stage eval shards through mounted bucket

eval_job.py

@@ -40,13 +40,15 @@ the Space's sharded submit path (UC3) to fan a large submission across
 several jobs. Steps 1-2 are identical, then the run dir is pruned to
 just this shard's fixtures, ``cadgenbench evaluate`` runs over that
 subset, and the resulting per-fixture dirs (``result.json`` + renders)
-are uploaded *verbatim* under ``reports/<id>/shards/<shard_id>/``. No
+are staged *verbatim*. If ``CADGENBENCH_SHARD_BUCKET_MOUNT`` is set, the
+shard copies them into that mounted bucket; otherwise it uploads under
+``reports/<id>/shards/<shard_id>/`` in the submissions dataset. No
 report HTML, ``report.json``, or gallery render is produced per shard:
-the Space downloads every shard's fixture dirs, merges them into one
-run dir, and builds the single ``run_summary`` + report + gallery from
-the merged whole (mirroring the orchestrator's ``_merge_eval``). Exit 0
-on success; any failure exits non-zero and the Space marks that shard
-ERROR and retries it.
+the Space reads every shard's fixture dirs, merges them into one run dir,
+and builds the single ``run_summary`` + report + gallery from the merged
+whole (mirroring the orchestrator's ``_merge_eval``). Exit 0 on success;
+any failure exits non-zero and the Space marks that shard ERROR and
+retries it.
 """
 from __future__ import annotations
 
@@ -71,6 +73,8 @@ REPORT_TIMEOUT_SECONDS = 5 * 60
 
 REPORTS_DIR_IN_REPO = "reports"
 RENDERS_DIR_IN_REPO = "renders"
+SHARD_BUCKET_MOUNT_ENV = "CADGENBENCH_SHARD_BUCKET_MOUNT"
+SHARD_BUCKET_PREFIX_ENV = "CADGENBENCH_SHARD_BUCKET_PREFIX"
 
 # Sub-prefix under ``reports/<id>/`` where each shard uploads its raw
 # per-fixture dirs in shard mode. The Space merges these and deletes the
@@ -280,15 +284,35 @@ def _upload_shard_artifacts(
 ) -> None:
     """Upload this shard's evaluated per-fixture dirs for the Space to merge.
 
-    Pushes the pruned ``run_dir`` (each ``<fixture>/`` with its
-    ``result.json`` + ``renders/`` + any overlay PNGs) verbatim to
-    ``reports/<id>/shards/<shard_id>/`` in one commit. The Space
-    downloads every shard's tree, copies the fixture dirs into a single
-    merged run dir, and builds the aggregate ``run_summary`` + report +
-    gallery from the whole. The per-shard ``run_summary.json`` written
-    by ``cadgenbench evaluate`` rides along harmlessly; the merge
-    recomputes it over the union and ignores the partials.
+    Persists the pruned ``run_dir`` (each ``<fixture>/`` with its
+    ``result.json`` + ``renders/`` + any overlay PNGs) verbatim. In
+    bucket mode, this is a filesystem copy into the mounted bucket; in
+    legacy mode, it is one dataset-repo commit under
+    ``reports/<id>/shards/<shard_id>/``. The Space reads every shard's
+    tree, copies the fixture dirs into a single merged run dir, and
+    builds the aggregate ``run_summary`` + report + gallery from the
+    whole. The per-shard ``run_summary.json`` written by
+    ``cadgenbench evaluate`` rides along harmlessly; the merge recomputes
+    it over the union and ignores the partials.
     """
+    bucket_mount = os.environ.get(SHARD_BUCKET_MOUNT_ENV)
+    if bucket_mount:
+        prefix = os.environ.get(SHARD_BUCKET_PREFIX_ENV, "submissions").strip("/")
+        dest = Path(bucket_mount) / prefix / submission_id / SHARDS_DIR_NAME / shard_id
+        if not Path(bucket_mount).is_dir():
+            raise RuntimeError(
+                f"{SHARD_BUCKET_MOUNT_ENV}={bucket_mount!r} is not a mounted directory."
+            )
+        if dest.exists():
+            shutil.rmtree(dest)
+        dest.parent.mkdir(parents=True, exist_ok=True)
+        shutil.copytree(run_dir, dest)
+        print(
+            f"[eval_job] staged shard {shard_id} -> {dest}",
+            flush=True,
+        )
+        return
+
     api = HfApi(token=token)
     path_in_repo = f"{REPORTS_DIR_IN_REPO}/{submission_id}/{SHARDS_DIR_NAME}/{shard_id}"
     api.upload_folder(