thumbnail's absence is not an error.
+    """
+    staged: list[tuple[Path, str]] = []
+    for fixture_dir in sorted(d for d in run_dir.iterdir() if d.is_dir()):
+        iso_png = fixture_dir / "renders" / f"{GALLERY_THUMB_VIEW}.png"
+        if iso_png.is_file():
+            staged.append((iso_png, fixture_dir.name))
+
+    if not staged:
+        print(
+            f"[eval_job] no gallery renders to upload for {submission_id}",
+            flush=True,
+        )
+        return
+
+    api = HfApi(token=token)
+    for iso_png, fixture_name in staged:
+        api.upload_file(
+            path_or_fileobj=str(iso_png),
+            path_in_repo=(
+                f"{RENDERS_DIR_IN_REPO}/{submission_id}/{fixture_name}.png"
+            ),
+            repo_id=submissions_repo,
+            repo_type="dataset",
+            commit_message=(
+                f"add gallery render {fixture_name} for {submission_id}"
+            ),
+        )
+    print(
+        f"[eval_job] uploaded {len(staged)} gallery render(s) under "
+        f"{RENDERS_DIR_IN_REPO}/{submission_id}/",
+        flush=True,
+    )
+
+
 if __name__ == "__main__":
     sys.exit(main())