diff --git a/scripts/common.py b/scripts/common.py
index c6a287cb..9db5d4f3 100644
--- a/scripts/common.py
+++ b/scripts/common.py
@@ -10,6 +10,7 @@ import contextlib
 import hashlib
 import json
 import os
+import tempfile
 import urllib.error
 import urllib.parse
 import urllib.request
@@ -1220,7 +1221,12 @@ def fetch_large_file(
             return cached
 
     os.makedirs(dest_dir, exist_ok=True)
-    tmp_path = cached + ".tmp"
+    # A per-process scratch name: two runs fetching the same asset into one
+    # shared path interleave their writes into a full-size, corrupt file.
+    tmp_fd, tmp_path = tempfile.mkstemp(
+        dir=dest_dir, prefix=os.path.basename(cached) + ".", suffix=".tmp"
+    )
+    os.close(tmp_fd)
     # GitHub rewrites spaces to dots in release asset names, so a file whose
     # name contains spaces is published under a dotted name.
     candidates = [name]
@@ -1250,6 +1256,8 @@ def fetch_large_file(
                 os.unlink(tmp_path)
 
     if not downloaded:
+        if os.path.exists(tmp_path):
+            os.unlink(tmp_path)
         return None
 
     if expected_sha1 or expected_md5:
diff --git a/tests/test_large_file_cache.py b/tests/test_large_file_cache.py
new file mode 100644
index 00000000..b47e2182
--- /dev/null
+++ b/tests/test_large_file_cache.py
@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+"""Large-file cache downloads.
+
+Two runs fetching the same asset used to stream into one shared scratch
+path, interleaving their writes into a full-size file with mixed content
+that then replaced the cache entry.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import io
+import os
+import sys
+import tempfile
+import threading
+import unittest
+import urllib.error
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(REPO_ROOT / "scripts"))
+
+import common  # noqa: E402
+
+PAYLOAD_A = b"A" * (256 * 1024)
+PAYLOAD_B = b"B" * (256 * 1024)
+
+
+class _SlowResponse(io.BytesIO):
+    """Serve a payload in chunks, yielding between them."""
+
+    def __init__(self, data: bytes, barrier: threading.Barrier | None = None):
+        super().__init__(data)
+        self._barrier = barrier
+        self._first = True
+
+    def read(self, size: int = -1) -> bytes:
+        chunk = super().read(min(size, 4096) if size and size > 0 else 4096)
+        if self._first and self._barrier is not None:
+            self._first = False
+            self._barrier.wait(timeout=10)
+        return chunk
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        self.close()
+        return False
+
+
+class LargeFileCacheTest(unittest.TestCase):
+    def setUp(self):
+        self.dir = tempfile.mkdtemp()
+        self._urlopen = common.urllib.request.urlopen
+
+    def tearDown(self):
+        common.urllib.request.urlopen = self._urlopen
+
+    def test_concurrent_fetches_do_not_mix(self):
+        barrier = threading.Barrier(2, timeout=10)
+        payloads = [PAYLOAD_A, PAYLOAD_B]
+        index = iter(range(2))
+        lock = threading.Lock()
+
+        def fake_urlopen(req, timeout=None):
+            with lock:
+                i = next(index)
+            return _SlowResponse(payloads[i], barrier)
+
+        common.urllib.request.urlopen = fake_urlopen
+        results: list[str | None] = [None, None]
+
+        def worker(slot: int):
+            results[slot] = common.fetch_large_file("asset.bin", dest_dir=self.dir)
+
+        threads = [threading.Thread(target=worker, args=(i,)) for i in range(2)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join(timeout=30)
+
+        cached = Path(self.dir) / "asset.bin"
+        self.assertTrue(cached.exists())
+        digest = hashlib.sha1(cached.read_bytes()).hexdigest()
+        accepted = {hashlib.sha1(p).hexdigest() for p in payloads}
+        # Whichever run lands last wins, but never a blend of the two
+        self.assertIn(digest, accepted)
+
+    def test_no_scratch_file_survives_a_successful_fetch(self):
+        common.urllib.request.urlopen = lambda req, timeout=None: _SlowResponse(
+            PAYLOAD_A
+        )
+        common.fetch_large_file("asset.bin", dest_dir=self.dir)
+        leftovers = [f for f in os.listdir(self.dir) if f.endswith(".tmp")]
+        self.assertEqual(leftovers, [])
+
+    def test_no_scratch_file_survives_a_failed_fetch(self):
+        def fail(req, timeout=None):
+            raise urllib.error.URLError("offline")
+
+        common.urllib.request.urlopen = fail
+        self.assertIsNone(common.fetch_large_file("asset.bin", dest_dir=self.dir))
+        self.assertEqual(os.listdir(self.dir), [])
+
+    def test_hash_mismatch_leaves_no_scratch_file(self):
+        common.urllib.request.urlopen = lambda req, timeout=None: _SlowResponse(
+            PAYLOAD_A
+        )
+        result = common.fetch_large_file(
+            "asset.bin", dest_dir=self.dir, expected_sha1="00" * 20
+        )
+        self.assertIsNone(result)
+        self.assertEqual(os.listdir(self.dir), [])
+
+    def test_cached_file_is_returned_without_download(self):
+        cached = Path(self.dir) / "asset.bin"
+        cached.write_bytes(PAYLOAD_A)
+
+        def fail(req, timeout=None):
+            raise AssertionError("must not download when the cache is valid")
+
+        common.urllib.request.urlopen = fail
+        self.assertEqual(
+            common.fetch_large_file("asset.bin", dest_dir=self.dir), str(cached)
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()