PyREUser3 0.7.0__tar.gz → 0.7.1__tar.gz

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyREUser3
-Version: 0.7.0
+Version: 0.7.1
 Summary: Pure Python tools for converting RE Engine .user.3 files to and from JSON.
 Author: Egg Targaryen
 License-Expression: MIT
@@ -47,6 +47,7 @@ pip install pyreuser3
 - `JSON -> .user.3` packing.
 - A reusable Python API through `REUser3Converter`.
 - CLI commands through `pyreuser3`.
+- Schema-free layout probing through `pyreuser3 probe` and the public probe API.
 - A local `.user.3` export Web UI through `pyreuser3-web`.
 - Automatic separation of the USR outer layout from the embedded RSZ header family;
   modern RSZ v4+ files preserve their original numeric version during repack.
@@ -59,6 +60,20 @@ Verified H30/modern layouts support repacking. Experimental physical H28 and leg
 RSZ v3 layouts are readable for analysis but intentionally blocked from repacking
 until real fixtures provide byte-for-byte validation.
 
+Version 0.7.1 fixes the 0.7.0 modern RSZ alignment regression. Modern userdata and
+data targets use absolute 16-byte file alignment even though their stored offsets
+remain RSZ-relative. Readable parsing accepts alignment-only deviations without
+hiding structural corruption, while the probe API reports them explicitly. Repack
+output blocks layouts carrying unverified diagnostics. The strict schema-free
+probe can validate large local corpora.
+
+Repeated operations on one `REUser3Converter` reuse schema and il2cpp metadata,
+with automatic file-signature invalidation and an explicit
+`clear_metadata_cache()` escape hatch. Batch patching reuses one exporter and
+packer. Signed `-1` null object references and very large fixed-width arrays are
+preserved losslessly; large arrays use a compact raw payload instead of millions
+of Python scalar objects.
+
 ## Requirements
 
 - Python 3.9 or newer.
@@ -66,6 +81,8 @@ until real fixtures provide byte-for-byte validation.
 - An `il2cpp_dump.json` file when exporting readable enum labels.
 - One or more unpacked `.user.3` files.
 
+`pyreuser3 probe` does not require a schema or `il2cpp_dump.json`.
+
 ## Usage
 
 Usage details may change as the package evolves. For the latest command-line and Python API examples, read the GitHub README:

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/PyREUser3.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyREUser3
-Version: 0.7.0
+Version: 0.7.1
 Summary: Pure Python tools for converting RE Engine .user.3 files to and from JSON.
 Author: Egg Targaryen
 License-Expression: MIT
@@ -47,6 +47,7 @@ pip install pyreuser3
 - `JSON -> .user.3` packing.
 - A reusable Python API through `REUser3Converter`.
 - CLI commands through `pyreuser3`.
+- Schema-free layout probing through `pyreuser3 probe` and the public probe API.
 - A local `.user.3` export Web UI through `pyreuser3-web`.
 - Automatic separation of the USR outer layout from the embedded RSZ header family;
   modern RSZ v4+ files preserve their original numeric version during repack.
@@ -59,6 +60,20 @@ Verified H30/modern layouts support repacking. Experimental physical H28 and leg
 RSZ v3 layouts are readable for analysis but intentionally blocked from repacking
 until real fixtures provide byte-for-byte validation.
 
+Version 0.7.1 fixes the 0.7.0 modern RSZ alignment regression. Modern userdata and
+data targets use absolute 16-byte file alignment even though their stored offsets
+remain RSZ-relative. Readable parsing accepts alignment-only deviations without
+hiding structural corruption, while the probe API reports them explicitly. Repack
+output blocks layouts carrying unverified diagnostics. The strict schema-free
+probe can validate large local corpora.
+
+Repeated operations on one `REUser3Converter` reuse schema and il2cpp metadata,
+with automatic file-signature invalidation and an explicit
+`clear_metadata_cache()` escape hatch. Batch patching reuses one exporter and
+packer. Signed `-1` null object references and very large fixed-width arrays are
+preserved losslessly; large arrays use a compact raw payload instead of millions
+of Python scalar objects.
+
 ## Requirements
 
 - Python 3.9 or newer.
@@ -66,6 +81,8 @@ until real fixtures provide byte-for-byte validation.
 - An `il2cpp_dump.json` file when exporting readable enum labels.
 - One or more unpacked `.user.3` files.
 
