download_missing must fetch record.json too (.bin-only filter broke restore); rehearsal pre-cleans stale test repo

build/hubsync.py

@@ -118,7 +118,11 @@ def download_missing(root, keys_wanted, repo, remote=None):
     remote = remote if remote is not None else remote_stage_files(repo)
     got = {}
     for key in keys_wanted:
-        want = sorted(p for p in remote if p.startswith(f"stage/{key}/") and p.endswith(".bin"))
+        # .json must come down too: record.json is what tells the restore the shard list, token count
+        # and drop stats. A .bin-only filter made every restored source look unknown and was the actual
+        # cause of "restored 0 staged source(s)" in rehearsal v3, with the files plainly present.
+        want = sorted(p for p in remote
+                      if p.startswith(f"stage/{key}/") and p.endswith((".bin", ".json")))
         if not want:
             got[key] = []
             continue

build/smoke_hub_resume.py

@@ -54,6 +54,20 @@ def main():
     base = ["build_mix.py", "--root", ROOT, "--only", TWO, "--cap-tokens-per-source", "1200000",
             "--hub-repo", REPO]
 
+    # Start from an empty stage repo. v3 of this rehearsal failed and, by design, KEPT the repo as
+    # evidence; without this the next run would restore those sources and skip the staging step it is
+    # meant to be testing.
+    def preclean():
+        import ounce100m_credentials
+        ounce100m_credentials.install()
+        from huggingface_hub import delete_repo
+        try:
+            delete_repo(repo_id=REPO, repo_type="dataset", token=os.environ["HF_TOKEN"])
+            return "deleted stale test repo"
+        except Exception as e:
+            return f"nothing to delete ({type(e).__name__})"
+
+    R["preclean"] = preclean()
     shutil.rmtree(ROOT, ignore_errors=True)
     r1 = sh(base + ["--stage-only"], "1. stage two sources and publish")
     R["stage"] = {"rc": r1["rc"], "json": grab_json(r1["out"], "BUILD_JSON_BEGIN", "BUILD_JSON_END")}