fix(phase2): 并行编译 worker 数按可用内存动态限制，修复内存不足 OOM

系统可用内存低（Trae CN/Java 占用后仅剩 ~1GB）时，ProcessPoolExecutor
按 cpu_count 开满 worker，多进程编译内存叠加在 StructGen.finalize 拼接
module IR 时 MemoryError（编译 StubMergerTopo.vp 报错即此）。新增
_PickWorkerCount：GlobalMemoryStatusEx 读 ullAvailPhys，按 per_worker_mb
估算 worker 上限（翻译 700MB / llc 400MB），内存不足自动降为单 worker
串行，内存充足回落 cpu_count。StructGen.finalize 拼接改中间变量拆分，
降低单 worker 峰值副本。

lib/Projectrans/Phase2Translator.py

@@ -23,6 +23,43 @@ import TransPyC
 from StubGen import PythonToStubConverter
 
 
+def _PickWorkerCount(n_files: int, per_worker_mb: int = 700) -> int:
+    """按可用物理内存动态限制并行 worker 数，避免多进程编译内存峰值 OOM。
+
+    单 worker 预估 per_worker_mb（Python 解释器 + llvmlite module + 共享
+    符号表快照）；系统可用内存不足时降为单 worker 串行，内存充足时回落到
+    cpu_count 并行。Windows 用 GlobalMemoryStatusEx 读 ullAvailPhys。
+    """
+    base = min(os.cpu_count() or 4, n_files)
+    if base <= 1:
+        return base
+    try:
+        import ctypes
+
+        class _MemStat(ctypes.Structure):
+            _fields_ = [
+                ("dwLength", ctypes.c_ulong),
+                ("dwMemoryLoad", ctypes.c_ulong),
+                ("ullTotalPhys", ctypes.c_ulonglong),
+                ("ullAvailPhys", ctypes.c_ulonglong),
+                ("ullTotalPageFile", ctypes.c_ulonglong),
+                ("ullAvailPageFile", ctypes.c_ulonglong),
+                ("ullTotalVirtual", ctypes.c_ulonglong),
+                ("ullAvailVirtual", ctypes.c_ulonglong),
+                ("ullAvailExtendedVirtual", ctypes.c_ulonglong),
+            ]
+
+        stat = _MemStat()
+        stat.dwLength = ctypes.sizeof(_MemStat)
+        if ctypes.windll.kernel32.GlobalMemoryStatusEx(ctypes.byref(stat)):
+            avail_mb = stat.ullAvailPhys // (1024 * 1024)
+            by_mem = max(1, avail_mb // per_worker_mb)
+            return max(1, min(base, by_mem))
+    except Exception:
+        pass
+    return base
+
+
 def _ExtractCDefineConstantsFromPyi(pyi_path: str, all_dc: dict[str, object]) -> None:
     """从 .pyi 文件中提取 CDefine 常量到 all_dc。文件不存在时静默跳过。
 
@@ -737,7 +774,7 @@ class Phase2Translator:
             _vlog().info(f"\n需要翻译 {len(need_translate)} 个文件")
             t_start = time.time()
 
-            n_workers = min(os.cpu_count() or 4, len(need_translate))
+            n_workers = _PickWorkerCount(len(need_translate))
 
             if n_workers > 1 and len(need_translate) > 1:
                 _vlog().info(f"  并行翻译 (workers={n_workers})")
@@ -2019,7 +2056,7 @@ class Phase2Translator:
                 need_compile.append(ll_path)
 
         if need_compile:
-            n_workers = min(os.cpu_count() or 4, len(need_compile))
+            n_workers = _PickWorkerCount(len(need_compile), 400)
             _vlog().info(f"  并行编译 {len(need_compile)} 个文件 (workers={n_workers})")
             errors = []
             with ThreadPoolExecutor(max_workers=n_workers) as executor:
@@ -2545,7 +2582,7 @@ class Phase2Translator:
                 shared_pickle_path = ''
 
             if shared_pickle_path and os.path.exists(shared_pickle_path):
-                n_workers = min(os.cpu_count() or 4, len(need_translate))
+                n_workers = _PickWorkerCount(len(need_translate))
                 errors = []
                 # 波次并行翻译：按依赖关系分批提交，确保被依赖的文件先翻译完成。
                 # 根因：_TryLoadStructFromStub 扫描 output 目录加载跨模块 struct 定义，
@@ -2784,7 +2821,7 @@ class Phase2Translator:
             return (task, result)
 
         if len(compile_tasks) > 1:
-            n_workers = min(os.cpu_count() or 4, len(compile_tasks))
+            n_workers = _PickWorkerCount(len(compile_tasks), 400)
             with ThreadPoolExecutor(max_workers=n_workers) as executor:
                 futures = {executor.submit(_compile_one_include, task): task for task in compile_tasks}
                 done_count = 0

lib/core/LLVMCG/StructGen.py

@@ -201,7 +201,9 @@ class StructGenMixin:
             if insert_pos == -1:
                 insert_pos = header_end
             struct_block: str = '\n'.join(struct_defs) + '\n'
-            ir_text = ir_text[:insert_pos + 1] + struct_block + ir_text[insert_pos + 1:]
+            _ir_head = ir_text[:insert_pos + 1]
+            _ir_tail = ir_text[insert_pos + 1:]
+            ir_text = _ir_head + struct_block + _ir_tail
         # 仅含特殊字符（数字开头或含点）的 struct 名需要加引号：
         # 合并为单个正则一次扫描全文本，替代逐 name 各自 re.sub（每个 name
         # 一次 O(text) 扫描）。lookahead 边界排除保证各备选互斥，语义等价。