+`pyreuser3 probe` does not require a schema or `il2cpp_dump.json`.
+
 ## Usage
 
 Usage details may change as the package evolves. For the latest command-line and Python API examples, read the GitHub README:

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/PyREUser3.egg-info/SOURCES.txt

@@ -42,6 +42,7 @@ pyreuser3/web/picker.py
 pyreuser3/web/runners.py
 pyreuser3/web/server.py
 pyreuser3/web/settings.py
+tests/test_api_cache.py
 tests/test_enum_binary.py
 tests/test_enum_codec.py
 tests/test_export_enums.py

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/README.md

@@ -31,6 +31,7 @@ from pyreuser3 import REUser3Converter
 - `JSON -> .user.3` packing.
 - A reusable Python API through `REUser3Converter`.
 - CLI commands through `pyreuser3`.
+- Schema-free layout probing for individual files and whole corpora.
 - A local `.user.3` export Web UI through `pyreuser3-web`.
 
 This PyPI package intentionally does not include game resources, dumped game data, RE_RSZ templates, `il2cpp_dump.json`,
@@ -43,6 +44,9 @@ or repository-specific helper scripts.
 - An `il2cpp_dump.json` file when exporting readable enum labels.
 - One or more unpacked `.user.3` files.
 
+Layout probing is the exception: `pyreuser3 probe` only inspects the USR/RSZ
+container and does not require a schema or `il2cpp_dump.json`.
+
 ## Command Line
 
 Export `.user.3` files to JSON:
@@ -76,6 +80,19 @@ pyreuser3 pack \
 
 The `-p/--il2cpp-dump-path` option is required for export and optional for pack. Passing it during pack is recommended when enum names need to be resolved back to numeric values.
 
+Probe a file or directory without loading game metadata:
+
+```bash
+pyreuser3 probe -i <input-user3-file-or-directory>
+pyreuser3 probe -i <input-user3-file-or-directory> --strict -o layout-report.json
+```
+
+The default probe uses the same safety policy as readable export: structural
+corruption remains fatal, while non-canonical alignment is reported as a warning.
+`--strict` rejects every layout deviation and is intended for corpus validation.
+The optional JSON report contains totals grouped by detected layout, RSZ version,
+and diagnostic code.
+
 Start the local `.user.3` export Web UI:
 
 ```bash
@@ -111,6 +128,29 @@ converter.pack_file(
 )
 ```
 
+`REUser3Converter` loads schema and il2cpp metadata lazily and caches it for the
+lifetime of that converter. Repeated readable, repack, pack, and patch calls no
+longer rescan large metadata files. Cache entries are invalidated automatically
+when the source path, size, or nanosecond modification time changes; callers that
+replace metadata in an unusual way can force a reload with:
+
+```python
+converter.clear_metadata_cache()
+```
+
+`patch_directory()` also reuses one prepared exporter and packer for the complete
+batch instead of constructing them once per file.
+
+Container layout probing is also available without constructing a converter or
+loading a schema:
+
+```python
+from pyreuser3 import probe_usr_file, probe_usr_path
+
+one_file = probe_usr_file("input/example.user.3")
+whole_tree = probe_usr_path("input/natives", policy="strict_probe")
+```
+
 Convert a `.user.3` file to an in-memory JSON-compatible Python object without writing a JSON file:
 
 ```python
@@ -139,10 +179,22 @@ dependency tables. Layout candidates and their read/repack capability status are
 declared in `pyreuser3/usr_layouts.py`; RSZ field definitions still come from the
 supplied REFramework-compatible schema. The verified modern header family accepts
 structurally valid RSZ v4+ files and preserves their original version instead of
-forcing MHWS version 16. Experimental physical H28 and legacy RSZ v3 candidates are
-read-only until real fixtures validate byte-for-byte repacking. V1 and v2 documents
-are recognized for diagnostics but must be re-exported as v3 before packing because
-they do not record the required layout metadata.
+forcing MHWS version 16. For this family, offset values remain relative to the RSZ
+section, but the userdata and data targets are aligned to absolute 16-byte file
+positions. Readable export accepts alignment-only deviations, and the probe API
+reports them as structured warnings. Repack export records the same diagnostics in
+`_warnings` and blocks packing through `_unsupported` until the layout is verified.
+Experimental physical H28 and legacy RSZ v3 candidates are read-only until real
+fixtures validate byte-for-byte repacking. V1 and v2 documents are recognized for
+diagnostics but must be re-exported as v3 before packing because they do not record
+the required layout metadata.
+
+Modern files may use signed `-1` as an explicit null `Object` reference; repack
+preserves that sentinel while continuing to reject every other missing instance
+ID. Fixed-width arrays above one million elements are kept compactly as
+`_raw_array_count` plus `_raw_array_hex`. This representation is intentionally
+opaque but lossless, avoids expanding millions of Python integers, and validates
+the count/payload length before packing.
 
 For stable patch-and-repack workflows, use `patch_file()` or `parse_pack_file()`:
 
@@ -165,6 +217,33 @@ converter.patch_file(
 )
 ```
 
+## Compatibility Validation
+
+The repository includes a schema-free corpus runner. Game files are not committed;
+point it at a locally unpacked `natives` tree:
+
+```bash
+python tests/corpus_probe.py D:/game/natives \
+  --expected-total 62768 \
+  --expected-version 16 \
+  --report layout-report.json
+```
+
+Version 0.7.1 fixes modern RSZ alignment regression found in 0.7.0 and folds the
+planned compatibility work into the bug-fix release: staged layout detection,
+structured diagnostics, safe readable parsing, verified-only repacking, and the
+schema-free probe command. The modern rule was validated against 62,768 MHWS
+`.user.3` files; four SystemSetting fixtures additionally passed readable export,
+repack export, byte-identical packing, and reparse checks.
+
+A second corpus from Monster Hunter Stories 3 was also validated: all 42,945 files
+passed strict layout detection and schema-driven repack export. All 42,945 rebuilt
+without exceptions; 38,548 were byte-identical. The remaining 4,397 are successful
+rebuilds but not byte-identical; observed causes include nonzero padding and
+alternate empty-string encodings, so they are not claimed as byte-identical
+fixtures. Two voxel files use the compact large-array representation; both rebuilt
+byte-identically, including the 6.55 MB `dg100_Root` payload.
+
 ## Build From Source
 
 ```bash

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/docs/PYPI.md

@@ -14,6 +14,7 @@ pip install pyreuser3
 - `JSON -> .user.3` packing.
 - A reusable Python API through `REUser3Converter`.
 - CLI commands through `pyreuser3`.
+- Schema-free layout probing through `pyreuser3 probe` and the public probe API.
 - A local `.user.3` export Web UI through `pyreuser3-web`.
 - Automatic separation of the USR outer layout from the embedded RSZ header family;
   modern RSZ v4+ files preserve their original numeric version during repack.
@@ -26,6 +27,20 @@ Verified H30/modern layouts support repacking. Experimental physical H28 and leg
 RSZ v3 layouts are readable for analysis but intentionally blocked from repacking
 until real fixtures provide byte-for-byte validation.
 
+Version 0.7.1 fixes the 0.7.0 modern RSZ alignment regression. Modern userdata and
+data targets use absolute 16-byte file alignment even though their stored offsets
+remain RSZ-relative. Readable parsing accepts alignment-only deviations without
+hiding structural corruption, while the probe API reports them explicitly. Repack
+output blocks layouts carrying unverified diagnostics. The strict schema-free
+probe can validate large local corpora.
+
+Repeated operations on one `REUser3Converter` reuse schema and il2cpp metadata,
+with automatic file-signature invalidation and an explicit
+`clear_metadata_cache()` escape hatch. Batch patching reuses one exporter and
+packer. Signed `-1` null object references and very large fixed-width arrays are
+preserved losslessly; large arrays use a compact raw payload instead of millions
+of Python scalar objects.
+
 ## Requirements
 
 - Python 3.9 or newer.
@@ -33,6 +48,8 @@ until real fixtures provide byte-for-byte validation.
 - An `il2cpp_dump.json` file when exporting readable enum labels.
 - One or more unpacked `.user.3` files.
 
+`pyreuser3 probe` does not require a schema or `il2cpp_dump.json`.
+
 ## Usage
 
 Usage details may change as the package evolves. For the latest command-line and Python API examples, read the GitHub README:

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/docs/README.zh-CN.md

@@ -33,6 +33,7 @@ PyREUser3 当前提供：
 - `JSON -> .user.3` 封包；
 - 面向其他项目调用的 `REUser3Converter` Python API；
 - `pyreuser3` 命令行工具；
+- 无需 schema 的单文件与整批布局探测；
 - `pyreuser3-web` 本地 Web 导出界面。
 
 PyPI 包不会包含游戏资源、游戏 dump、RE_RSZ 模板、`il2cpp_dump.json`、或特定仓库脚本。调用方需要自己准备与目标游戏版本匹配的数据文件。
@@ -44,6 +45,9 @@ PyPI 包不会包含游戏资源、游戏 dump、RE_RSZ 模板、`il2cpp_dump.js
 - 导出可读枚举标签时需要 `il2cpp_dump.json`；
 - 一个或多个已解包的 `.user.3` 文件。
 
+布局探测是例外：`pyreuser3 probe` 只检查 USR/RSZ 容器，不需要 schema 或
+`il2cpp_dump.json`。
+
 ## 命令行使用
 
 导出 `.user.3` 为 JSON：
@@ -82,6 +86,17 @@ pyreuser3 pack \
 - `-s/--schema-path` 必须指向具体 schema JSON 文件，不应传目录；
 - `-i`、`-j` 都可以传单个文件或目录，目录会递归处理。
 
+无需加载游戏元数据即可探测单个文件或整个目录：
+
+```bash
+pyreuser3 probe -i <输入的-user3-文件或目录>
+pyreuser3 probe -i <输入的-user3-文件或目录> --strict -o layout-report.json
+```
+
+默认探测采用与 readable 导出相同的安全策略：结构损坏仍会直接失败，只有非规范对齐会记录为
+警告。`--strict` 会拒绝任何布局偏差，适合整批语料验证。可选的 JSON 报告会按布局、RSZ
+版本和诊断代码汇总结果。
+
 可用 `--user-magic` 和 `--rsz-magic` 覆盖默认 magic，例如：
 
