fix: build only recipes whose members match

scripts/romset_recipes.py

@@ -165,15 +165,38 @@ class AtomPool:
 def identify_archive(
     path: Path, recipes: dict[str, Recipe], pool: AtomPool
 ) -> str | None:
-    """Return the recipe label that reproduces *path* byte for byte."""
+    """Return the recipe label that reproduces *path* byte for byte.
+
+    Equal bytes need equal (name, crc32) members, which the archive's
+    central directory gives without inflating anything: building every
+    recipe at DEFLATE level 9 first took 94 s and 1 GB at pipeline step 1c.
+    """
+    held = _member_list(path)
+    if held is None:
+        return None
     original = path.read_bytes()
     for label, recipe in sorted(recipes.items()):
+        if sorted((name, crc.lower()) for name, crc in recipe) != held:
+            continue
         blob = pool.build(recipe)
         if blob is not None and blob == original:
             return label
     return None
 
 
+def _member_list(path: Path) -> list[tuple[str, str]] | None:
+    """Sorted (name, crc32) of an archive's members, from its directory."""
+    try:
+        with zipfile.ZipFile(path) as archive:
+            return sorted(
+                (info.filename, format(info.CRC, "08x"))
+                for info in archive.infolist() if not info.is_dir()
+            )
+    except (zipfile.BadZipFile, OSError) as exc:
+        print(f"  UNREADABLE: {path}: {exc}", file=sys.stderr)
+        return None
+
+
 def reconstruct(
     target_md5: str, recipes: dict[str, Recipe], pool: AtomPool
 ) -> tuple[str, bytes] | None: