perf(compile): 批次加权分桶消除长尾 + 共享符号表去重复序列化 + typedef 收集走 AST 缓存

- Phase2 批次切分：原按文件数均分（batch_size=ceil(N/n)），大文件集中批次拖长尾、worker 空等；改为按源文件大小 LPT 贪心分桶，各批次总成本均匀（任务颗粒度不均匀）
- 共享符号表序列化：shared_data 同时含 symbol_table 对象与 symbol_table_bytes 双重序列化（_shared_sym.json 4.65MB 近半冗余）；统一为仅传 symbol_table_bytes，worker/批次端从字节反序列化（includes 与 App 两阶段同步修改）
- Phase1 _collect_typedef_map / _process_file_pyi：直读+ast.parse 改为 parse_python_file mtime AST 缓存，Phase1 内重复收集与 Phase2 主进程解析共享同一缓存

lib/Projectrans/Phase1Generator.py

@@ -357,9 +357,9 @@ class Phase1Generator:
                 continue
             pyi_path: str = os.path.join(self.temp_dir, fname)
             try:
-                with open(pyi_path, 'r', encoding='utf-8') as f:
-                    content: str = f.read()
-                tree: ast.Module = ast.parse(content)
+                content, tree = parse_python_file(pyi_path)
+                if tree is None:
+                    continue
                 for node in ast.iter_child_nodes(tree):
                     if isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name):
                         var_name: str = node.target.id
@@ -676,9 +676,9 @@ class Phase1Generator:
                 continue
             pyi_path: str = os.path.join(self.temp_dir, fname)
             try:
-                with open(pyi_path, 'r', encoding='utf-8') as f:
-                    content: str = f.read()
-                tree: ast.Module = ast.parse(content)
+                content, tree = parse_python_file(pyi_path)
+                if tree is None:
+                    continue
                 for node in ast.iter_child_nodes(tree):
                     if isinstance(node, ast.ClassDef):
                         is_enum: bool = False

lib/Projectrans/Phase2Translator.py

@@ -96,7 +96,9 @@ def _parallel_translate_worker(src_path: str, out_path: str, src_root: str, temp
             sys.setrecursionlimit(max(sys.getrecursionlimit(), 100000))
             with open(shared_sym_pickle_path, 'rb') as f:
                 shared_data = pickle.load(f)
-            trans._shared_symbol_table = shared_data['symbol_table']
+            _sym_bytes: bytes | None = shared_data.get('symbol_table_bytes')
+            if _sym_bytes is not None:
+                trans._shared_symbol_table = pickle.loads(_sym_bytes)
             trans._shared_source_module_sig_files = shared_data['source_module_sig_files']
             trans._shared_all_dc = shared_data['all_dc']
             trans._shared_export_extern_funcs = shared_data['export_extern_funcs']
@@ -150,7 +152,8 @@ def _parallel_translate_batch(batch: list, src_root: str, temp_dir: str, output_
             with open(shared_sym_pickle_path, 'rb') as f:
                 shared_data = pickle.load(f)
             symbol_table_bytes = shared_data.get('symbol_table_bytes')
-            trans._shared_symbol_table = shared_data['symbol_table']
+            if symbol_table_bytes is not None:
+                trans._shared_symbol_table = pickle.loads(symbol_table_bytes)
             trans._shared_source_module_sig_files = shared_data['source_module_sig_files']
             trans._shared_all_dc = shared_data['all_dc']
             trans._shared_export_extern_funcs = shared_data['export_extern_funcs']
@@ -687,7 +690,6 @@ class Phase2Translator:
                     # 写入的符号污染后文件 → 跨模块 struct 类型前缀错位）。
                     symbol_table_bytes: bytes = pickle.dumps(self._shared_symbol_table)
                     shared_data = {
-                        'symbol_table': self._shared_symbol_table,
                         'symbol_table_bytes': symbol_table_bytes,
                         'source_module_sig_files': self._shared_source_module_sig_files,
                         'all_dc': self._shared_all_dc,
@@ -726,9 +728,26 @@ class Phase2Translator:
                     # 相比"一文件一任务"：减少进程池调度/通信/构造开销
                     # （200+ 文件 → n_workers 个批次任务）。
                     batches: list[list] = []
-                    batch_size: int = max(1, (len(need_translate) + n_workers - 1) // n_workers)
-                    for bi in range(0, len(need_translate), batch_size):
-                        batches.append(need_translate[bi:bi + batch_size])
+                    # 按源文件大小加权分桶（LPT 贪心：最大文件优先放入当前总成本最小的桶），
+                    # 使各批次总成本均匀。原按文件数均分（batch_size=ceil(N/n)）导致大文件
+                    # 集中的批次拖长尾，其他 worker 提前完成空等（任务颗粒度不均匀）。
+                    if len(need_translate) <= n_workers:
+                        batches = [[item] for item in need_translate]
+                    else:
+                        sized_items: list = []
+                        for item in need_translate:
+                            try:
+                                _size: int = os.path.getsize(item[1])
+                            except OSError:
+                                _size = 0
+                            sized_items.append((item, _size))
+                        buckets: list[list] = [[] for _ in range(n_workers)]
+                        bucket_cost: list[int] = [0] * n_workers
+                        for item, _size in sorted(sized_items, key=lambda t: t[1], reverse=True):
+                            target: int = min(range(n_workers), key=lambda w: bucket_cost[w])
+                            buckets[target].append(item)
+                            bucket_cost[target] += _size
+                        batches = [b for b in buckets if b]
                     futures = {}
                     for bidx, batch in enumerate(batches):
                         rel_first = os.path.relpath(batch[0][1], self.src_root)
@@ -2421,8 +2440,10 @@ class Phase2Translator:
             # 序列化共享符号表供 worker 使用
             shared_pickle_path = os.path.join(self.output_dir, '_shared_sym_includes.pickle')
             try:
+                # symbol_table 只以 pickle 字节传递（symbol_table_bytes），不重复序列化
+                # 对象本身（原双重序列化使文件体积/时间翻倍，4.65MB 中近半冗余）。
                 shared_data = {
-                    'symbol_table': self._shared_symbol_table,
+                    'symbol_table_bytes': pickle.dumps(self._shared_symbol_table),
                     'source_module_sig_files': self._shared_source_module_sig_files,
                     'all_dc': self._shared_all_dc,
                     'export_extern_funcs': self._shared_export_extern_funcs,