 ```bash
@@ -139,6 +154,27 @@ converter.pack_file(
 )
 ```
 
+`REUser3Converter` 会延迟加载 schema 和 il2cpp metadata，并在该 converter 的生命周期内复用。
+连续执行 readable、repack、pack 或 patch 时，不再反复扫描大型 metadata 文件。源文件的绝对
+路径、大小或纳秒修改时间变化后缓存会自动失效；如果调用方以特殊方式原地替换文件，也可以强制
+清理：
+
+```python
+converter.clear_metadata_cache()
+```
+
+`patch_directory()` 现在也会让整批文件复用同一个已准备好的 exporter 和 packer，不再为每个
+文件重复构造。
+
+也可以不创建 converter、不加载 schema，直接调用容器布局探测 API：
+
+```python
+from pyreuser3 import probe_usr_file, probe_usr_path
+
+one_file = probe_usr_file("input/example.user.3")
+whole_tree = probe_usr_path("input/natives", policy="strict_probe")
+```
+
 不写入 JSON 文件，直接把 `.user.3` 转成内存中的 JSON 兼容 Python 对象：
 
 ```python
@@ -162,9 +198,17 @@ repack_data = converter.user3_to_json(
 USR 外层布局、RSZ 头族和文件中的真实 RSZ 版本，并保留 resource 与 userdata 依赖表。布局
 候选及其读取/回封能力状态集中声明在 `pyreuser3/usr_layouts.py`，RSZ 字段定义仍来自传入的
 REFramework 兼容模板。已验证的现代头族接受通过完整结构校验的 RSZ v4+ 文件，并原样保留
-版本号，不再固定为 MHWS 的版本 16。实验性的物理 H28 与 legacy RSZ v3 候选暂时只读，获得
-真实样本并完成逐字节回封验证后才能启用 repack。v1 和 v2 文档仍可识别以便诊断，但由于缺少
-必要的布局元数据，封回前必须从源文件重新导出为 v3。
+版本号，不再固定为 MHWS 的版本 16。这个头族中的 offset 数值仍以 RSZ 段为基准，但 userdata
+与 data 的目标位置按整个文件的绝对 16 字节边界对齐。readable 导出会接受仅对齐偏差，probe
+API 则把它们报告为结构化警告；repack 导出会把同一诊断写入 `_warnings`，并通过
+`_unsupported` 阻止未经验证的布局被封回。实验性的物理 H28 与 legacy RSZ v3 候选暂时只读，
+获得真实样本并完成逐字节回封验证后才能启用 repack。v1 和 v2 文档仍可识别以便诊断，但由于
+缺少必要的布局元数据，封回前必须从源文件重新导出为 v3。
+
+部分现代文件会用有符号 `-1` 明确表示空 `Object` 引用；repack 会原样保留这个哨兵，同时继续
+拒绝其他不存在的实例 ID。超过一百万项的定长数组会紧凑保存为 `_raw_array_count` 和
+`_raw_array_hex`。这种表示不提供逐项可读性，但能够无损回封、避免展开数百万个 Python 整数，
+并会在封包前校验 count 与 payload 长度是否匹配。
 
