mix_fingerprint: fold the manifest's per-shard sha256 into the corpus identity (names+sizes could collide across rebuilds); phase4_session repins shard_dataset

train/shard_dataset.py

@@ -151,12 +151,20 @@ def save_cursor(dirpath, cur):
 
 
 def mix_fingerprint(root, manifest=None):
-    """Cheap, stable identity for the exact data being trained on: sha of the manifest's shard list
-    plus each shard's size. Used by the trainer to refuse a resume onto a different corpus."""
+    """Cheap, stable identity for the exact data being trained on. Used by the trainer to refuse a resume
+    onto a different corpus, so it has to be computed the same way in every session.
+
+    Names alone (or names and sizes, as this function did until the E-031 review) are too weak: a rebuilt
+    mix whose shards happen to land on the same sizes is a different corpus, and the whole point of the
+    check is to catch a corpus swap between sessions. The manifest already certifies a sha256 per shard,
+    so folding it in costs nothing and needs no file reads; the local size is included as well so a
+    truncated file on disk cannot hide behind a manifest that still describes the good one.
+    """
     import hashlib
     man = manifest or json.load(open(os.path.join(root, "manifest.json")))
     h = hashlib.sha256()
     for s in man["shards"]:
         p = os.path.join(root, s["file"])
-        h.update(f"{s['file']}:{os.path.getsize(p) if os.path.exists(p) else 'missing'}\n".encode())
+        h.update(f"{s['file']}:{s.get('sha256')}:"
+                 f"{os.path.getsize(p) if os.path.exists(p) else 'missing'}\n".encode())
     return h.hexdigest()[:16]