eval_job: upload interface overlay + link GT/input in hosted report

Upload each fixture's interface_overlay.png to the public render bucket
alongside the turntables so every report artifact lands together, and pass
--gt-base-url/--input-base-url to `report single` so the hosted report links
GT/input/overlay (lazy) instead of base64-inlining them. Bump CADGENBENCH_SHA
to 25943a0 (the report generator that understands the new flags).

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

Dockerfile

@@ -51,7 +51,7 @@ RUN python -m venv "$VIRTUAL_ENV" \
 # rebuild picks up the latest evaluator dependencies (including Open3D for
 # alignment). Lock to a commit SHA at the v1 release for reproducible scores
 # (see space-setup/post-gt-swap.md Stage F).
-ARG CADGENBENCH_SHA=3d49822
+ARG CADGENBENCH_SHA=25943a0
 RUN python -m pip install --no-cache-dir \
         "cadgenbench @ git+https://github.com/huggingface/cadgenbench.git@${CADGENBENCH_SHA}"
 

eval_job.py

@@ -86,6 +86,14 @@ RENDER_BUCKET = os.environ.get(
 ).strip()
 HF_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co").rstrip("/")
 
+# Space-relative proxy roots the hosted report references for the *private*
+# assets it can't link from the public bucket. The report is served by the
+# Space at ``/reports/<id>.html``, so these absolute-path URLs resolve against
+# the Space origin and the token-holding proxy streams the bytes. Kept in sync
+# with the routes registered in the leaderboard Space's ``app.py``.
+GT_PROXY_BASE_URL = "/gt"
+INPUT_PROXY_BASE_URL = "/task-input"
+
 
 def _render_base_url(submission_id: str) -> str:
     """Public ``.../resolve/renders/<id>`` base; report appends ``/<fixture>/<file>``."""
@@ -103,17 +111,23 @@ def _upload_renders_to_bucket(
     """
     add: list[tuple[str, str]] = []
     for fixture_dir in sorted(d for d in run_dir.iterdir() if d.is_dir()):
+        dest_prefix = (
+            f"{RENDERS_DIR_IN_REPO}/{submission_id}/{fixture_dir.name}"
+        )
         renders_dir = fixture_dir / "renders"
-        if not renders_dir.is_dir():
-            continue
-        for render_path in sorted(renders_dir.iterdir()):
-            if render_path.suffix.lower() not in {".png", ".webp"}:
-                continue
-            add.append((
-                str(render_path),
-                f"{RENDERS_DIR_IN_REPO}/{submission_id}/"
-                f"{fixture_dir.name}/{render_path.name}",
-            ))
+        if renders_dir.is_dir():
+            for render_path in sorted(renders_dir.iterdir()):
+                if render_path.suffix.lower() not in {".png", ".webp"}:
+                    continue
+                add.append((str(render_path), f"{dest_prefix}/{render_path.name}"))
+        # The interface overlay is a per-fixture report artifact that lives at
+        # the fixture-dir root (not under renders/). It must ride to the bucket
+        # alongside the turntables so the hosted report can reference it by URL
+        # instead of base64-inlining it; without this the report would have a
+        # broken overlay link. Uploaded under the same per-fixture prefix.
+        overlay = fixture_dir / "interface_overlay.png"
+        if overlay.is_file():
+            add.append((str(overlay), f"{dest_prefix}/{overlay.name}"))
     if not add:
         return []
     HfApi(token=token).batch_bucket_files(RENDER_BUCKET, add=add, token=token)
@@ -435,7 +449,14 @@ def _run_report(
         str(run_dir), "-o", str(html_out),
     ]
     if render_base_url:
-        cmd += ["--render-base-url", render_base_url]
+        cmd += [
+            "--render-base-url", render_base_url,
+            # GT + inputs are private, so they link through the Space proxy
+            # rather than the public bucket. Passed alongside the render base
+            # so the whole hosted report is lazy-loaded links, not base64.
+            "--gt-base-url", GT_PROXY_BASE_URL,
+            "--input-base-url", INPUT_PROXY_BASE_URL,
+        ]
     print(f"[eval_job] {' '.join(cmd)}", flush=True)
     proc = subprocess.run(
         cmd,