 批量处理目录：
 
@@ -212,6 +256,29 @@ converter.patch_file(
 )
 ```
 
+## 兼容性验证
+
+仓库内提供无需 schema 的语料测试脚本。游戏文件不会提交到仓库，请让脚本指向本地已解包的
+`natives` 目录：
+
+```bash
+python tests/corpus_probe.py D:/game/natives \
+  --expected-total 62768 \
+  --expected-version 16 \
+  --report layout-report.json
+```
+
+0.7.1 修复了 0.7.0 引入的现代 RSZ 对齐回归，并把原计划中的兼容性工作合并进本次 bugfix：
+分阶段布局探测、结构化诊断、安全 readable 解析、仅验证布局可 repack，以及无需 schema 的
+probe 命令。新的现代布局规则已用 62,768 个 MHWS `.user.3` 文件验证；其中四个
+SystemSetting 样本还完成了 readable 导出、repack 导出、逐字节一致回封和二次解析验证。
+
+另外还验证了 Monster Hunter Stories 3 语料：42,945 个文件全部通过严格 layout 探测和
+schema 驱动的 repack 导出；42,945 个文件全部能够无异常地重新构建，其中 38,548 个逐字节
+一致。其余 4,397 个可以成功重建但并非逐字节一致；已观察到的原因包括非零 padding、空字符串
+的另一种物理编码，因此不把它们宣称为逐字节 fixture。两个使用紧凑超大数组表示的 voxel 文件
+都能逐字节一致回封，其中包括 6.55 MB 的 `dg100_Root` payload。
+
 ## 常见注意事项
 
 - schema JSON、`il2cpp_dump.json` 和 `.user.3` 文件应来自同一个游戏版本；

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "PyREUser3"
-version = "0.7.0"
+version = "0.7.1"
 description = "Pure Python tools for converting RE Engine .user.3 files to and from JSON."
 readme = "docs/PYPI.md"
 requires-python = ">=3.9"

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/pyreuser3/__init__.py

@@ -9,7 +9,7 @@ from __future__ import annotations
 from importlib import import_module
 from typing import Any
 
-__version__ = "0.7.0"
+__version__ = "0.7.1"
 
 __all__ = [
     "__version__",
@@ -18,9 +18,12 @@ __all__ = [
     "FieldDef",
     "PackError",
     "ParseError",
+    "LayoutValidationIssue",
     "TypeDB",
     "REUser3Converter",
     "RSZ_MAGIC",
+    "probe_usr_file",
+    "probe_usr_path",
     "User3Exporter",
     "User3Packer",
     "USR_MAGIC",
@@ -31,6 +34,9 @@ _EXPORT_MODULES = {
     "ParseError": ".core",
     "RSZ_MAGIC": ".core",
     "USR_MAGIC": ".core",
+    "LayoutValidationIssue": ".usr_container",
+    "probe_usr_file": ".usr_container",
+    "probe_usr_path": ".usr_container",
     "ClassDef": ".schema",
     "FieldDef": ".schema",
     "TypeDB": ".schema",

{pyreuser3-0.7.0 → pyreuser3-0.7.1}/pyreuser3/api.py

@@ -10,12 +10,15 @@ from __future__ import annotations
 import inspect
 import json
 import re
+import threading
 from pathlib import Path
 from typing import Any, Callable, Literal, Optional
 
-from .core import RSZ_MAGIC, USR_MAGIC
+from .core import RSZ_MAGIC, USR_MAGIC, resolve_schema_path
 from .export import User3Exporter
 from .pack import User3Packer
+from .schema import TypeDB
+from .usr_container import probe_usr_file, probe_usr_path
 
 # Preserve the exported JSON structure so external scripts and hand-edited files remain
 # compatible across workflows.
@@ -67,6 +70,11 @@ class REUser3Converter:
         self.tree_depth = tree_depth
         self.user_magic = int(user_magic)
         self.rsz_magic = int(rsz_magic)
+        self._metadata_cache_lock = threading.RLock()
+        self._typedb_cache: tuple[tuple[str, int, int], TypeDB] | None = None
+        self._il2cpp_metadata_cache: (
+            tuple[tuple[str, int, int], tuple[dict, dict]] | None
+        ) = None
 
     def export_directory(
         self,
@@ -94,6 +102,7 @@ class REUser3Converter:
             exclude_regexes,
             json_format=self._normalize_json_format(json_format),
         )
+        self._prepare_exporter_metadata(exporter)
         return exporter.run()
 
     def export_file(
@@ -189,10 +198,42 @@ class REUser3Converter:
         Returns:
             JsonTree: JSON-compatible tree used by export, editing, or packing workflows.
         """
-        exporter = self._new_exporter(user3_path, Path.cwd(), [])
+        exporter = self._new_exporter(
+            user3_path,
+            Path.cwd(),
+            [],
+            json_format="repack",
+        )
         self._prepare_exporter_metadata(exporter)
         return exporter._parse_user3_pack(Path(user3_path))
 
+    def probe_user3(self, user3_path: str | Path, strict: bool = False) -> dict[str, Any]:
+        """Inspect one container without loading schema or enum metadata."""
+
+        return probe_usr_file(
+            user3_path,
+            user_magic=self.user_magic,
+            rsz_magic=self.rsz_magic,
+            policy="strict_probe" if strict else "safe_read",
+        )
+
+    def probe_directory(
+        self,
+        user3_root: str | Path,
+        *,
+        strict: bool = False,
+        include_successes: bool = False,
+    ) -> dict[str, Any]:
+        """Inspect a file tree without loading schema or enum metadata."""
+
+        return probe_usr_path(
+            user3_root,
+            user_magic=self.user_magic,
+            rsz_magic=self.rsz_magic,
+            policy="strict_probe" if strict else "safe_read",
+            include_successes=include_successes,
+        )
+
     @staticmethod
     def _normalize_json_format(json_format: str) -> str:
         """Normalize the public in-memory JSON format selector.
@@ -310,7 +351,8 @@ class REUser3Converter:
         include_patterns = [re.compile(p) for p in (include_regexes or [])]
         exclude_patterns = [re.compile(p) for p in (exclude_regexes or [])]
 
-        total = success = failed = skipped = 0
+        selected: list[tuple[Path, Path]] = []
+        skipped = 0
         for file_path in files:
             # Resolve and validate paths at the boundary so later code never guesses
             # relative to a surprising working directory.
@@ -327,22 +369,36 @@ class REUser3Converter:
             if any(pattern.search(rel) for pattern in exclude_patterns):
                 skipped += 1
                 continue
+            selected.append((file_path, Path(rel)))
 
-            total += 1
-            output_path = target_root / (
-                file_path.name
-                if source_root.is_file()
-                else file_path.relative_to(source_root)
-            )
+        if not selected:
+            return {"total": 0, "success": 0, "failed": 0, "skipped": skipped}
+
+        exporter = self._new_exporter(
+            source_root,
+            target_root,
+            [],
+            json_format="repack",
+        )
+        self._prepare_exporter_metadata(exporter)
+        packer = self._new_packer(target_root)
+        success = failed = 0
+        for file_path, relative_path in selected:
+            output_path = target_root / relative_path
             try:
                 # Treat each file independently so one malformed resource is reported
                 # but does not stop the rest of the batch.
-                self.patch_file(file_path, output_path, callback)
+                data = exporter._parse_user3_pack(file_path)
+                modified = self._run_callback(callback, data, file_path)
+                if modified is None:
+                    modified = data
+                output_path.parent.mkdir(parents=True, exist_ok=True)
+                output_path.write_bytes(packer.pack(modified))
                 success += 1
             except Exception:
                 failed += 1
         return {
-            "total": total,
+            "total": len(selected),
             "success": success,
             "failed": failed,
             "skipped": skipped,
@@ -372,6 +428,8 @@ class REUser3Converter:
         """
         if self.il2cpp_dump_path is None:
             raise FileNotFoundError("il2cpp_dump_path is required for exporting JSON")
+        typedb = self._get_cached_typedb()
+        il2cpp_metadata = self._get_cached_il2cpp_metadata()
         return User3Exporter(
             user3_root=user3_root,
             schema_dir=self.schema_path,
@@ -382,6 +440,8 @@ class REUser3Converter:
             user_magic=self.user_magic,
             rsz_magic=self.rsz_magic,
             json_format=json_format,
+            preloaded_typedb=typedb,
+            preloaded_il2cpp_metadata=il2cpp_metadata,
         )
 
     def _new_packer(self, output_root: str | Path | None) -> User3Packer:
@@ -399,6 +459,8 @@ class REUser3Converter:
             output_root=output_root,
             user_magic=self.user_magic,
             rsz_magic=self.rsz_magic,
+            preloaded_typedb=self._get_cached_typedb(),
+            preloaded_il2cpp_metadata=self._get_cached_il2cpp_metadata(),
         )
 
     def _prepare_exporter_metadata(self, exporter: User3Exporter) -> None:
@@ -419,14 +481,58 @@ class REUser3Converter:
         # parsing keeps the same lookup only in memory.
         # Keep the generated metadata attached to the exporter so field parsing can
         # format enum names consistently.
-        enums_internal, enum_context = exporter.export_il2cpp_metadata_from_path(
-            self.il2cpp_dump_path
-        )
+        metadata = self._get_cached_il2cpp_metadata()
+        if metadata is None:
+            raise FileNotFoundError("il2cpp_dump_path is required for parsing JSON")
+        enums_internal, enum_context = metadata
         exporter.enum_lookup = exporter._build_enum_lookup_from_enums_internal(
             enums_internal
         )
         exporter._apply_enum_context(enum_context)
         exporter._ensure_enum_lookup()
+        exporter._metadata_prepared = True
+
+    @staticmethod
+    def _metadata_file_signature(path: Path) -> tuple[str, int, int]:
+        """Return an inexpensive cache key that changes when a metadata file changes."""
+
+        resolved = path.resolve(strict=True)
+        stat = resolved.stat()
+        return str(resolved), stat.st_size, stat.st_mtime_ns
+
+    def _get_cached_typedb(self) -> TypeDB:
+        """Load the schema once per converter and invalidate it after file changes."""
+
+        schema_path = resolve_schema_path(self.schema_path)
+        signature = self._metadata_file_signature(schema_path)
+        with self._metadata_cache_lock:
+            if self._typedb_cache is None or self._typedb_cache[0] != signature:
+                self._typedb_cache = (signature, TypeDB.load(schema_path))
+            return self._typedb_cache[1]
+
+    def _get_cached_il2cpp_metadata(self) -> tuple[dict, dict] | None:
+        """Stream the il2cpp dump once per converter and invalidate after changes."""
+
+        if self.il2cpp_dump_path is None:
+            return None
+        signature = self._metadata_file_signature(self.il2cpp_dump_path)
+        with self._metadata_cache_lock:
+            if (
+                self._il2cpp_metadata_cache is None
+                or self._il2cpp_metadata_cache[0] != signature
+            ):
+                metadata = User3Exporter.export_il2cpp_metadata_from_path(
+                    self.il2cpp_dump_path
+                )
+                self._il2cpp_metadata_cache = (signature, metadata)
+            return self._il2cpp_metadata_cache[1]
+
+    def clear_metadata_cache(self) -> None:
+        """Discard cached schema and il2cpp metadata for this converter."""
+
+        with self._metadata_cache_lock:
+            self._typedb_cache = None
+            self._il2cpp_metadata_cache = None
 
     @staticmethod
     def _discover_user3_files(user3_root: Path) -> list[Path]: