Split web/ into senbei-wasm crate + static assets

The Rust bindings move from web/src to a top-level senbei-wasm crate (still outside the workspace, own Cargo.lock), matching the other senbei-* crates. web/ keeps only the static frontend; wasm-pack emits the JS/wasm package into web/pkg/ via --out-dir. JS glue renamed senbei_web -> senbei_wasm with the crate.
Merge bfloat16-senbei workspace restructure, bump to 1.1.0
2026-09-19 03:57:59 -04:00 · 2026-08-30 22:44:15 +08:00 · 2026-08-30 22:31:21 +08:00 · 2026-08-16 05:51:39 +08:00 · 2026-08-16 05:47:16 +08:00 · 2026-08-16 05:33:44 +08:00
58 changed files with 6239 additions and 3959 deletions
@@ -19,7 +19,7 @@ jobs:
    runs-on: windows-latest
    steps:
      - uses: actions/checkout@v4
-      - run: cargo clippy --all-targets -- -D warnings
+      - run: cargo clippy --workspace --all-targets -- -D warnings

  test:
    # The test suite exercises Windows path semantics, so it runs on Windows.
@@ -28,7 +28,7 @@ jobs:
    runs-on: windows-latest
    steps:
      - uses: actions/checkout@v4
-      - run: cargo test --release
+      - run: cargo test --release --workspace

  check-portable:
    # Build-only portability gate: non-Windows host and the wasm target the
@@ -39,8 +39,8 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
-      - run: cargo clippy --all-targets -- -D warnings
-      - run: cargo check --target wasm32-unknown-unknown
+      - run: cargo clippy --workspace --all-targets -- -D warnings
+      - run: cargo check --workspace --target wasm32-unknown-unknown

  cli:
    strategy:
@@ -82,11 +82,12 @@ jobs:
    steps:
      - uses: actions/checkout@v4
      - run: cargo install wasm-pack --locked
-      # `-- --locked` forwards to cargo: web/Cargo.lock is committed on
+      # `-- --locked` forwards to cargo: senbei-wasm/Cargo.lock is committed on
      # purpose, so the wasm build must be pinned by it rather than silently
      # re-resolving (which is how it drifted out of sync with the manifest).
-      - run: wasm-pack build --target web --release -- --locked
-        working-directory: web
+      # --out-dir emits the JS/wasm package into the static frontend's web/pkg/.
+      - run: wasm-pack build --target web --release --out-dir ../web/pkg -- --locked
+        working-directory: senbei-wasm
      - uses: actions/upload-artifact@v4
        with:
          name: senbei-web
@@ -62,11 +62,12 @@ jobs:
    steps:
      - uses: actions/checkout@v4
      - run: cargo install wasm-pack --locked
-      # `-- --locked` forwards to cargo: web/Cargo.lock is committed on
+      # `-- --locked` forwards to cargo: senbei-wasm/Cargo.lock is committed on
      # purpose, so the wasm build must be pinned by it rather than silently
      # re-resolving (which is how it drifted out of sync with the manifest).
-      - run: wasm-pack build --target web --release -- --locked
-        working-directory: web
+      # --out-dir emits the JS/wasm package into the static frontend's web/pkg/.
+      - run: wasm-pack build --target web --release --out-dir ../web/pkg -- --locked
+        working-directory: senbei-wasm
      - name: Stage static site
        run: |
          mkdir dist
@@ -153,6 +153,8 @@ target/
 !/samples/README.md

 ### senbei web build ###
+# The wasm crate builds out-of-workspace in senbei-wasm/ and wasm-pack emits
+# the JS/wasm package into web/pkg/ for the static frontend.
+/senbei-wasm/target/
 /web/pkg/
-/web/target/
 /web/.playwright-cli
@@ -4,19 +4,22 @@ Guidance for AI coding agents (and human contributors) working in this repo.

 ## Project

-Senbei is a static unpacker for Crackproof-protected PE files: a pure,
-panic-free, no-I/O unpacker core (`src/unpacker/`) plus a thin CLI shell
-(`src/`), an il2cpp metadata de-obfuscator (`src/metadata.rs`), and a
-WebAssembly browser frontend (`web/`). Read `docs/design.md` first.
+Senbei is a static unpacker for Crackproof-protected PE files: a Cargo
+workspace with a pure, panic-free, no-I/O unpacker core (`senbei-pe/`, built
+on `senbei-crypto/`), an il2cpp metadata de-obfuscator (`senbei-metadata/`),
+filesystem/CLI orchestration (`senbei-io/`), the `senbei` binary
+(`senbei-cli/`), WebAssembly bindings (`senbei-wasm/`, outside the workspace;
+builds into `web/pkg/`), and the static browser frontend assets (`web/`).
+Read `docs/design.md` first.

 ## Commands

 ```cmd
-cargo build --release            :: CLI
-cargo test --release             :: full suite (golden corpus: samples/, git-ignored)
-cargo clippy --all-targets -- -D warnings
+cargo build --release            :: CLI (default member: senbei-cli)
+cargo test --release --workspace :: full suite (golden corpus: samples/, git-ignored)
+cargo clippy --workspace --all-targets -- -D warnings
 cargo fmt --all
-cd web && wasm-pack build --target web --release   :: browser build
+cd senbei-wasm && wasm-pack build --target web --release --out-dir ../web/pkg   :: browser build
 ```

 The `samples/` corpus is user-managed and absent on CI; without it the
@@ -62,21 +62,21 @@ checksum = "da7c62ceae207dd37ea5b845da6a0696c799f85e97da1ab5b7910be3c1c80223"

 [[package]]
 name = "futures-core"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2cd50c473c80f6d7c3670a752354b8e569b1a7cbfdc0419ec88e5edad85e0dc7"
+checksum = "92d699e522242e69e3003b94ecc1f960f3a5e015aa7c5d7486e65ad01dd94f5e"

 [[package]]
 name = "futures-task"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b231ed28831efb4a61a08580c4bc233ec56bc009f4cd8f52da2c3cb97df0c109"
+checksum = "cd417de3d1d015fc3bfd2b1ea46dfc7bab72ef86f1cc7cc9c78e728b34a6d1fd"

 [[package]]
 name = "futures-util"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a77a90a256fce34da66415271e30f94ee91c57b04b8a2c042d9cf3220179deaa"
+checksum = "0d50a92467f8ba5dd6e3ee5d4bd04d73ab2e4e1c44474a0674821dfce14b79bc"
 dependencies = [
 "futures-core",
 "futures-task",
@@ -110,9 +110,9 @@ dependencies = [

 [[package]]
 name = "js-sys"
-version = "0.3.103"
+version = "0.3.104"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "53b44bfcdb3f8d5837a46dae1ca9660a837176eee74a28b229bc626816589102"
+checksum = "0e0c1080212aad755ea003d18543e8768dd432c48819efd73a7bf1e39b7a5a3a"
 dependencies = [
 "cfg-if",
 "futures-util",
@@ -139,9 +139,9 @@ checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"

 [[package]]
 name = "owo-colors"
-version = "4.3.0"
+version = "4.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d211803b9b6b570f68772237e415a029d5a50c65d382910b879fb19d3271f94d"
+checksum = "13c45bb4a6ae1280ec0803b1ef9d3455eb50f01efbbe1447ab020f1d54fba9d8"

 [[package]]
 name = "pin-project-lite"
@@ -151,9 +151,9 @@ checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"

 [[package]]
 name = "portable-atomic"
-version = "1.14.0"
+version = "1.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d20d5497ef88037a52ff98267d066e7f11fcc5e99bbfbd58a42336193aacec3"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"

 [[package]]
 name = "proc-macro2"
@@ -208,19 +208,48 @@ dependencies = [
 ]

 [[package]]
-name = "senbei"
-version = "1.0.0"
+name = "senbei-cli"
+version = "1.1.0"
+dependencies = [
+ "senbei-io",
+ "senbei-metadata",
+ "tempfile",
+]
+
+[[package]]
+name = "senbei-crypto"
+version = "1.1.0"
+dependencies = [
+ "thiserror",
+]
+
+[[package]]
+name = "senbei-io"
+version = "1.1.0"
 dependencies = [
 "anyhow",
 "indicatif",
 "libc",
 "owo-colors",
+ "senbei-metadata",
+ "senbei-pe",
 "tempfile",
- "thiserror",
 "walkdir",
 "windows",
 ]

+[[package]]
+name = "senbei-metadata"
+version = "1.1.0"
+
+[[package]]
+name = "senbei-pe"
+version = "1.1.0"
+dependencies = [
+ "senbei-crypto",
+ "thiserror",
+]
+
 [[package]]
 name = "slab"
 version = "0.4.12"
@@ -240,9 +269,9 @@ dependencies = [

 [[package]]
 name = "syn"
-version = "3.0.3"
+version = "3.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "53e9bae58849f64dfa4f5d5ae372c8341f7305f82a3868709269343628b659a3"
+checksum = "e6275cddf4610d1775e6d1fe9469b2e77d0f39fd98fb7450901b821e0c53649f"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -264,22 +293,22 @@ dependencies = [

 [[package]]
 name = "thiserror"
-version = "2.0.19"
+version = "2.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09a43598840e33d5b0331f38c5e30d13bb11c11210a4b58f0d9b18a5a5eefcd9"
+checksum = "ec86235f5fcc2a73650310756d2ac5b138a5780bbbdfae3eeccec992c435ba4f"
 dependencies = [
 "thiserror-impl",
 ]

 [[package]]
 name = "thiserror-impl"
-version = "2.0.19"
+version = "2.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "43cbfe0cf76104d42a574802844187e84a305e531ed54455f11fbde0f10541cd"
+checksum = "bc04cd3e1236dd4a98afca4569f2deb3f120e5422a4023be2cb683f8486292af"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 3.0.3",
+ "syn 3.0.4",
 ]

 [[package]]
@@ -312,9 +341,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4b067c0c11094aef6b7a801c1e34a26affafdf3d051dba08456b868789aaf9a4"
+checksum = "1b70935747edd64d89de3efa29d73789b806c15798f8e7dca4d8ac356b50ce70"
 dependencies = [
 "cfg-if",
 "once_cell",
@@ -325,9 +354,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "167ce5e579f6bcf889c4f7175a8a5a585de84e8ff93976ce393efa5f2837aab1"
+checksum = "77775f8f3f7217702089053b94958f8f54061a3f663417df76e19cbdcca29bc1"
 dependencies = [
 "quote",
 "wasm-bindgen-macro-support",
@@ -335,9 +364,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f3997c7839262f4ef12cf90b818d6340c18e80f263f1a94bf157d0ec4420380e"
+checksum = "e11d33f857dc2fb11b8bc75aee111aa9cbeb12cd9f25efd3d4c2a3dd4e235284"
 dependencies = [
 "bumpalo",
 "proc-macro2",
@@ -348,9 +377,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc1b4cb0cc549fcf58d7dfc081778139b3d283a081644e833e84682ad71cea24"
+checksum = "7ef64dbcc55df09c7e5a46182d181c2cfa3e925f3da937ea764728b4bbb9dcbf"
 dependencies = [
 "unicode-ident",
 ]
@@ -1,39 +1,39 @@
-[package]
-name = "senbei"
-version = "1.0.0"
+[workspace]
+members = [
+    "senbei-cli",
+    "senbei-crypto",
+    "senbei-io",
+    "senbei-metadata",
+    "senbei-pe",
+]
+default-members = ["senbei-cli"]
+# The wasm bindings crate is its own crate (own Cargo.lock, cdylib) and stays
+# outside the workspace.
+exclude = ["senbei-wasm"]
+resolver = "2"
+
+[workspace.package]
+version = "1.1.0"
 edition = "2024"
-description = "Static unpacker for Crackproof-protected PE files"
 license = "AGPL-3.0-only"
-keywords = ["unpacker", "reverse-engineering", "pe", "security-research"]
-categories = ["command-line-utilities"]

-[lib]
-name = "senbei"
-path = "src/lib.rs"
-
-[[bin]]
-name = "senbei"
-path = "src/main.rs"
-
-[dependencies]
+[workspace.dependencies]
 anyhow = "1"
+indicatif = "0.18"
+libc = "0.2"
+owo-colors = "4"
+tempfile = "3"
 thiserror = "2"
 walkdir = "2"
-indicatif = "0.18"
-owo-colors = "4"
-
-[target.'cfg(windows)'.dependencies]
 windows = { version = "0.62", features = [
    "Win32_Foundation",
    "Win32_System_Console",
    "Win32_System_SystemInformation",
 ] }
-
-[target.'cfg(all(not(windows), not(target_arch = "wasm32")))'.dependencies]
-libc = "0.2"
-
-[dev-dependencies]
-tempfile = "3"
+senbei-crypto = { path = "senbei-crypto" }
+senbei-io = { path = "senbei-io" }
+senbei-metadata = { path = "senbei-metadata" }
+senbei-pe = { path = "senbei-pe" }

 [profile.release]
 opt-level = 3
@@ -41,7 +41,8 @@ lives in [`web/`](web/).

 | Kind | Description |
 | --- | --- |
-| `Exe` | Crackproof-protected executable (PE32+ and PE32). |
+| `NativeExe` | Crackproof-protected native executable (PE32+ and PE32). |
+| `ManagedExe` | Protected .NET executable (has a CLR data directory). |
 | `NativeDll` | Protected native (unmanaged) DLL. |
 | `ManagedDll` | Protected .NET assembly (has a CLR data directory). |
 | `._` companion | Stub + external encrypted payload layout, spliced automatically. |
@@ -7,53 +7,75 @@ no driver or proxy DLL is involved.

 ## Crate layout

-The crate is split into a pure core and a thin CLI shell:
+Senbei is a Cargo workspace split into a pure core and thin shells around it:

- **`src/unpacker/`** — the core. Pure functions over byte slices: no file
-  I/O, no environment access (beyond a few debugging overrides, see
+- **`senbei-pe/`** — the core. Pure functions over byte slices: no file I/O,
+  no environment access (beyond a few debugging overrides, see
  [development.md](development.md)), panic-free at the public boundary (all
  internal panics are trapped and converted to `UnpackError::Corrupt`). This
  is what the WebAssembly build embeds.
- **`src/` (top level)** — the CLI shell: argument parsing, recursive folder
-  scanning, per-run log file, progress bar, Explorer-friendly exit pause, and
-  the single-file/folder orchestration in `job.rs`.
- **`src/metadata.rs`** — il2cpp `global-metadata.dat` method-token
+- **`senbei-crypto/`** — cryptographic, checksum, compression, and bytecode
+  primitives the core is built from. Same purity rules as `senbei-pe`.
+- **`senbei-metadata/`** — il2cpp `global-metadata.dat` method-token
  de-obfuscation (format version 31; other versions are left untouched).
+- **`senbei-io/`** — filesystem and orchestration: recursive folder scanning,
+  per-run log file, progress bar, Explorer-friendly exit pause, and the
+  single-file/folder orchestration in `job.rs` (incl. the wasm-safe in-memory
+  byte API used by the web frontend).
+- **`senbei-cli/`** — the `senbei` binary: argument parsing + dispatch. The
+  integration test suite (incl. the golden corpus test) lives in
+  `senbei-cli/tests/`.

 ```
-src/
-├── main.rs                argument parsing + dispatch
-├── lib.rs                 module roots
+senbei-cli/
+└── src/main.rs            argument parsing + dispatch
+senbei-io/src/
 ├── job.rs                 single-file + folder orchestration, out-naming,
 │                          companion splice, stub overlay/TLS restore,
 │                          pipeline routing (incl. the wasm-safe byte API)
 ├── scan.rs                recursive Crackproof + metadata discovery
-├── metadata.rs            il2cpp global-metadata.dat de-obfuscation
 ├── logfile.rs             per-run timestamped log
 ├── ui.rs                  progress bar + status lines
-├── pause.rs               Explorer-friendly exit pause
-└── unpacker/              pure, panic-free, no-I/O core
-    ├── mod.rs             detection + unpack_auto dispatch
-    ├── exe.rs             EXE pipeline (PE32+ and PE32)
-    ├── dll.rs             native + managed DLL pipeline
-    ├── integrity.rs       static post-unpack sanity check
-    ├── primitives.rs      decrypt_data* steps, key/shift selection
-    ├── bytecode.rs        bytecode VM
-    ├── parallel.rs        deterministic block-parallel fan-out
-    ├── tables.rs          constant tables
-    └── crc32.rs           checksum
+└── pause.rs               Explorer-friendly exit pause
+senbei-metadata/src/
+└── metadata.rs            il2cpp global-metadata.dat de-obfuscation
+senbei-crypto/src/
+├── primitives.rs          decrypt_data* steps, key derivation
+├── bytecode.rs            bytecode VM
+├── tables.rs              constant tables
+└── crc32.rs               checksum
+senbei-pe/src/engine/      pure, panic-free, no-I/O core
+├── mod.rs                 detection + unpack_auto dispatch
+├── error.rs               structured error taxonomy
+├── integrity.rs           static post-unpack sanity check
+├── parallel.rs            deterministic block-parallel fan-out
+├── layout/                layout discovery + validation
+│   ├── dd8.rs             .text dd8 key-formula + shift selection
+│   ├── discovery.rs       layout candidate discovery (trial-and-validate)
+│   └── image.rs           PE image reconstruction helpers
+├── exe/
+│   ├── pipeline.rs        EXE pipeline (PE32+ and PE32 orchestration)
+│   └── pipeline/pe32.rs   PE32-specific EXE restore
+└── dll/
+    └── pipeline.rs        native + managed DLL pipeline
 ```

 ## Detection and routing

 Detection is content-based (`unpacker::detect`), never extension-based: the
 key table is derived from the file header and checked against the format
-magic, then the PE characteristics classify the input as EXE, native DLL, or
-managed DLL.
+magic, then the PE characteristics classify the input as EXE or DLL and the
+CLR data directory splits each into native vs managed (`NativeExe` /
+`ManagedExe` / `NativeDll` / `ManagedDll`).

 `unpack_auto` then dispatches:

- `Exe` → the EXE pipeline (handles both PE32+ and PE32).
+- `NativeExe` / `ManagedExe` → the EXE pipeline (handles both PE32+ and
+  PE32). Managed EXEs take the same path: their import-string table is null
+  (imports are the CLR bootstrap stub), the entry point comes from the
+  protected header (the config block stores 0 for managed images), and the
+  COR20 header, BSJB metadata stream, and CLR resources are restored verbatim
+  from the protected file, mirroring the managed-DLL restore.
 - `NativeDll` / `ManagedDll` → the DLL pipeline first; on failure, the EXE
  pipeline as a fallback. Two DLL layouts exist in the wild: an older layout
  the DLL pipeline parses, and a newer one that protects DLLs with the
@@ -60,9 +60,9 @@ since binaries are not committed).

 ## Conventions

- The `src/unpacker/` core is pure: no file I/O, no panics across the public
-  boundary, no `unsafe`. Keep it that way — it is what the WebAssembly build
-  embeds.
+- The `senbei-pe/` core (and its `senbei-crypto/` base) is pure: no file I/O,
+  no panics across the public boundary, no `unsafe`. Keep it that way — it is
+  what the WebAssembly build embeds.
 - Layout heuristics must **trial-and-validate**: never pick a candidate offset
  on shape alone and trust it; validate by decryption/checksum and fall
  through to the next candidate on failure. A silent wrong offset produces a
@@ -81,12 +81,18 @@ since binaries are not committed).

 ```
 senbei/
-├── Cargo.toml                 senbei lib + bin package
+├── Cargo.toml                 workspace root (members: the senbei-* crates)
 ├── rust-toolchain.toml        pinned toolchain + targets
-├── src/                       CLI shell + pure unpacker core (see docs/design.md)
-├── tests/                     CLI, detection, golden, and folder tests
+├── senbei-cli/                senbei binary (default member)
+│   └── tests/                 CLI, detection, golden, and folder tests
+├── senbei-pe/                 pure unpacker core (see docs/design.md)
+├── senbei-crypto/             crypto/compression primitives
+├── senbei-metadata/           il2cpp metadata de-obfuscation
+├── senbei-io/                 filesystem, scanning, CLI orchestration
+├── senbei-wasm/               WebAssembly bindings crate (own Cargo.lock,
+│                              outside the workspace; builds into web/pkg/)
 ├── samples/                   local-only test corpus (git-ignored)
-├── web/                       WebAssembly browser build
+├── web/                       static browser frontend assets (+ built pkg/)
 ├── docs/                      usage, design, and development documentation
 └── .github/                   CI workflows and issue templates
 ```
@@ -96,8 +102,8 @@ senbei/
 See [web/README.md](../web/README.md). In short:

 ```cmd
-cd web
-wasm-pack build --target web --release
+cd senbei-wasm
+wasm-pack build --target web --release --out-dir ../web/pkg
 ```

 then serve `web/` statically and open `index.html`. Everything runs
@@ -0,0 +1,20 @@
+[package]
+name = "senbei-cli"
+version.workspace = true
+edition.workspace = true
+description = "Command-line entry point for Senbei"
+license.workspace = true
+keywords = ["unpacker", "reverse-engineering", "pe", "security-research"]
+categories = ["command-line-utilities"]
+
+[[bin]]
+name = "senbei"
+path = "src/main.rs"
+
+[dependencies]
+senbei-io.workspace = true
+
+[dev-dependencies]
+senbei-io.workspace = true
+senbei-metadata.workspace = true
+tempfile.workspace = true
@@ -1,4 +1,4 @@
-use senbei::{job, pause};
+use senbei_io::{job, pause, scan};
 use std::path::Path;

 fn main() -> std::process::ExitCode {
@@ -27,9 +27,6 @@ fn main() -> std::process::ExitCode {
            "--no-log" => no_log = true,
            "--scan-all" => scan_all = true,
            "--out" => match args.next() {
-                // Reject a missing value (and a following flag swallowed as the
-                // value): previously `--out` at end of argv silently fell back
-                // to the default output directory.
                Some(v) if !v.starts_with('-') => out = Some(v),
                _ => {
                    eprintln!("error: --out requires a directory argument");
@@ -42,7 +39,6 @@ fn main() -> std::process::ExitCode {
                return std::process::ExitCode::from(2);
            }
            other => {
-                // Previously the last positional silently won.
                if let Some(prev) = &path {
                    eprintln!("error: multiple input paths given ('{prev}' and '{other}')");
                    return std::process::ExitCode::from(2);
@@ -63,33 +59,36 @@ fn main() -> std::process::ExitCode {
            }
            let p = Path::new(&p);
            let out_path = out.as_deref().map(Path::new);
-            let r = if p.is_dir() {
+            let result = if p.is_dir() {
                job::run_folder_opts(
                    p,
                    out_path,
                    quiet,
                    verbose,
                    no_log,
-                    scan_all || senbei::scan::scan_all_env(),
+                    scan_all || scan::scan_all_env(),
                )
            } else {
                job::run_file_v(p, out_path, quiet, verbose, no_log)
            };
-            match r {
-                Ok(s) => {
+            match result {
+                Ok(summary) => {
                    if quiet < 2 {
                        println!(
                            "{} unpacked · {} skipped · {} errors · {} suspect · {} metadata",
-                            s.unpacked, s.skipped, s.errors, s.suspect, s.metadata
+                            summary.unpacked,
+                            summary.skipped,
+                            summary.errors,
+                            summary.suspect,
+                            summary.metadata
                        );
-                        println!("done in {} ms", s.duration_ms);
+                        println!("done in {} ms", summary.duration_ms);
                    }
-                    if s.errors > 0 { 1 } else { 0 }
+                    if summary.errors > 0 { 1 } else { 0 }
                }
-                Err(e) => {
-                    // Fatal: out-dir/log create, etc.
+                Err(error) => {
                    if quiet < 2 {
-                        eprintln!("error: {e:#}");
+                        eprintln!("error: {error:#}");
                    }
                    1
                }
@@ -107,7 +106,7 @@ fn print_help() {
    );
    println!(
        "  --scan-all   probe every file in a folder, including ones the scan\n\
-         \x20              pre-filter skips (under 4128 bytes, or a bulk-asset\n\
-         \x20              extension like .ab/.xml/.acb). Much slower on game trees."
+         \x20              pre-filter skips (under 4128 bytes, extensionless,\n\
+         \x20              or a bulk-asset extension). Much slower on large trees."
    );
 }
@@ -0,0 +1,11 @@
+//! Shared test fixtures.
+#![allow(dead_code)]
+
+use std::path::PathBuf;
+
+/// Path to the workspace-root `samples/` — the user-managed corpus dropped in
+/// by hand. Git-ignored except its README; tests here run against whatever is
+/// present. `CARGO_MANIFEST_DIR` is `senbei-cli/`, so go one level up.
+pub fn samples_dir() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../samples")
+}
@@ -1,4 +1,4 @@
-use senbei::job::{default_out_root_for_file, out_name};
+use senbei_io::job::{default_out_root_for_file, out_name};
 use std::path::Path;

 #[test]
@@ -1,4 +1,4 @@
-use senbei::logfile::{Log, local_stamp_compact, local_stamp_display};
+use senbei_io::logfile::{Log, local_stamp_compact, local_stamp_display};

 #[test]
 fn local_stamp_compact_matches_shape() {
@@ -1,4 +1,4 @@
-use senbei::job;
+use senbei_io::job;
 use std::path::Path;

 fn list_logs(dir: &Path) -> Vec<std::path::PathBuf> {
@@ -8,7 +8,7 @@
 //! - golden present, bytes differ     -> FAIL (the test fails)
 //! - no golden                        -> WARNING (printed; needs a manual check)
 //!
-//! Inputs go through [`senbei::job::unpack_bytes`], the same routing the CLI
+//! Inputs go through [`senbei_io::job::unpack_bytes`], the same routing the CLI
 //! uses, **not** `unpack_auto` directly. That matters: `unpack_auto` alone
 //! cannot reach the external-companion layout, whose stub is meaningless
 //! without its `<name>._` payload — a corpus wired to `unpack_auto` silently
@@ -17,7 +17,7 @@
 //! samples folder is picked up automatically, exactly as it is on disk.
 //!
 //! An input whose bytes carry the il2cpp metadata magic is routed through
-//! [`senbei::metadata::deobfuscate`] instead, giving the method-token remap
+//! [`senbei_metadata::deobfuscate`] instead, giving the method-token remap
 //! real-world coverage (its unit tests only build synthetic layouts).
 //!
 //! The folder is git-ignored (see `senbei/samples/README.md`), so the set of
@@ -116,10 +116,10 @@ fn samples_unpack_against_goldens() {
            }
        };

-        let got = if senbei::metadata::is_metadata(&bytes) {
+        let got = if senbei_metadata::is_metadata(&bytes) {
            // il2cpp metadata: method-token de-obfuscation, no PE pipeline and
            // no integrity check (the output is not a PE image).
-            match senbei::metadata::deobfuscate(&bytes) {
+            match senbei_metadata::deobfuscate(&bytes) {
                Ok((out, _report)) => out,
                Err(e) => {
                    failures.push(format!("{name}: de-obfuscation failed: {e}"));
@@ -140,7 +140,7 @@ fn samples_unpack_against_goldens() {
                },
                None => None,
            };
-            let image = match senbei::job::unpack_bytes(&bytes, companion.as_deref()) {
+            let image = match senbei_io::job::unpack_bytes(&bytes, companion.as_deref()) {
                Ok(img) => img,
                Err(e) => {
                    failures.push(format!("{name}: unpack failed: {e:?}"));
@@ -0,0 +1,9 @@
+[package]
+name = "senbei-crypto"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+description = "Cryptographic and compression primitives for Senbei"
+
+[dependencies]
+thiserror.workspace = true
@@ -61,8 +61,8 @@ impl OpsLut {
 pub fn generate(data: &[u8], offset: u32) -> Option<Vec<Op>> {
    // Bounds-checked cursor: a corrupt `data_offset` (bad decrypt_data6 / the
    // alignment fallback) must yield `None`, not an out-of-bounds panic — the
-    // panic path would surface as a misleading `UnpackError::Corrupt` instead
-    // of the precise `BytecodeGenFailed`, and any future caller without a
+    // panic path would surface as a misleading `UnpackError::InternalPanic` instead
+    // of the precise `BytecodeGenerationFailed`, and any future caller without a
    // `catch_unwind` wrapper would abort outright.
    let mut pos = offset as usize;
    let mut next = move || {
@@ -0,0 +1,77 @@
+//! Cryptographic, checksum, compression, and bytecode primitives.
+
+pub mod bytecode;
+pub mod crc32;
+pub mod primitives;
+mod tables;
+
+/// Maximum buffer size accepted by allocation-sensitive transforms.
+pub const MAX_IMAGE_SIZE: u64 = 1 << 30;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum BufferOperation {
+    Read,
+    CopySource,
+    CopyDestination,
+    ZeroFill,
+}
+
+impl std::fmt::Display for BufferOperation {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::Read => "read",
+            Self::CopySource => "copy source",
+            Self::CopyDestination => "copy destination",
+            Self::ZeroFill => "zero-fill",
+        })
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum Error {
+    #[error(
+        "{operation} range out of bounds (offset {offset}, size {size}, buffer length {buffer_len})"
+    )]
+    BufferRangeOutOfBounds {
+        operation: BufferOperation,
+        offset: usize,
+        size: usize,
+        buffer_len: usize,
+    },
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, thiserror::Error)]
+#[non_exhaustive]
+pub enum DecompressionFailure {
+    #[error("compressed source size {size} exceeds limit {max}")]
+    SourceTooLarge { size: u32, max: u64 },
+    #[error("Huffman code length {bits} is invalid")]
+    InvalidCodeLength { bits: u8 },
+    #[error("Huffman tree traversal exceeded 64 levels")]
+    HuffmanTraversalLimit,
+    #[error("pending length accumulator overflowed at {pending}")]
+    PendingLengthOverflow { pending: u32 },
+    #[error("output step {step} at byte {written} exceeds expected size {expected}")]
+    OutputOverflow {
+        written: u32,
+        step: u32,
+        expected: u32,
+    },
+    #[error("run-fill width {width} reads before output offset 0x{destination:08X}")]
+    RunFillBeforeOutput { width: u32, destination: u32 },
+    #[error("run-fill width {width} is unsupported")]
+    InvalidRunFillWidth { width: u32 },
+    #[error("back-reference distance {distance} exceeds {written} written bytes")]
+    InvalidBackReference { distance: u32, written: u32 },
+    #[error("Huffman symbol consumed no input and produced no output")]
+    NoProgress,
+    #[error(
+        "output size mismatch (wrote {written}/{expected} bytes after consuming {consumed}/{source_size})"
+    )]
+    OutputSizeMismatch {
+        written: u32,
+        expected: u32,
+        consumed: u32,
+        source_size: u32,
+    },
+}
@@ -146,11 +146,11 @@ mod tests {
    #[test]
    fn generated_tables_match_committed_bytes() {
        assert_eq!(COLUMMIX1.len(), 1024);
-        assert_eq!(super::super::crc32::compute(&COLUMMIX1), 0x7e8d_5d5f);
-        assert_eq!(super::super::crc32::compute(&COLUMMIX2), 0xfcc4_acfc);
-        assert_eq!(super::super::crc32::compute(&COLUMMIX3), 0x637a_f0cd);
-        assert_eq!(super::super::crc32::compute(&COLUMMIX4), 0x1e7b_c381);
-        assert_eq!(super::super::crc32::compute(&SBOX), 0x10fd_6dc1);
+        assert_eq!(crate::crc32::compute(&COLUMMIX1), 0x7e8d_5d5f);
+        assert_eq!(crate::crc32::compute(&COLUMMIX2), 0xfcc4_acfc);
+        assert_eq!(crate::crc32::compute(&COLUMMIX3), 0x637a_f0cd);
+        assert_eq!(crate::crc32::compute(&COLUMMIX4), 0x1e7b_c381);
+        assert_eq!(crate::crc32::compute(&SBOX), 0x10fd_6dc1);
        // Spot-check the first dword of each (matches the original first row).
        assert_eq!(&COLUMMIX1[..4], &[0x50, 0xa7, 0xf4, 0x51]);
        assert_eq!(&COLUMMIX2[..4], &[0xa7, 0xf4, 0x51, 0x50]);
@@ -0,0 +1,23 @@
+[package]
+name = "senbei-io"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+description = "Filesystem, scanning, logging, and CLI orchestration for Senbei"
+
+[dependencies]
+anyhow.workspace = true
+indicatif.workspace = true
+owo-colors.workspace = true
+senbei-metadata.workspace = true
+senbei-pe.workspace = true
+walkdir.workspace = true
+
+[target.'cfg(windows)'.dependencies]
+windows.workspace = true
+
+[target.'cfg(all(not(windows), not(target_arch = "wasm32")))'.dependencies]
+libc.workspace = true
+
+[dev-dependencies]
+tempfile.workspace = true
@@ -1,4 +1,4 @@
-use crate::unpacker;
+use senbei_pe as unpacker;
 use std::path::{Path, PathBuf};

 /// Crackproof header key table lives at this fixed file offset. For the
@@ -387,9 +387,9 @@ pub fn run_folder_v(
 ///
 /// When `scan_all` is true every regular file under `root` is opened and
 /// content-probed, instead of skipping ones the free directory metadata already
-/// rules out (too small to hold a Crackproof key table, or a bulk-asset
-/// extension). See [`crate::scan::find_targets_opts`] — exhaustive scanning is
-/// dramatically slower on asset-heavy game trees and finds the same targets.
+/// rules out (extensionless, too small to hold a Crackproof key table, or a
+/// bulk-asset extension). See [`crate::scan::find_targets_opts`] — exhaustive
+/// scanning is dramatically slower on asset-heavy trees.
 pub fn run_folder_opts(
    root: &Path,
    out_dir: Option<&Path>,
@@ -518,7 +518,7 @@ pub fn run_folder_opts(
    // il2cpp metadata pass. Crackproof's `-GMD` option obfuscates the method
    // tokens in `global-metadata.dat`; de-obfuscate any we find so the unpacked
    // il2cpp game assembly resolves methods instead of indexing its per-module
-    // tables out of bounds (see [`crate::metadata`]). This is additive to the
+    // tables out of bounds (see [`senbei_metadata`]). This is additive to the
    // Crackproof module unpack above — the metadata blob is not itself a
    // Crackproof file.
    for meta in metas {
@@ -660,7 +660,7 @@ pub fn run_file_v(
        let mut buf = [0u8; 4];
        std::fs::File::open(input)
            .and_then(|mut f| f.read_exact(&mut buf))
-            .map(|_| crate::metadata::is_metadata(&buf))
+            .map(|_| senbei_metadata::is_metadata(&buf))
            .unwrap_or(false)
    };

@@ -801,13 +801,13 @@ fn panic_payload(panic: &(dyn std::any::Any + Send)) -> String {
    }
 }

-/// If `e`'s chain contains [`crate::metadata::Error::UnsupportedVersion`],
+/// If `e`'s chain contains [`senbei_metadata::Error::UnsupportedVersion`],
 /// return the version. Used to apply the folder-mode "leave untouched, don't
 /// fail the run" policy to metadata versions this build can't de-obfuscate.
 fn unsupported_version(e: &anyhow::Error) -> Option<u32> {
    for cause in e.chain() {
-        if let Some(crate::metadata::Error::UnsupportedVersion(v)) =
-            cause.downcast_ref::<crate::metadata::Error>()
+        if let Some(senbei_metadata::Error::UnsupportedVersion(v)) =
+            cause.downcast_ref::<senbei_metadata::Error>()
        {
            return Some(*v);
        }
@@ -831,19 +831,14 @@ fn write_atomic(dest: &Path, bytes: &[u8]) -> std::io::Result<()> {
    r
 }

-/// Detect `bytes` and run the right pipeline. The EXE pipeline is invoked
-/// directly (no DLL-pipeline probe) when the input was spliced from an
-/// external companion (`spliced`) or when the caller forces it (`force_exe`
-/// — the web app's recovery path after a DLL-probe trap; see
-/// [`unpack_bytes_force_exe`]).
+/// Detect `bytes` and run the right pipeline. Spliced external companions use
+/// the EXE pipeline directly because that layout is definitionally EXE-style.
 ///
 /// Routing spliced inputs straight to the EXE pipeline is safe: the
 /// companion layout is definitionally the EXE-style shell (the runtime
 /// loader maps the companion and runs the standard shell unpack), so the DLL
-/// pipeline probe can never be right for it — and probing is not a no-op on
-/// targets without unwinding (wasm), where the probe's caught panic becomes
-/// a fatal trap. Output bytes are identical to the dll-first + exe-fallback
-/// route for every input that route handles.
+/// pipeline probe can never be right for it. Output bytes are identical to the
+/// DLL-first + EXE-fallback route for every input that route handles.
 fn unpack_spliced_or_auto(
    bytes: &[u8],
    spliced: bool,
@@ -880,10 +875,9 @@ pub struct UnpackedImage {
 /// Unpack in-memory `input` bytes, optionally paired with an external
 /// companion payload `companion` (the `<input>._` file's contents).
 ///
-/// This is the I/O-free counterpart of [`unpack_one_v`], used by the
-/// WebAssembly build: splice (when the companion's first 32 bytes match the
-/// stub header), unpack, overlay the export table and TLS directory from the
-/// stub, then run the static integrity check.
+/// This is the in-memory counterpart of [`unpack_one_v`]: splice a matching
+/// companion, unpack, overlay the export table and TLS directory from the stub,
+/// then run the static integrity check.
 pub fn unpack_bytes(
    input: &[u8],
    companion: Option<&[u8]>,
@@ -960,7 +954,7 @@ pub fn unpack_one_v(
 /// into a sparse, original-metadata-style value; il2cpp expects the contiguous
 /// per-module index it indexes its codegen tables with, so a statically-unpacked
 /// il2cpp game assembly reads garbage and crashes during init. This rewrites
-/// the tokens back to their canonical form (see [`crate::metadata::deobfuscate`]).
+/// the tokens back to their canonical form (see [`senbei_metadata::deobfuscate`]).
 ///
 /// The output is written only when something actually changed
 /// (`report.remapped > 0`); an already-clean metadata is left untouched and no
@@ -970,11 +964,11 @@ pub fn deobfuscate_metadata_to(
    input: &Path,
    dest: &Path,
    verbose: bool,
-) -> anyhow::Result<crate::metadata::Report> {
+) -> anyhow::Result<senbei_metadata::Report> {
    let data = std::fs::read(input)?;
    // Preserve the metadata::Error in the chain (rather than stringifying it)
    // so the folder driver can apply its unsupported-version policy.
-    let (out, report) = crate::metadata::deobfuscate(&data)
+    let (out, report) = senbei_metadata::deobfuscate(&data)
        .map_err(|e| anyhow::Error::new(e).context(format!("{input:?}")))?;
    if report.remapped > 0 {
        if let Some(parent) = dest.parent() {
@@ -1,7 +1,7 @@
+//! Filesystem and command-line orchestration.
+
 pub mod job;
 pub mod logfile;
-pub mod metadata;
 pub mod pause;
 pub mod scan;
 pub mod ui;
-pub mod unpacker;
@@ -1,4 +1,4 @@
-use crate::unpacker::detect;
+use senbei_pe::detect;
 use std::io::Read;
 use std::path::{Path, PathBuf};
 use walkdir::WalkDir;
@@ -16,24 +16,23 @@ const DETECT_PREFIX: u64 = 8 * 1024;
 /// Smallest file that can possibly be a target, so anything shorter is skipped
 /// without ever being opened.
 ///
-/// A Crackproof module needs ≥ 4128 bytes for [`crate::unpacker::detect`]'s key
+/// A Crackproof module needs ≥ 4128 bytes for [`senbei_pe::detect`]'s key
 /// table (it reads the dword at 4124), so the bound is exact for the unpack
 /// path. An il2cpp `global-metadata.dat` only needs 4 bytes to match its magic,
 /// but its header alone runs to offset 0xB0 and the images/types/methods tables
 /// it indexes make every real one megabytes long — a sub-4 KiB "metadata" could
-/// only ever fail [`crate::metadata::deobfuscate`] with `Malformed`, so nothing
+/// only ever fail [`senbei_metadata::deobfuscate`] with `Malformed`, so nothing
 /// processable is lost.
 const MIN_SIZE: u64 = 4128;

 /// File extensions that are bulk data by construction and can never be a PE
 /// image or an il2cpp metadata blob.
 ///
-/// This is deliberately a **deny**-list, not an allow-list: the default is to
-/// probe, so anything unrecognised is still opened. Targets are recognised by
-/// content, not extension, and can carry arbitrary names — there is no closed
-/// set of target extensions an allow-list of `exe`/`dll` could enumerate.
-/// Only extensions that are bulk asset or text formats by construction appear
-/// here.
+/// This is deliberately a **deny**-list, not an executable allow-list: unknown
+/// extensions are still probed. Extensionless files are handled separately by
+/// [`denied_name`] because asset stores commonly contain tens of thousands of
+/// extensionless chunks; exhaustive probing remains available through
+/// `--scan-all`.
 ///
 /// Set `SENBEI_SCAN_ALL=1` (or pass `--scan-all`) to probe every file regardless.
 const DENY_EXT: &[&str] = &[
@@ -90,11 +89,12 @@ const DENY_EXT: &[&str] = &[
    "sr",
 ];

-/// Whether `path`'s extension is on [`DENY_EXT`]. Extensionless files are never
-/// denied (they could be anything).
-fn denied_ext(path: &Path) -> bool {
+/// Whether `path` can be skipped from its name alone. Extensionless files and
+/// files whose extension is on [`DENY_EXT`] are not opened during a default
+/// scan. `--scan-all` remains available when exhaustive probing is required.
+fn denied_name(path: &Path) -> bool {
    let Some(ext) = path.extension() else {
-        return false;
+        return true;
    };
    let Some(ext) = ext.to_str() else {
        return false;
@@ -206,12 +206,17 @@ pub fn find_targets_opts(root: &Path, scan_all: bool) -> (Vec<PathBuf>, Vec<Path
            continue;
        }
        if !scan_all {
+            // Name checks come first so extensionless asset chunks never
+            // trigger even an explicit metadata query.
+            if denied_name(entry.path()) {
+                continue;
+            }
            // Skip on directory metadata alone — never open these.
            let too_small = entry
                .metadata()
                .map(|m| m.len() < MIN_SIZE)
                .unwrap_or(false);
-            if too_small || denied_ext(entry.path()) {
+            if too_small {
                continue;
            }
        }
@@ -223,7 +228,7 @@ pub fn find_targets_opts(root: &Path, scan_all: bool) -> (Vec<PathBuf>, Vec<Path
    // `Some(Class::None)` means "probed, matched neither detector".
    let n = paths.len();
    let mut class: Vec<Option<Class>> = vec![Some(Class::None); n];
-    let workers = crate::unpacker::parallel::thread_cap().clamp(1, n.max(1));
+    let workers = senbei_pe::thread_cap().clamp(1, n.max(1));
    if workers <= 1 {
        for (p, c) in paths.iter().zip(class.iter_mut()) {
            *c = classify(p);
@@ -304,7 +309,7 @@ fn classify(path: &Path) -> Option<Class> {
    let r = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
        if detect(&head).is_some() {
            Class::Crackproof
-        } else if crate::metadata::is_metadata(&head) {
+        } else if senbei_metadata::is_metadata(&head) {
            Class::Metadata
        } else {
            Class::None
@@ -329,29 +334,49 @@ mod tests {
    #[test]
    fn denies_bulk_asset_extensions_case_insensitively() {
        for p in ["a.ab", "a.XML", "a.Acb", "a.ma2", "a.manifest", "a.PNG"] {
-            assert!(denied_ext(Path::new(p)), "{p} should be denied");
+            assert!(denied_name(Path::new(p)), "{p} should be denied");
+        }
+    }
+
+    #[test]
+    fn denies_extensionless_files() {
+        for p in ["asset", "level0", "0123456789abcdef"] {
+            assert!(denied_name(Path::new(p)), "{p} should be denied");
        }
    }

    #[test]
    fn never_denies_what_a_target_can_be_named() {
-        // Targets are recognised by content, not name — a protected module
-        // can carry any extension, or none — so names like these must always
-        // be probed. An allow-list would have skipped them.
+        // Unknown extensions must still be probed. This keeps the filter a
+        // narrow deny-list rather than an executable-extension allow-list.
        for p in [
            "app.exe.bak",
            "managed.dll.bak",
            "daemon.exe",
            "GameLib.dll",
            "global-metadata.dat",
-            "noextension",
            "a.so",
            "a.bin",
        ] {
-            assert!(!denied_ext(Path::new(p)), "{p} must still be probed");
+            assert!(!denied_name(Path::new(p)), "{p} must still be probed");
        }
    }

+    #[test]
+    fn extensionless_targets_require_exhaustive_scan() {
+        let td = tempfile::tempdir().unwrap();
+        let root = td.path();
+        let mut blob = vec![0u8; MIN_SIZE as usize + 1];
+        blob[..4].copy_from_slice(&0xFAB1_1BAFu32.to_le_bytes());
+        std::fs::write(root.join("metadata"), &blob).unwrap();
+
+        let (_, filtered, _) = find_targets_opts(root, false);
+        assert!(filtered.is_empty());
+
+        let (_, exhaustive, _) = find_targets_opts(root, true);
+        assert_eq!(exhaustive.len(), 1);
+    }
+
    /// A file below the Crackproof key-table bound is skipped without being
    /// opened, but a large non-asset file is still probed.
    #[test]
@@ -1,6 +1,6 @@
-use crate::unpacker::{IntegrityReport, Kind};
 use indicatif::{ProgressBar, ProgressStyle};
 use owo_colors::OwoColorize;
+use senbei_pe::{IntegrityReport, Kind};
 use std::path::Path;

 /// Create a progress bar for `n` items. Hidden when `quiet` is true.
@@ -0,0 +1,6 @@
+[package]
+name = "senbei-metadata"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+description = "Unity il2cpp metadata de-obfuscation for Senbei"
@@ -0,0 +1,5 @@
+//! Unity il2cpp metadata de-obfuscation.
+
+mod metadata;
+
+pub use metadata::*;
@@ -0,0 +1,10 @@
+[package]
+name = "senbei-pe"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+description = "PE detection, unpacking, and validation for Senbei"
+
+[dependencies]
+senbei-crypto.workspace = true
+thiserror.workspace = true
@@ -0,0 +1,3 @@
+mod pipeline;
+
+pub use pipeline::*;
@@ -13,9 +13,12 @@
 //!   CalculateChecksumWithSizeXor        -> primitives::calculate_checksum
 //!   CalculateCrc32                      -> crc32::compute (via above)

-use super::UnpackError;
-use super::bytecode::{Op, OpsLut, generate};
-use super::primitives::{self, *};
+use super::super::{
+    BufferOperation, BytecodeStage, DecompressionStage, DescriptorTable, SectionPipeline,
+    UnpackError,
+};
+use senbei_crypto::bytecode::{Op, OpsLut, generate};
+use senbei_crypto::primitives::{self, *};

 /// Read a signed 32-bit little-endian value.
 fn get_i32(d: &[u8], offset: i32) -> i32 {
@@ -68,6 +71,7 @@ fn decrypt_data4(
    key: i32,
    decomp_params: &[i32; 4],
    transform: Option<&[Op]>,
+    stage: DecompressionStage,
 ) -> Result<(), UnpackError> {
    let addr = get_i32(d, offset);
    let size = get_i32(d, offset + 4);
@@ -84,19 +88,17 @@ fn decrypt_data4(
        OpsLut::new(ops).map_region(d, addr as usize, size as usize);
    }

-    if size != decompressed_size {
-        // decompress reports corruption (after partial writes) via its bool;
-        // surface it instead of shipping a garbage block.
-        if !decompress(
+    if size != decompressed_size
+        && let Err(reason) = primitives::decompress_detailed(
            d,
            addr as u32,
            compressed_addr as u32,
            decomp_params[1] as u32,
            size as u32,
            decompressed_size as u32,
-        ) {
-            return Err(UnpackError::DecompressFailed);
-        }
+        )
+    {
+        return Err(UnpackError::StageDecompressionFailed { stage, reason });
    }
    Ok(())
 }
@@ -218,7 +220,11 @@ fn decrypt_and_decompress_data(
        // Guard: need 16 bytes at section_data_offset in `d`
        let off = section_data_offset as usize;
        if off.saturating_add(16) > d.len() {
-            return Err(UnpackError::OutOfBounds(off));
+            return Err(UnpackError::DescriptorOutOfBounds {
+                table: DescriptorTable::DllSectionBlocks,
+                offset: off,
+                image_len: d.len(),
+            });
        }
        decrypt_data6_shift6(d, section_data_offset, 16);
        let dest_offset = get_i32(d, section_data_offset);
@@ -246,10 +252,10 @@ fn decrypt_and_decompress_data(
        let lut = OpsLut::new(decrypt_func);
        let ko0 = decomp_params[0];
        let ko2 = decomp_params[2];
-        let ks_snap =
-            primitives::aes_schedule_snapshot(d, ko2 as u32).ok_or(UnpackError::Corrupt)?;
+        let ks_snap = primitives::aes_schedule_snapshot(d, ko2 as u32)
+            .ok_or(UnpackError::InvalidAesKeySchedule { offset: ko2 as u32 })?;
        let tab_snap = primitives::huffman_table_snapshot(d, ko0 as u32)
-            .ok_or(UnpackError::DecompressFailed)?;
+            .ok_or(UnpackError::InvalidHuffmanTable { offset: ko0 as u32 })?;
        let spans: Vec<(usize, usize)> = blocks
            .iter()
            .map(|b| {
@@ -277,12 +283,15 @@ fn decrypt_and_decompress_data(
                    b.size as u32,
                    b.expected_crc as u32,
                ) {
-                    return Err(UnpackError::DecompressFailed);
+                    return Err(UnpackError::SectionDecompressionFailed {
+                        pipeline: SectionPipeline::Dll,
+                        block: i,
+                    });
                }
            }
            Ok(())
        };
-        super::parallel::parallel_for(d, &spans, 1, do_block)?;
+        super::super::parallel::parallel_for(d, &spans, 1, do_block)?;
    }

    // Zero-fill loop.
@@ -292,7 +301,11 @@ fn decrypt_and_decompress_data(
        // decrypts 16 too, so guard 16 (an 8-byte guard would let
        // decrypt_data6_shift6 index past the end of a truncated descriptor).
        if off.saturating_add(16) > d.len() {
-            return Err(UnpackError::OutOfBounds(off));
+            return Err(UnpackError::DescriptorOutOfBounds {
+                table: DescriptorTable::DllZeroFill,
+                offset: off,
+                image_len: d.len(),
+            });
        }
        decrypt_data6_shift6(d, section_data_offset, 16);
        let zero_offset = get_i32(d, section_data_offset);
@@ -305,7 +318,12 @@ fn decrypt_and_decompress_data(
        for i in 0..zero_size {
            let idx = (zero_offset + i) as usize;
            if idx >= d.len() {
-                return Err(UnpackError::OutOfBounds(idx));
+                return Err(UnpackError::BufferRangeOutOfBounds {
+                    operation: BufferOperation::ZeroFill,
+                    offset: idx,
+                    size: 1,
+                    buffer_len: d.len(),
+                });
            }
            d[idx] = 0;
        }
@@ -324,12 +342,16 @@ pub fn unpack_dll(input: &[u8]) -> Result<Vec<u8>, UnpackError> {
 pub fn unpack_dll_v(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError> {
    // Trap any out-of-bounds panic from a truncated/garbled file and report it
    // as a clean error so the public API stays panic-free.
-    super::catch_unpack(move || unpack_dll_inner(input, verbose))
+    super::super::catch_unpack(move || unpack_dll_inner(input, verbose))
 }

 fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError> {
-    if input.len() < 4096 {
-        return Err(UnpackError::InputTooShort(input.len()));
+    const HEADER_LEN: usize = 4128;
+    if input.len() < HEADER_LEN {
+        return Err(UnpackError::InputTooShort {
+            actual: input.len(),
+            required: HEADER_LEN,
+        });
    }

    // `file_data` and `original_file_data` both borrow the same protected input.
@@ -347,15 +369,18 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        println!("  keys[6] anchor    = 0x{:08X}", keys[6] as u32);
    }

-    if !super::is_supported_magic(keys[1] as u32) {
-        return Err(UnpackError::DllUnpack(
-            "Not a Crackproof protected file (KONN magic mismatch)".into(),
-        ));
+    if !super::super::is_supported_magic(keys[1] as u32) {
+        return Err(UnpackError::HeaderMagicMismatch {
+            found: keys[1] as u32,
+        });
    }

    let pe_offset = get_i32(file_data, 60);
    if pe_offset < 0 || (pe_offset as usize).saturating_add(84) > file_data.len() {
-        return Err(UnpackError::DllUnpack("implausible PE offset".into()));
+        return Err(UnpackError::InvalidPeOffset {
+            offset: i64::from(pe_offset),
+            input_len: file_data.len(),
+        });
    }
    // This pipeline is PE32+-only: its header fixups write the data
    // directories at PE32+ offsets (pe+144..180, pe+136 for the DD blob). On a
@@ -363,14 +388,18 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
    // structurally plausible but unloadable file. Reject early with a clear
    // error so `unpack_auto`'s EXE-pipeline fallback handles PE32 DLLs (that
    // path is PE32-aware — see run_pe32), instead of us mangling them here.
-    if get_i32(file_data, pe_offset + 24) & 0xFFFF != 0x20B {
-        return Err(UnpackError::DllUnpack(
-            "not a PE32+ image (the DLL pipeline handles 64-bit only)".into(),
-        ));
+    let optional_magic = get_u16(file_data, (pe_offset + 24) as u32);
+    if optional_magic != 0x20B {
+        return Err(UnpackError::UnsupportedDllPeMagic {
+            found: optional_magic,
+        });
    }
    let size_of_image = get_i32(file_data, pe_offset + 80);
-    if size_of_image <= 0 || size_of_image as u64 > super::MAX_IMAGE_SIZE {
-        return Err(UnpackError::DllUnpack("implausible SizeOfImage".into()));
+    if size_of_image <= 0 || size_of_image as u64 > super::super::MAX_IMAGE_SIZE {
+        return Err(UnpackError::InvalidImageSize {
+            size: i64::from(size_of_image),
+            max: super::super::MAX_IMAGE_SIZE,
+        });
    }
    let mut out = vec![0u8; size_of_image as usize];
    let base_offset = keys[6] - keys[3] + 0x2000;
@@ -438,6 +467,16 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        println!("  checksum1       = 0x{:08X}", checksum1 as u32);
        println!("  decrypted_addr1 = 0x{:08X}", decrypted_addr1 as u32);
    }
+    let primary_end = decrypted_addr1.checked_add(3856);
+    if decrypted_addr1 < keys[3]
+        || primary_end.is_none_or(|end| end < 0 || end as usize > out.len())
+    {
+        return Err(UnpackError::InvalidDllPrimaryDescriptor {
+            address: decrypted_addr1 as u32,
+            minimum: keys[3] as u32,
+            image_len: out.len(),
+        });
+    }
    let import_offset = get_i32(&out, decrypted_addr1 + 3444);
    let decrypted_addr2_size = get_i32(&out, decrypted_addr1 + 3632);
    decrypt_data3(
@@ -512,6 +551,7 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        table_val ^ checksum2 ^ (xor_accumulator as i32),
        &decomp_params,
        None,
+        DecompressionStage::DllCodeBlock1,
    )?;

    let addr3b = get_i32(&out, decrypted_addr1 + 3728);
@@ -527,7 +567,7 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
    let crc_val = {
        let a = crc_data_addr as usize;
        let n = crc_data_size as usize;
-        super::crc32::compute(&out[a..a + n]) as i32
+        senbei_crypto::crc32::compute(&out[a..a + n]) as i32
    };
    let crc_xored = crc_data_size ^ crc_val;
    let trailing_val = get_i32(&out, crc_data_addr + crc_data_size - 4);
@@ -537,6 +577,7 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        crc_xored ^ (xor_accumulator as i32) ^ trailing_val,
        &decomp_params,
        None,
+        DecompressionStage::DllCodeBlock2,
    )?;

    let checksum3 = calculate_checksum(&out, (decrypted_addr1 + 3480) as u32) as i32;
@@ -549,6 +590,7 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        (not_val ^ (xor_key as u32)) as i32,
        &decomp_params,
        None,
+        DecompressionStage::DllCodeBlock3,
    )?;

    let addr4 = get_i32(&out, addr4_offset);
@@ -574,8 +616,9 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        lfsr_seed_val = lfsr_seed_val.wrapping_add(k);
    }

-    let decrypt_func = generate(&out, lfsr as u32)
-        .ok_or_else(|| UnpackError::DllUnpack("Failed to build decryption expression".into()))?;
+    let decrypt_func = generate(&out, lfsr as u32).ok_or(UnpackError::BytecodeGenerationFailed(
+        BytecodeStage::DllPrimaryDecryptor,
+    ))?;

    let addr5_offset = decrypted_addr1 + 3840;
    let addr5 = get_i32(&out, addr5_offset);
@@ -585,6 +628,7 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
        lfsr_seed_val ^ xor_key ^ checksum4,
        &decomp_params,
        Some(&decrypt_func),
+        DecompressionStage::DllCodeBlock4,
    )?;
    if verbose {
        println!("[7/9] Decrypting code block 4 (addr5)...");
@@ -603,9 +647,9 @@ fn unpack_dll_inner(input: &[u8], verbose: bool) -> Result<Vec<u8>, UnpackError>
    let lfsr2 = metadata_offset + 88;
    decrypt_data6(&mut out, lfsr2 as u32);

-    let decrypt_func2 = generate(&out, lfsr2 as u32).ok_or_else(|| {
-        UnpackError::DllUnpack("Failed to build second decryption expression".into())
-    })?;
+    let decrypt_func2 = generate(&out, lfsr2 as u32).ok_or(
+        UnpackError::BytecodeGenerationFailed(BytecodeStage::DllSectionDecryptor),
+    )?;

    let section_image_base = 4095 - get_i32(original_file_data, 4224);
    let section_data_offset = get_i32(&out, addr5 + 11976);
@@ -0,0 +1,253 @@
+pub use senbei_crypto::{BufferOperation, DecompressionFailure};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum DecompressionStage {
+    ExeStage3,
+    ExeStage3Secondary,
+    ExeStage4,
+    ExeStage5,
+    Pe32FourthStage,
+    Pe32FifthStage,
+    Pe32SeventhStage,
+    DllCodeBlock1,
+    DllCodeBlock2,
+    DllCodeBlock3,
+    DllCodeBlock4,
+}
+
+impl std::fmt::Display for DecompressionStage {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::ExeStage3 => "EXE stage3",
+            Self::ExeStage3Secondary => "EXE secondary stage3",
+            Self::ExeStage4 => "EXE stage4",
+            Self::ExeStage5 => "EXE stage5",
+            Self::Pe32FourthStage => "PE32 fourth stage",
+            Self::Pe32FifthStage => "PE32 fifth stage",
+            Self::Pe32SeventhStage => "PE32 seventh stage",
+            Self::DllCodeBlock1 => "DLL code block 1",
+            Self::DllCodeBlock2 => "DLL code block 2",
+            Self::DllCodeBlock3 => "DLL code block 3",
+            Self::DllCodeBlock4 => "DLL code block 4",
+        })
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum BytecodeStage {
+    ExeStage4,
+    ExeStage5,
+    Pe32CustomDecryptor,
+    Pe32FileDecryptor,
+    DllPrimaryDecryptor,
+    DllSectionDecryptor,
+}
+
+impl std::fmt::Display for BytecodeStage {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::ExeStage4 => "EXE stage4",
+            Self::ExeStage5 => "EXE stage5",
+            Self::Pe32CustomDecryptor => "PE32 custom decryptor",
+            Self::Pe32FileDecryptor => "PE32 file decryptor",
+            Self::DllPrimaryDecryptor => "DLL primary decryptor",
+            Self::DllSectionDecryptor => "DLL section decryptor",
+        })
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum SectionPipeline {
+    ExePe32Plus,
+    ExePe32,
+    Dll,
+}
+
+impl std::fmt::Display for SectionPipeline {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::ExePe32Plus => "PE32+ EXE",
+            Self::ExePe32 => "PE32 EXE",
+            Self::Dll => "DLL",
+        })
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum DescriptorTable {
+    DllSectionBlocks,
+    DllZeroFill,
+}
+
+impl std::fmt::Display for DescriptorTable {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::DllSectionBlocks => "DLL section-block",
+            Self::DllZeroFill => "DLL zero-fill",
+        })
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+#[non_exhaustive]
+pub enum UnpackError {
+    #[error("input too short (need at least {required} bytes, got {actual})")]
+    InputTooShort { actual: usize, required: usize },
+
+    #[error("decrypted header magic mismatch (got 0x{found:08X})")]
+    HeaderMagicMismatch { found: u32 },
+
+    #[error("anchor field not found — corrupt data or wrong offset")]
+    AnchorNotFound,
+
+    #[error("stage1 descriptor not found near anchor 0x{anchor:08X}")]
+    Stage1DescriptorNotFound { anchor: u32 },
+
+    #[error("stage2 field not found — corrupt data or wrong offset")]
+    Stage2NotFound,
+
+    #[error("chk_src_start not found — corrupt data or wrong offset")]
+    ChkSrcStartNotFound,
+
+    #[error("table_start not found — corrupt data or wrong offset")]
+    TableStartNotFound,
+
+    #[error("{0} bytecode generation failed — corrupt data or wrong offset")]
+    BytecodeGenerationFailed(BytecodeStage),
+
+    #[error("stage5 marker not found — this build's layout is not supported by this unpacker")]
+    Stage5MarkerNotFound,
+
+    #[error("not a Crackproof-protected file")]
+    NotCrackproof,
+
+    #[error("invalid PE header offset {offset} for {input_len}-byte input")]
+    InvalidPeOffset { offset: i64, input_len: usize },
+
+    #[error("DLL pipeline requires PE32+ optional-header magic, got 0x{found:04X}")]
+    UnsupportedDllPeMagic { found: u16 },
+
+    #[error(
+        "DLL primary descriptor address 0x{address:08X} is below layout base 0x{minimum:08X} or outside {image_len}-byte image"
+    )]
+    InvalidDllPrimaryDescriptor {
+        address: u32,
+        minimum: u32,
+        image_len: usize,
+    },
+
+    #[error("invalid SizeOfImage {size}; expected 1..={max}")]
+    InvalidImageSize { size: i64, max: u64 },
+
+    #[error(
+        "{operation} range out of bounds (offset {offset}, size {size}, buffer length {buffer_len})"
+    )]
+    BufferRangeOutOfBounds {
+        operation: BufferOperation,
+        offset: usize,
+        size: usize,
+        buffer_len: usize,
+    },
+
+    #[error(
+        "EXE checksum descriptor at 0x{descriptor:08X} points outside input (offset {offset}, size {size}, input length {image_len})"
+    )]
+    ExeChecksumRangeOutOfBounds {
+        descriptor: u32,
+        offset: usize,
+        size: usize,
+        image_len: usize,
+    },
+
+    #[error(
+        "{table} descriptor out of bounds (offset {offset}, size 16, image length {image_len})"
+    )]
+    DescriptorOutOfBounds {
+        table: DescriptorTable,
+        offset: usize,
+        image_len: usize,
+    },
+
+    #[error("PE32 tbl not found — corrupt data or wrong offset")]
+    Pe32TblNotFound,
+
+    #[error("PE32 thirdStage decrypt failed — corrupt data or wrong offset")]
+    Pe32ThirdStageFailed,
+
+    #[error("PE32 customDecryptor not found in sevenStage")]
+    Pe32CustomDecryptorNotFound,
+
+    #[error("PE32 eighthStageKey not found")]
+    Pe32EighthKeyNotFound,
+
+    #[error("PE32 file LFSR not found in eighthStage")]
+    Pe32FileLfsrNotFound,
+
+    #[error("{stage} decompression failed: {reason}")]
+    StageDecompressionFailed {
+        stage: DecompressionStage,
+        reason: DecompressionFailure,
+    },
+
+    #[error("{pipeline} section block {block} decompression failed")]
+    SectionDecompressionFailed {
+        pipeline: SectionPipeline,
+        block: usize,
+    },
+
+    #[error("AES key schedule is outside the image at offset {offset}")]
+    InvalidAesKeySchedule { offset: u32 },
+
+    #[error("Huffman table is outside the image at offset {offset}")]
+    InvalidHuffmanTable { offset: u32 },
+
+    #[error("DLL pipeline failed: {dll}; EXE fallback failed: {exe}")]
+    PipelineFallbackFailed {
+        dll: Box<UnpackError>,
+        exe: Box<UnpackError>,
+    },
+
+    #[error(
+        "PE32 second-stage range is invalid (offset {offset}, size {size}, image length {image_len})"
+    )]
+    Pe32SecondStageRangeInvalid {
+        offset: u32,
+        size: u32,
+        image_len: usize,
+    },
+
+    #[error("PE32 relocation-data descriptor not found")]
+    Pe32RelocationDataNotFound,
+
+    #[error("file decryptor candidate failed structural validation")]
+    FileDecryptorValidationFailed,
+
+    #[error("PE32 memory image could not be rebuilt as a file-layout PE")]
+    Pe32OutputLayoutInvalid,
+
+    #[error("internal panic at {file}:{line}:{column}: {message}")]
+    InternalPanic {
+        message: String,
+        file: String,
+        line: u32,
+        column: u32,
+    },
+}
+
+impl From<senbei_crypto::Error> for UnpackError {
+    fn from(error: senbei_crypto::Error) -> Self {
+        match error {
+            senbei_crypto::Error::BufferRangeOutOfBounds {
+                operation,
+                offset,
+                size,
+                buffer_len,
+            } => Self::BufferRangeOutOfBounds {
+                operation,
+                offset,
+                size,
+                buffer_len,
+            },
+        }
+    }
+}
@@ -0,0 +1,3 @@
+mod pipeline;
+
+pub use pipeline::*;
@@ -77,6 +77,49 @@ fn rva_to_off(secs: &[Section], file_len: usize, rva: u32, need: u32) -> Option<
    None
 }

+fn is_executable_rva(secs: &[Section], rva: u32) -> bool {
+    secs.iter().any(|section| {
+        let span = section.vsize.max(section.raw_size);
+        rva >= section.va
+            && rva < section.va.wrapping_add(span)
+            && (section.chars & 0x2000_0000) != 0
+    })
+}
+
+fn check_common_entry_branches(
+    stub: &[u8],
+    ep: u32,
+    secs: &[Section],
+    report: &mut IntegrityReport,
+) {
+    if stub.len() < 18
+        || stub[0..3] != [0x48, 0x83, 0xEC]
+        || stub[4] != 0xE8
+        || stub[9..12] != [0x48, 0x83, 0xC4]
+        || stub[12] != stub[3]
+        || stub[13] != 0xE9
+    {
+        return;
+    }
+    for (name, rel_off, instruction_len) in [("call", 5usize, 9i64), ("jump", 14usize, 18i64)] {
+        let rel = i32::from_le_bytes([
+            stub[rel_off],
+            stub[rel_off + 1],
+            stub[rel_off + 2],
+            stub[rel_off + 3],
+        ]) as i64;
+        let target = i64::from(ep) + instruction_len + rel;
+        let valid = u32::try_from(target)
+            .ok()
+            .is_some_and(|rva| is_executable_rva(secs, rva));
+        if !valid {
+            report.issues.push(format!(
+                "entry point {name} target 0x{target:X} is outside executable sections (DD8 selection is likely wrong)"
+            ));
+        }
+    }
+}
+
 /// Inspect an unpacked PE image and report any defect that would make the OS
 /// loader fault at runtime. `out` is the bytes the unpacker produced.
 pub fn check(out: &[u8]) -> IntegrityReport {
@@ -253,6 +296,9 @@ pub fn check(out: &[u8]) -> IntegrityReport {
                        "entry point RVA 0x{ep:X} is not in an executable section"
                    ));
                }
+                if let Some(entry_stub) = out.get(off as usize..off as usize + 18) {
+                    check_common_entry_branches(entry_stub, ep, &secs, &mut r);
+                }
            }
        }
    }
@@ -363,3 +409,40 @@ fn looks_like_dll_name(d: &[u8], off: u32) -> bool {
    }
    d[start..end].iter().all(|&b| (0x20..0x7F).contains(&b))
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn executable_text() -> Vec<Section> {
+        vec![Section {
+            va: 0x1000,
+            vsize: 0x4000,
+            raw_ptr: 0x1000,
+            raw_size: 0x4000,
+            chars: 0x6000_0020,
+        }]
+    }
+
+    #[test]
+    fn common_entry_stub_rejects_out_of_image_branches() {
+        let stub = [
+            0x48, 0x83, 0xEC, 0x28, 0xE8, 0x5B, 0x02, 0x41, 0x00, 0x48, 0x83, 0xC4, 0x28, 0xE9,
+            0x7A, 0xFE, 0x54, 0xFF,
+        ];
+        let mut report = IntegrityReport::default();
+        check_common_entry_branches(&stub, 0x1264, &executable_text(), &mut report);
+        assert_eq!(report.issues.len(), 2);
+    }
+
+    #[test]
+    fn common_entry_stub_accepts_executable_branches() {
+        let stub = [
+            0x48, 0x83, 0xEC, 0x28, 0xE8, 0x5B, 0x02, 0x00, 0x00, 0x48, 0x83, 0xC4, 0x28, 0xE9,
+            0x7A, 0xFE, 0xFF, 0xFF,
+        ];
+        let mut report = IntegrityReport::default();
+        check_common_entry_branches(&stub, 0x1264, &executable_text(), &mut report);
+        assert!(report.ok());
+    }
+}
@@ -0,0 +1,14 @@
+//! Internal PE layout discovery and image reconstruction.
+
+mod dd8;
+mod discovery;
+mod image;
+
+pub(super) use dd8::{select_dd8_formula_pe32, select_dd8_shift};
+pub(super) use discovery::{
+    discover_eighth_slots, find_bytecode_offset, find_lfsr_block, find_str_pos, find_tbl_pe32,
+    find_v_after_pad, find_v4_offset, get_string_to_null, section_name, trial_decrypt5_u32,
+};
+pub(super) use image::{
+    compact_memory_image_to_pe, move_pe32_imports_to_kmiat, pe32_imports_already_match_idata_layout,
+};
@@ -0,0 +1,609 @@
+//! Validation-driven selection for per-page text transforms.
+
+use super::discovery::trial_decrypt5_u32;
+
+/// PE32 `.text` dd8 key-formula selection with a skip decision. The packer keys
+/// the per-page XOR either with `page+1` or `0x8000*(page+1)`; the formula is
+/// not recorded. Replays the dd8 page pass on a scratch copy of sample pages
+/// (25/50/75% of `.text`) under each formula and counts how many positions
+/// decode to `0xCC` (int3 padding).
+///
+/// Returns `Some(true)` for the `0x8000*(page+1)` formula, `Some(false)` for
+/// `page+1`, or `None` when `.text` must NOT be dd8-decrypted at all. The packer
+/// dd8-encrypts `.text` on EXEs (so unpacking must replay it) but leaves a native
+/// DLL's `.text` plaintext; replaying dd8 there scrambles ~1 byte per 16-byte
+/// block. The decision: dd8 only *restores* int3 padding when `.text` was
+/// genuinely encrypted, so apply it only when the chosen formula's whole-page
+/// 0xCC count rises *clearly* above the no-dd8 baseline; otherwise skip.
+///
+/// "Clearly" matters: dd8 XORs 255 positions per page with pseudo-random bytes,
+/// so on an already-plaintext `.text` it manufactures ~1 spurious `0xCC` per
+/// sampled page for free (255/256 expected). A bare `best > baseline` test is
+/// therefore biased towards *applying* dd8 on exactly the inputs that must skip
+/// it — and a wrongly-applied dd8 is silent: it scrambles ~1 byte per 16 with no
+/// error and nothing downstream (not even `integrity::check`, which only reads
+/// 16 bytes at the entry point) notices. The [`MIN_DD8_NET_GAIN`] floor below is
+/// the PE32 counterpart of the margin+floor `select_dd8_shift` already applies
+/// on PE32+ for the same failure mode.
+pub fn select_dd8_formula_pe32(data: &[u8], text_off: u32, text_size: u32) -> Option<bool> {
+    let num_pages_total = text_size / 0x1000;
+    let mut sample_pages: Vec<u32> = Vec::new();
+    for frac in [0.25f64, 0.5, 0.75] {
+        let pg = (num_pages_total as f64 * frac) as u32;
+        if pg > 0 && pg < num_pages_total {
+            sample_pages.push(pg);
+        }
+    }
+    if sample_pages.is_empty() && num_pages_total > 1 {
+        sample_pages.push(num_pages_total / 2);
+    }
+    let score = |big: bool| -> i64 {
+        let mut total = 0i64;
+        for &sp in &sample_pages {
+            let pg_off = (text_off + sp * 0x1000) as usize;
+            if pg_off + 0x1000 > data.len() {
+                continue;
+            }
+            let mut buf = [0u8; 0x1000];
+            buf.copy_from_slice(&data[pg_off..pg_off + 0x1000]);
+            let pk = if big {
+                0x8000u32.wrapping_mul(sp.wrapping_add(1))
+            } else {
+                sp.wrapping_add(1)
+            };
+            let mut k = pk;
+            let rk = k.rotate_right(15);
+            k = rk;
+            for bi in 1..256u32 {
+                let rk = k.rotate_right(15);
+                let ri = rk.wrapping_add(bi);
+                k = ri.wrapping_add(bi);
+                let tidx = (bi.wrapping_mul(16).wrapping_add(ri & 0xF)) as usize;
+                if tidx < buf.len() {
+                    buf[tidx] ^= k as u8;
+                }
+            }
+            total += buf.iter().filter(|&&b| b == 0xCC).count() as i64;
+        }
+        total
+    };
+    let s_small = score(false);
+    let s_big = score(true);
+    // Baseline: whole-page 0xCC over the same sample pages with NO dd8. dd8 only
+    // rewrites 255 bytes per page, so comparing the chosen formula's whole-page
+    // 0xCC against this baseline reveals whether dd8 *restores* int3 padding
+    // (count rises -> .text was packer-encrypted, apply) or merely scrambles
+    // already-plaintext code (count falls -> native-DLL .text left intact, skip).
+    let mut baseline: i64 = 0;
+    for &sp in &sample_pages {
+        let pg_off = (text_off + sp * 0x1000) as usize;
+        if pg_off + 0x1000 > data.len() {
+            continue;
+        }
+        baseline += data[pg_off..pg_off + 0x1000]
+            .iter()
+            .filter(|&&b| b == 0xCC)
+            .count() as i64;
+    }
+    let big = s_big > s_small;
+    let best = s_small.max(s_big);
+    // Minimum net 0xCC gain over the baseline before dd8 is applied. Noise on an
+    // already-plaintext `.text` is ~1 manufactured 0xCC per sampled page (3 pages
+    // -> ~3); every corpus build that genuinely needs dd8 gains +154 or more
+    // (observed +154 and +312), and the one native DLL that must skip scores -18.
+    // A floor of 32 sits ~10x above the noise and ~5x below the smallest true
+    // positive, so it changes no existing decision.
+    const MIN_DD8_NET_GAIN: i64 = 32;
+    let apply = best.saturating_sub(baseline) >= MIN_DD8_NET_GAIN;
+    if std::env::var("SEL_DIAG").is_ok() {
+        eprintln!(
+            "SEL pe32 dd8 s_small={} s_big={} baseline={} gain={} big={} apply={}",
+            s_small,
+            s_big,
+            baseline,
+            best - baseline,
+            big,
+            apply
+        );
+    }
+    // When no interior pages could be sampled (tiny .text) we cannot measure the
+    // effect; preserve the historical behavior of applying dd8.
+    if sample_pages.is_empty() || apply {
+        Some(big)
+    } else {
+        None
+    }
+}
+
+// ---------------------------------------------------------------------------
+// dd8 page-XOR shift selection.
+//
+// The packer scrambles ~1 byte per 16-byte block of .text via decrypt_data8,
+// keyed by `page_idx << shift` (absolute page index = text_va >> 12). Observed
+// shifts are 0 and 15. The shift is NOT stored in any header/config field, so
+// the decision must be validated against the resulting .text content.
+//
+// A recognised CRT entry stub is the strongest oracle: decode skip/0/15 and
+// require both of its direct rel32 branches to land in executable .text. This
+// includes the call/jump displacement bytes themselves; an older entry oracle
+// wildcarded those bytes and could accept a stub whose opcodes looked right but
+// whose branch targets were outside the image.
+//
+// Other entry shapes fall back to padding statistics over a few sample pages
+// (head/tail margin skipped: entry/exit regions have atypical padding density).
+// The primary signal is a *structural* fingerprint: the MSVC function-end
+// padding pattern, a 0xC3 RET opcode followed by a run of >= 4 0xCC int3 bytes.
+// dd8 XORs one pseudo-random byte per 16-byte block, so an already-plaintext
+// page keeps its padding runs only under "no dd8", while a packer-encrypted
+// page restores them only under the correct shift — a wrong candidate destroys
+// every run it touches and essentially never manufactures a RET followed by a
+// long int3 run by chance. This separates the states far more cleanly than a
+// bare 0xCC count, which a wrong candidate inflates for free (~255 coincidences
+// per page at p=1/256).
+//
+// When no candidate produces any RET-anchored padding (sampled pages with
+// dense code and no padded epilogues), the fingerprint is silent, so the
+// decision falls back to the older mutated-position 0xCC count. Both signals
+// use the same decision rule: a candidate must beat the no-dd8 baseline by a
+// clear margin AND an absolute floor, otherwise dd8 is skipped — a wrongly
+// applied dd8 scrambles ~1 byte per 16 with no error surfaced downstream.
+// ---------------------------------------------------------------------------
+pub fn select_dd8_shift(data: &[u8], text_va: u32, text_size: u32, info3: u32) -> u32 {
+    if let Some((shift, scores)) = select_dd8_by_entry_stub(data, text_va, text_size, info3) {
+        if std::env::var("SEL_DIAG").is_ok() {
+            eprintln!(
+                "SEL dd8 entry best_shift={} none={} s0={} s15={}",
+                shift, scores[0], scores[1], scores[2]
+            );
+        }
+        return shift;
+    }
+    let num_pages_total = text_size >> 12;
+    // Fewer than two pages: nothing meaningful to sample; preserve the
+    // historical behavior (shift 0 — the dd8 loop is empty or single-page).
+    if num_pages_total < 2 {
+        return 0;
+    }
+    let text_off = text_va as usize;
+
+    // Sample up to 4 pages, skipping a head/tail margin (entry/exit regions
+    // have atypical padding density). Small .text: sample every page.
+    let mut sample_pages: Vec<u32> = Vec::new();
+    if num_pages_total <= 4 {
+        sample_pages.extend(0..num_pages_total);
+    } else {
+        let margin = (num_pages_total / 8).max(1);
+        let lo = margin;
+        let hi = num_pages_total - margin;
+        if hi <= lo {
+            sample_pages.extend(0..num_pages_total);
+        } else {
+            let step = ((hi - lo) / 4).max(1);
+            let mut i = 0;
+            while i < 4 {
+                let p = lo + i * step;
+                if p < num_pages_total {
+                    sample_pages.push(p);
+                }
+                i += 1;
+            }
+        }
+    }
+    if sample_pages.is_empty() {
+        return 0;
+    }
+
+    let abs_base = text_va >> 12;
+    // Require a clear 2x margin over the already-plaintext baseline AND an
+    // absolute floor. The 2x test alone trips on noise when the counts are
+    // tiny: an external-companion DLL whose .text is already plaintext scores
+    // s15=4 vs none=1 — a spurious 4x — and gets dd8 wrongly applied,
+    // corrupting ~1 byte per 16. The floor rejects that noise while sitting
+    // far below every genuinely-encrypted build's score.
+    const MIN_DD8_HITS: u32 = 8;
+    let margin_pick = |none: u32, s0: u32, s15: u32| -> u32 {
+        let mut best_score = none;
+        let mut best_shift = 99u32; // 99 == skip dd8
+        for (shift, hits) in [(0u32, s0), (15u32, s15)] {
+            if hits > best_score {
+                best_score = hits;
+                best_shift = shift;
+            }
+        }
+        if best_shift != 99 && (best_score < none * 2 || best_score < MIN_DD8_HITS) {
+            best_shift = 99;
+        }
+        best_shift
+    };
+
+    // Primary: RET+int3 padding fingerprint. The fingerprint is diluted across
+    // the whole page (dd8 touches only 255 of 4096 bytes, so even an encrypted
+    // page keeps most of its padding runs), so instead of the fallback's 2x
+    // margin the gate is a *positive delta* over the no-dd8 baseline: on an
+    // already-plaintext .text each wrong shift destroys runs (scores below the
+    // baseline), while the correct shift on an encrypted page restores them
+    // (scores above it). The floor on the delta rejects noise-level gains.
+    let r_none = fingerprint_score(data, text_off, abs_base, &sample_pages, None);
+    let r0 = fingerprint_score(data, text_off, abs_base, &sample_pages, Some(0));
+    let r15 = fingerprint_score(data, text_off, abs_base, &sample_pages, Some(15));
+    // Fallback: mutated-position 0xCC count, for pages whose code has no
+    // RET-anchored padding at all (the fingerprint is silent there).
+    let (none_hits, s0, s15);
+    let best_shift = if r_none != 0 || r0 != 0 || r15 != 0 {
+        none_hits = 0;
+        s0 = 0;
+        s15 = 0;
+        let mut best_score = r_none;
+        let mut shift = 99u32;
+        for (s, score) in [(0u32, r0), (15u32, r15)] {
+            if score > best_score {
+                best_score = score;
+                shift = s;
+            }
+        }
+        if shift != 99 && best_score.saturating_sub(r_none) < MIN_DD8_HITS {
+            shift = 99;
+        }
+        shift
+    } else {
+        none_hits = score_dd8_baseline(data, text_off, &sample_pages);
+        s0 = score_dd8_shift(data, text_off, text_va, &sample_pages, 0);
+        s15 = score_dd8_shift(data, text_off, text_va, &sample_pages, 15);
+        margin_pick(none_hits, s0, s15)
+    };
+    if std::env::var("SEL_DIAG").is_ok() {
+        eprintln!(
+            "SEL dd8 best_shift={} fp=({},{},{}) cc=({},{},{}) samples={:?}",
+            best_shift, r_none, r0, r15, none_hits, s0, s15, sample_pages
+        );
+    }
+    best_shift
+}
+
+/// Minimum 0xCC run length after a RET for the run to count as MSVC
+/// function-end padding.
+const MIN_CC_RUN: u32 = 4;
+
+/// Total length of MSVC function-end padding runs in a page: each 0xC3 byte
+/// followed by >= [`MIN_CC_RUN`] 0xCC bytes contributes the run length.
+fn ret_int3_score(page: &[u8]) -> u32 {
+    let mut total = 0u32;
+    let mut i = 0;
+    while i < page.len() {
+        if page[i] == 0xC3 {
+            let mut j = i + 1;
+            while j < page.len() && page[j] == 0xCC {
+                j += 1;
+            }
+            let run = (j - i - 1) as u32;
+            if run >= MIN_CC_RUN {
+                total += run;
+            }
+            i = j;
+        } else {
+            i += 1;
+        }
+    }
+    total
+}
+
+/// Replay the dd8 page-XOR in place on one sample page.
+fn dd8_apply(buf: &mut [u8; 0x1000], abs_page: u32, shift: u32) {
+    let mut key = abs_page << shift;
+    for bi in 0..256u32 {
+        let mixed = key.rotate_right(15).wrapping_add(bi);
+        key = mixed.wrapping_add(bi);
+        // The packer's dd8 loop does not XOR block i=0 (see decrypt_data8).
+        if bi == 0 {
+            continue;
+        }
+        let tidx = (bi.wrapping_mul(16).wrapping_add(mixed & 0xF)) as usize;
+        buf[tidx] ^= key as u8;
+    }
+}
+
+/// Sum the RET+int3 fingerprint over the sample pages for one candidate
+/// (`None` = the no-dd8 baseline, page as-is).
+fn fingerprint_score(
+    data: &[u8],
+    text_off: usize,
+    abs_base: u32,
+    sample_pages: &[u32],
+    shift: Option<u32>,
+) -> u32 {
+    let mut total = 0u32;
+    for &sp in sample_pages {
+        let pg_off = text_off + (sp as usize) * 0x1000;
+        if pg_off + 0x1000 > data.len() {
+            continue;
+        }
+        let mut page = [0u8; 0x1000];
+        page.copy_from_slice(&data[pg_off..pg_off + 0x1000]);
+        if let Some(sh) = shift {
+            dd8_apply(&mut page, abs_base.wrapping_add(sp), sh);
+        }
+        total += ret_int3_score(&page);
+    }
+    total
+}
+
+/// Select DD8 from the common CRT entry stub when its direct call and jump
+/// provide a stronger oracle than sparse padding statistics. The candidate is
+/// accepted only when it is the sole one whose two branch targets stay inside
+/// `.text`; unrecognised entry code falls through to the padding selector.
+fn select_dd8_by_entry_stub(
+    data: &[u8],
+    text_va: u32,
+    text_size: u32,
+    info3: u32,
+) -> Option<(u32, [u8; 3])> {
+    for entry in entry_candidates(data, text_va, text_size, info3) {
+        let [Some(none), Some(s0), Some(s15)] = [None, Some(0), Some(15)]
+            .map(|shift| entry_stub_branch_score(data, text_va, text_size, entry, shift))
+        else {
+            continue;
+        };
+        let scores = [none, s0, s15];
+        let best = scores.iter().copied().max()?;
+        if best == 2 && scores.iter().filter(|&&score| score == best).count() == 1 {
+            let index = scores.iter().position(|&score| score == best)?;
+            return Some(([99, 0, 15][index], scores));
+        }
+    }
+    None
+}
+
+fn entry_candidates(data: &[u8], text_va: u32, text_size: u32, info3: u32) -> Vec<u32> {
+    let text_end = text_va.saturating_add(text_size);
+    let mut entries = Vec::with_capacity(3);
+    if let Some(pe) = read_u32(data, 0x3C)
+        && let Some(entry) = pe.checked_add(40).and_then(|offset| read_u32(data, offset))
+        && (text_va..text_end).contains(&entry)
+    {
+        entries.push(entry);
+    }
+    for metadata_off in [32u32, 64] {
+        let Some(end) = info3
+            .checked_add(metadata_off)
+            .and_then(|offset| offset.checked_add(8))
+        else {
+            continue;
+        };
+        if end as usize > data.len() {
+            continue;
+        }
+        let entry = trial_decrypt5_u32(data, info3 + metadata_off);
+        let image_base = trial_decrypt5_u32(data, info3 + metadata_off + 4);
+        if image_base == info3 && (text_va..text_end).contains(&entry) && !entries.contains(&entry)
+        {
+            entries.push(entry);
+        }
+    }
+    entries
+}
+
+fn entry_stub_branch_score(
+    data: &[u8],
+    text_va: u32,
+    text_size: u32,
+    entry: u32,
+    shift: Option<u32>,
+) -> Option<u8> {
+    let text_end = text_va.checked_add(text_size)?;
+    if entry < text_va || entry.checked_add(18)? > text_end {
+        return None;
+    }
+
+    let mut stub = [0u8; 18];
+    for (offset, byte) in stub.iter_mut().enumerate() {
+        *byte = dd8_candidate_byte(data, entry + offset as u32, shift)?;
+    }
+    if stub[0..3] != [0x48, 0x83, 0xEC]
+        || stub[4] != 0xE8
+        || stub[9..12] != [0x48, 0x83, 0xC4]
+        || stub[12] != stub[3]
+        || stub[13] != 0xE9
+    {
+        return None;
+    }
+
+    let call_rel = i32::from_le_bytes(stub[5..9].try_into().ok()?) as i64;
+    let jump_rel = i32::from_le_bytes(stub[14..18].try_into().ok()?) as i64;
+    let call_target = i64::from(entry) + 9 + call_rel;
+    let jump_target = i64::from(entry) + 18 + jump_rel;
+    let in_text = |target: i64| target >= i64::from(text_va) && target < i64::from(text_end);
+    Some(u8::from(in_text(call_target)) + u8::from(in_text(jump_target)))
+}
+
+fn dd8_candidate_byte(data: &[u8], rva: u32, shift: Option<u32>) -> Option<u8> {
+    let mut byte = *data.get(rva as usize)?;
+    let Some(shift) = shift else {
+        return Some(byte);
+    };
+    let page = rva >> 12;
+    let block = (rva & 0xFFF) >> 4;
+    let mut key = page << shift;
+    for index in 0..=block {
+        let mixed = key.rotate_right(15).wrapping_add(index);
+        key = mixed.wrapping_add(index);
+        if index != 0 {
+            let target = (page << 12)
+                .wrapping_add(index << 4)
+                .wrapping_add(mixed & 0xF);
+            if target == rva {
+                byte ^= key as u8;
+            }
+        }
+    }
+    Some(byte)
+}
+
+fn read_u32(data: &[u8], offset: u32) -> Option<u32> {
+    let start = offset as usize;
+    let bytes = data.get(start..start.checked_add(4)?)?;
+    Some(u32::from_le_bytes(bytes.try_into().ok()?))
+}
+
+// Baseline: count int3 pads already present at the first byte of each 16-byte
+// block, i.e. the positions dd8 would target if its in-block offset were 0.
+fn score_dd8_baseline(data: &[u8], text_off: usize, sample_pages: &[u32]) -> u32 {
+    let mut hits = 0u32;
+    for &sp in sample_pages {
+        let pg_off = text_off + (sp as usize) * 0x1000;
+        if pg_off + 0x1000 > data.len() {
+            continue;
+        }
+        for bi in 1..256usize {
+            if data[pg_off + bi * 16] == 0xCC {
+                hits += 1;
+            }
+        }
+    }
+    hits
+}
+
+// Replay decrypt_data8 on each sample page under `shift` and count how many of
+// the 255 mutated positions decode to 0xCC.
+fn score_dd8_shift(
+    data: &[u8],
+    text_off: usize,
+    text_va: u32,
+    sample_pages: &[u32],
+    shift: u32,
+) -> u32 {
+    let abs_base = text_va >> 12;
+    let mut hits = 0u32;
+    for &sp in sample_pages {
+        let pg_off = text_off + (sp as usize) * 0x1000;
+        if pg_off + 0x1000 > data.len() {
+            continue;
+        }
+        let abs_page = abs_base.wrapping_add(sp);
+        let mut key = abs_page << shift;
+        for bi in 0..256u32 {
+            let mixed = key.rotate_right(15).wrapping_add(bi);
+            key = mixed.wrapping_add(bi);
+            if bi == 0 {
+                continue;
+            }
+            let tidx = (bi.wrapping_mul(16).wrapping_add(mixed & 0xF)) as usize;
+            if tidx < 0x1000 {
+                let mutated = data[pg_off + tidx] ^ (key as u8);
+                if mutated == 0xCC {
+                    hits += 1;
+                }
+            }
+        }
+    }
+    hits
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn entry_stub_fixture() -> Vec<u8> {
+        let mut data = vec![0u8; 0x5000];
+        data[0x3C..0x40].copy_from_slice(&0x100u32.to_le_bytes());
+        data[0x128..0x12C].copy_from_slice(&0x1264u32.to_le_bytes());
+        data[0x1264..0x1276].copy_from_slice(&[
+            0x48, 0x83, 0xEC, 0x28, 0xE8, 0x5B, 0x02, 0x00, 0x00, 0x48, 0x83, 0xC4, 0x28, 0xE9,
+            0x7A, 0xFE, 0xFF, 0xFF,
+        ]);
+        data
+    }
+
+    fn apply_dd8_page(data: &mut [u8], page_rva: u32, shift: u32) {
+        let mut key = (page_rva >> 12) << shift;
+        for index in 0..256u32 {
+            let mixed = key.rotate_right(15).wrapping_add(index);
+            key = mixed.wrapping_add(index);
+            if index == 0 {
+                continue;
+            }
+            let target = page_rva.wrapping_add(index << 4).wrapping_add(mixed & 0xF) as usize;
+            data[target] ^= key as u8;
+        }
+    }
+
+    #[test]
+    fn entry_stub_selects_plaintext_and_both_dd8_shifts() {
+        let plain = entry_stub_fixture();
+        assert_eq!(select_dd8_shift(&plain, 0x1000, 0x4000, 0), 99);
+
+        for expected in [0u32, 15] {
+            let mut encrypted = plain.clone();
+            apply_dd8_page(&mut encrypted, 0x1000, expected);
+            assert_eq!(select_dd8_shift(&encrypted, 0x1000, 0x4000, 0), expected);
+        }
+    }
+
+    /// Seed the first `count` dd8-targeted positions of each sampled page with
+    /// the byte that decodes to `0xCC` under the `page+1` formula — i.e. an
+    /// encrypted `.text` whose plaintext is int3 padding. Positions whose key
+    /// byte would make the *ciphertext* itself `0xCC` are skipped so the
+    /// fixture contains no `0xCC` at all and every post-dd8 `0xCC` is a genuine
+    /// gain over a zero baseline.
+    fn seed_dd8_int3(data: &mut [u8], text_off: u32, pages: &[u32], count: u32) {
+        for &sp in pages {
+            let pg_off = (text_off + sp * 0x1000) as usize;
+            let mut k = sp.wrapping_add(1);
+            k = k.rotate_right(15);
+            let mut planted = 0u32;
+            for bi in 1..256u32 {
+                let ri = k.rotate_right(15).wrapping_add(bi);
+                k = ri.wrapping_add(bi);
+                if planted >= count {
+                    continue;
+                }
+                let ct = 0xCCu8 ^ (k as u8);
+                if ct == 0xCC {
+                    continue;
+                }
+                let tidx = (bi.wrapping_mul(16).wrapping_add(ri & 0xF)) as usize;
+                data[pg_off + tidx] = ct;
+                planted += 1;
+            }
+        }
+    }
+
+    /// Review regression: a near-plaintext `.text` must NOT be dd8-decrypted.
+    /// dd8 XORs 255 positions per page with pseudo-random bytes, so it
+    /// manufactures a few `0xCC` for free — under the old bare
+    /// `best > baseline` test any positive gain was enough to "apply" dd8 and
+    /// scramble ~1 byte per 16 of a native DLL's already-plaintext code,
+    /// silently (nothing downstream, including the integrity check, notices).
+    /// Here the gain is real but small; the floor must still reject it.
+    #[test]
+    fn pe32_dd8_skips_text_whose_gain_is_only_noise_sized() {
+        let text_off: u32 = 0x1000;
+        let text_size: u32 = 8 * 0x1000;
+        let mut data = vec![0u8; (text_off + text_size) as usize];
+        seed_dd8_int3(&mut data, text_off, &[2, 4, 6], 5);
+        assert!(
+            !data.contains(&0xCC),
+            "fixture must have a zero 0xCC baseline"
+        );
+        assert_eq!(
+            select_dd8_formula_pe32(&data, text_off, text_size),
+            None,
+            "a gain this small is indistinguishable from dd8's own noise"
+        );
+    }
+
+    /// Control for the above: a `.text` whose dd8 pass restores a large amount
+    /// of int3 padding clears the floor and is decrypted. Same fixture shape,
+    /// only the amount of restored padding differs.
+    #[test]
+    fn pe32_dd8_applies_when_padding_is_restored() {
+        let text_off: u32 = 0x1000;
+        let text_size: u32 = 8 * 0x1000;
+        let mut data = vec![0u8; (text_off + text_size) as usize];
+        seed_dd8_int3(&mut data, text_off, &[2, 4, 6], 255);
+        assert_eq!(
+            select_dd8_formula_pe32(&data, text_off, text_size),
+            Some(false),
+            "encrypted .text must be decrypted with the page+1 formula"
+        );
+    }
+}
@@ -0,0 +1,507 @@
+//! Structural locators for protected PE stages.
+
+use senbei_crypto::primitives::{get_u32, lfsr_keystream};
+
+/// Find the 4-byte v_val that follows the LAST occurrence of `48 EB 01 B9`
+/// (REX.W jmp+1; mov ecx,imm32) plus any 0xCC padding. Used to locate
+/// stage4's accum2 seed. Works across builds even when API-name anchors are
+/// absent.
+pub fn find_v_after_pad(data: &[u8], base: u32, len: u32) -> Option<u32> {
+    let start = base as usize;
+    let end = (base.saturating_add(len)) as usize;
+    if end > data.len() {
+        return None;
+    }
+    let sig = [0x48u8, 0xEB, 0x01, 0xB9];
+    let slice = &data[start..end];
+    // last occurrence
+    let mut last = None;
+    let mut i = 0usize;
+    while i + sig.len() <= slice.len() {
+        if slice[i..i + sig.len()] == sig {
+            last = Some(i);
+        }
+        i += 1;
+    }
+    let pos = last?;
+    // skip CCs after the `48 EB 01 B9`
+    let mut after = pos + sig.len();
+    while after < slice.len() && slice[after] == 0xCC {
+        after += 1;
+    }
+    if after + 4 > slice.len() {
+        return None;
+    }
+    Some((start + after) as u32)
+}
+
+/// Predict the 4 bytes that DecryptData5(va, size) would produce at va+0..va+4
+/// without mutating the buffer. The cipher's per-byte transform depends only
+/// on the byte itself and the low 8 bits of (va+i), with no cross-byte state,
+/// so each byte can be decrypted in isolation. Used to detect the EP/DD layout
+/// offset before committing to the actual call.
+pub fn trial_decrypt5_u32(data: &[u8], va: u32) -> u32 {
+    let mut out = [0u8; 4];
+    for i in 0..4u32 {
+        let b3 = data[(va + i) as usize];
+        let b = (va + i) as u8;
+        let b2 = b.wrapping_add(1);
+        let b4 = b3.rotate_left(2) ^ b2;
+        let b5 = b4.rotate_left(2) ^ b;
+        out[i as usize] = b5.rotate_left(2);
+    }
+    u32::from_le_bytes(out)
+}
+
+/// Scan stage4/stage5 for the encrypted custom-decryptor bytecode block. The
+/// raw byte at p+95 is used by decrypt_data6 as the iteration count. We trial-
+/// decrypt that many bytes with the LFSR keystream and accept the first
+/// position where the byte stream parses as a valid opcode sequence ending in
+/// 195 (ret).
+pub fn find_bytecode_offset(data: &[u8], base: u32, len: u32) -> Option<u32> {
+    let start = base as usize;
+    let end = (base.saturating_add(len)) as usize;
+    if end > data.len() {
+        return None;
+    }
+    let mut ks = [0u8; 256];
+    lfsr_keystream(&mut ks);
+    // Scan forward from `start+16` on 16-byte boundaries relative to `start`.
+    // The bytecode block is positioned a fixed offset into stage4/stage5; the
+    // lowest parseable candidate is the real one (later ones are coincidental
+    // parses of trailing filler bytes that happen to map to valid opcodes).
+    // The enclosing buffer isn't necessarily 16-aligned to its absolute
+    // address in newer builds, so we anchor the stride to `start`.
+    let mut p = start + 16;
+    while p + 96 <= end {
+        let count = data[p + 95] as usize;
+        if count >= 8 && p + count <= end {
+            let mut buf = [0u8; 256];
+            let take = count.min(256);
+            for i in 0..take {
+                buf[i] = data[p + i] ^ ks[i];
+            }
+            if let Some(nops) = parse_bytecode_check(&buf[..take])
+                && nops >= 4
+            {
+                return Some(p as u32);
+            }
+        }
+        p += 16;
+    }
+    None
+}
+
+/// Validate bytecode structure without allocating a `Vec` of ops. Returns
+/// `Some(non_nop_op_count)` if the byte stream parses successfully as a valid
+/// opcode sequence ending in 195 (ret), `None` otherwise. Allows non-trivial
+/// bytecode filtering by op count.
+pub fn parse_bytecode_check(buf: &[u8]) -> Option<usize> {
+    let mut i = 0usize;
+    let mut nops: usize = 0;
+    while i < buf.len() {
+        let b = buf[i];
+        i += 1;
+        match b {
+            4 | 44 | 52 => {
+                if i >= buf.len() {
+                    return None;
+                }
+                i += 1;
+                nops += 1;
+            }
+            144 => {}
+            192 | 254 => {
+                if i >= buf.len() {
+                    return None;
+                }
+                let mb = buf[i];
+                i += 1;
+                let rm = mb & 7;
+                let mod_ = (mb >> 6) & 3;
+                let reg = (mb >> 3) & 7;
+                if mod_ != 3 || rm != 0 {
+                    return None;
+                }
+                if reg > 1 {
+                    return None;
+                }
+                if b == 192 {
+                    if i >= buf.len() {
+                        return None;
+                    }
+                    i += 1;
+                }
+                nops += 1;
+            }
+            195 => return Some(nops),
+            _ => return None,
+        }
+    }
+    None
+}
+
+/// Locate stage3's v4_val: the last non-zero dword in the buffer, anchored
+/// by the `C3 CC CC CC` (ret + 3 int3) immediately before it.
+pub fn find_v4_offset(data: &[u8], base: u32, len: u32) -> Option<u32> {
+    let start = base as usize;
+    let end = (base.saturating_add(len)) as usize;
+    if end > data.len() || end < start + 4 {
+        return None;
+    }
+    // walk backwards looking for the first non-zero byte
+    let mut i = end;
+    while i > start && data[i - 1] == 0 {
+        i -= 1;
+    }
+    if i < start + 4 {
+        return None;
+    }
+    // v_val occupies the 4 bytes ending at i (rounded up to dword boundary)
+    let v_end = i;
+    let v_start = ((v_end + 3) & !3).saturating_sub(4);
+    // require that the 4 bytes preceding v_val match `C3 CC CC CC`
+    if v_start < start + 4 || data[v_start - 4..v_start] != [0xC3, 0xCC, 0xCC, 0xCC] {
+        return None;
+    }
+    Some(v_start as u32)
+}
+
+/// Scan a sub-buffer for an ASCII needle; return its absolute position.
+pub fn find_str_pos(data: &[u8], base: u32, len: u32, needle: &[u8]) -> Option<u32> {
+    let start = base as usize;
+    let end = (base.saturating_add(len)) as usize;
+    if end > data.len() || needle.is_empty() {
+        return None;
+    }
+    data[start..end]
+        .windows(needle.len())
+        .position(|w| w == needle)
+        .map(|rel| (start + rel) as u32)
+}
+
+pub fn get_string_to_null(data: &[u8], offset: u32) -> String {
+    let start = offset as usize;
+    if start >= data.len() {
+        return String::new();
+    }
+    // Bounded: an unterminated run must never walk off the end of the buffer
+    // (panic) or scan unboundedly into unrelated data.
+    let limit = start.saturating_add(4096).min(data.len());
+    let mut i = start;
+    while i < limit && data[i] != 0 {
+        i += 1;
+    }
+    String::from_utf8_lossy(&data[start..i]).into_owned()
+}
+
+/// Read a PE section-name field: exactly 8 bytes, NOT necessarily
+/// NUL-terminated (a full-width name like `.textbss` has no NUL at all).
+/// Returns the name with trailing NULs stripped. Using `get_string_to_null`
+/// here would run past the field into the VirtualSize/VirtualAddress dwords.
+pub fn section_name(data: &[u8], offset: u32) -> String {
+    let start = offset as usize;
+    let Some(field) = data.get(start..start + 8) else {
+        return String::new();
+    };
+    let end = field.iter().position(|&b| b == 0).unwrap_or(8);
+    String::from_utf8_lossy(&field[..end]).into_owned()
+}
+
+// ---------------------------------------------------------------------------
+// PE32 (32-bit) helpers
+// ---------------------------------------------------------------------------
+
+/// PE32 shell-table locator. Walks the shell region (`info[6]`) for a dword
+/// equal to `info[6]` followed by a plausible shell size, returning the table
+/// base (`candidate = off - 0x88`) when `candidate+0x58` holds a valid pointer.
+pub fn find_tbl_pe32(data: &[u8], info: &[u32; 8]) -> Option<u32> {
+    let shell = info[6];
+    if (data.len() as u64) < 0x100 {
+        return None;
+    }
+    let hi = (shell as u64)
+        .saturating_add(0x3000)
+        .min(data.len() as u64 - 0x100) as u32;
+    let mut off = shell;
+    while off < hi {
+        if off as usize + 8 <= data.len() {
+            let candidate = off.wrapping_sub(0x88);
+            if candidate >= shell && get_u32(data, off) == info[6] {
+                let shell_size_val = get_u32(data, off.wrapping_add(4));
+                if shell_size_val > 0x1000 && shell_size_val < 0x100000 {
+                    let v58_off = candidate.wrapping_add(0x58);
+                    if (v58_off as usize + 4) <= data.len() {
+                        let v58 = get_u32(data, v58_off);
+                        if v58 > 0 && (v58 as usize) < data.len() {
+                            return Some(candidate);
+                        }
+                    }
+                }
+            }
+        }
+        off = off.wrapping_add(4);
+    }
+    None
+}
+
+/// Locate an LFSR-encrypted bytecode block (decrypt_data6 form) in a region.
+/// `start_off` is the byte offset to begin scanning at, `scan_backward`
+/// controls direction. Returns the relative offset of the block. Includes full
+/// opcode-walk validation of candidate blocks.
+pub fn find_lfsr_block(
+    data: &[u8],
+    base: u32,
+    size: u32,
+    start_off: u32,
+    scan_backward: bool,
+) -> Option<u32> {
+    if size < 96 {
+        return None;
+    }
+    let mut ks = [0u8; 128];
+    lfsr_keystream(&mut ks);
+    let check = |scan_off: u32| -> bool {
+        let abs_off = base.wrapping_add(scan_off) as usize;
+        if abs_off + 96 > data.len() {
+            return false;
+        }
+        let sz = data[abs_off + 95] as usize;
+        if !(10..=95).contains(&sz) {
+            return false;
+        }
+        let mut decoded = [0u8; 95];
+        for bi in 0..sz {
+            decoded[bi] = data[abs_off + bi] ^ ks[bi];
+        }
+        // Full bytecode validation (shared with the stage4/5 locator): every
+        // opcode must decode with a valid ModR/M and the stream must REACH a
+        // RET (0xC3) as an opcode. The previous check only required a 0xC3
+        // byte *anywhere* in the window and accepted a walk that ran off the
+        // end without hitting RET — a `0x04 0xC3` (ADD 0xC3) tail passed, so
+        // coincidental LFSR-shaped garbage was accepted as a decryptor block.
+        parse_bytecode_check(&decoded[..sz]).is_some()
+    };
+    if scan_backward {
+        let hi = size - 96;
+        if hi >= start_off {
+            let mut scan_off = hi;
+            loop {
+                if check(scan_off) {
+                    return Some(scan_off);
+                }
+                if scan_off == start_off {
+                    break;
+                }
+                scan_off -= 1;
+            }
+        }
+    } else {
+        let hi = size - 95;
+        let mut scan_off = start_off;
+        while scan_off < hi {
+            if check(scan_off) {
+                return Some(scan_off);
+            }
+            scan_off += 1;
+        }
+    }
+    None
+}
+
+/// Slots discovered in the eighthStage for the marker-less layout.
+pub struct EighthSlots {
+    /// Absolute address of the file-data decryptor LFSR bytecode block. The
+    /// fileCS chain pointer is derived downstream as `file_lfsr - 0x58`.
+    pub file_lfsr: u32,
+    /// Absolute address of the compressedInfo (ptr,size) table pointer slot.
+    pub compressed_info_ptr: u32,
+}
+
+/// Marker-independent eighthStage slot discovery (PE32+ branch).
+///
+/// Newer Crackproof builds (e.g. some native/managed DLLs) omit the
+/// `pm\0\0cm\0\0` and `00 00 00 40 01 00 00 00` markers that the older layout's
+/// walk3/walk4/walk5 slot derivation relies on. Instead this discovers the
+/// slots structurally:
+///   * Scan the eighthStage for every LFSR (decrypt_data6) bytecode block.
+///   * The file decryptor is the LFSR block whose `fileCS = lfsr - 0x58` holds
+///     a pointer sitting just past `info[3]` (smallest positive distance).
+///   * `compressedInfo` is the pointer slot whose 16-byte target, after a
+///     trial `decrypt_data5`, parses as a plausible (src,sSize,dst,dSize)
+///     descriptor.
+///
+/// Returns `None` if no plausible file LFSR is found. `eighth_start`/`eighth_dsz`
+/// bound the search region; `info3` is `info[3]`; `compress_data_offset` is
+/// `(!u32(file_data,0x1080)) + 0x1000`; `file_data_len` is the protected file
+/// length.
+#[allow(clippy::too_many_arguments)]
+pub fn discover_eighth_slots(
+    data: &[u8],
+    eighth_start: u32,
+    eighth_dsz: u32,
+    info3: u32,
+    compress_data_offset: u32,
+    file_data_len: u32,
+) -> Option<EighthSlots> {
+    // Collect all LFSR candidates (forward scan).
+    //
+    // Advance by 1 after each hit, NOT by 96. A false-positive LFSR match can sit
+    // just before the real file-decryptor block (observed on an il2cpp game
+    // assembly build, 2026-07-13: junk at rel=0x31C1, real block at 0x3210).
+    // Stepping by the LFSR body size then skips the real block and discovery
+    // fails. Byte-stepping is cheap: eighthStage is only a few KB.
+    let mut all_lfsrs: Vec<u32> = Vec::new();
+    let mut scan_off: u32 = 0;
+    while scan_off + 95 < eighth_dsz {
+        match find_lfsr_block(data, eighth_start, eighth_dsz, scan_off, false) {
+            Some(found) => {
+                all_lfsrs.push(found);
+                scan_off = found + 1;
+            }
+            None => break,
+        }
+    }
+
+    // Pick the file LFSR: prefer the candidate whose fileCS pointer sits the
+    // smallest positive distance past info[3].
+    let mut off_file_lfsr: Option<u32> = None;
+    let mut best_dist: Option<u32> = None;
+    for &lfsr_off in &all_lfsrs {
+        if lfsr_off < 0x58 {
+            continue;
+        }
+        let cs_off = lfsr_off - 0x58;
+        let cs_val = get_u32(data, eighth_start.wrapping_add(cs_off));
+        if !(0x1000 < cs_val && (cs_val as usize) < data.len()) {
+            continue;
+        }
+        if cs_val < info3 {
+            continue;
+        }
+        let dist = cs_val - info3;
+        if best_dist.is_none_or(|b| dist < b) {
+            best_dist = Some(dist);
+            off_file_lfsr = Some(lfsr_off);
+        }
+    }
+    // Fallback: last LFSR with any in-image fileCS pointer.
+    if off_file_lfsr.is_none() {
+        for &lfsr_off in all_lfsrs.iter().rev() {
+            if lfsr_off < 0x58 {
+                continue;
+            }
+            let cs_val = get_u32(data, eighth_start.wrapping_add(lfsr_off - 0x58));
+            if 0x1000 < cs_val && (cs_val as usize) < data.len() {
+                off_file_lfsr = Some(lfsr_off);
+                break;
+            }
+        }
+    }
+    let off_file_lfsr = off_file_lfsr?;
+    let off_file_cs = off_file_lfsr - 0x58;
+
+    // Trial-decrypt to find compressedInfo: the pointer slot in the data area
+    // (between fileCS region start and the LFSR) whose target parses as a valid
+    // (src,sSize,dst,dSize) descriptor after a transient decrypt_data5.
+    let scan_from = off_file_lfsr.saturating_sub(0x400);
+    let mut off_compressed_info: Option<u32> = None;
+    let mut doff = scan_from;
+    while doff < off_file_lfsr {
+        if doff == off_file_cs {
+            doff += 4;
+            continue;
+        }
+        let ptr_val = get_u32(data, eighth_start.wrapping_add(doff));
+        if !(0x1000 < ptr_val && (ptr_val as usize) < data.len().saturating_sub(16)) {
+            doff += 4;
+            continue;
+        }
+        // Predict decrypt_data5(ptr_val, 16) without mutating: each dword is
+        // position-keyed and independent, so trial_decrypt5_u32 per dword.
+        let src2 = trial_decrypt5_u32(data, ptr_val);
+        let s_sz2 = trial_decrypt5_u32(data, ptr_val + 4);
+        let dst2 = trial_decrypt5_u32(data, ptr_val + 8);
+        let d_sz2 = trial_decrypt5_u32(data, ptr_val + 12);
+        let src_file_off = src2.wrapping_add(compress_data_offset);
+        let valid = s_sz2 > 0
+            && s_sz2 < 0x200000
+            && (src_file_off as u64 + s_sz2 as u64) <= file_data_len as u64
+            && dst2 >= 0x1000
+            && (dst2 as u64 + d_sz2 as u64) <= data.len() as u64
+            && d_sz2 >= s_sz2
+            && d_sz2 < 0x200000;
+        if valid {
+            off_compressed_info = Some(doff);
+            break;
+        }
+        doff += 4;
+    }
+    let off_compressed_info = off_compressed_info?;
+
+    Some(EighthSlots {
+        file_lfsr: eighth_start.wrapping_add(off_file_lfsr),
+        compressed_info_ptr: eighth_start.wrapping_add(off_compressed_info),
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Task 4.1 regression: build a synthetic buffer whose valid bytecode block
+    /// sits PAST `len` but within `len*2`. Assert that the smaller window misses
+    /// it and the doubled window finds it.
+    #[test]
+    fn bytecode_locate_double_window_retry() {
+        // We place the block at offset (base + len + 16) which is inside
+        // the len*2 window but outside the len window.
+        let base: u32 = 0;
+        let len: u32 = 256;
+        // Block sits at base + len + 16 = 272, aligned to 16.
+        let block_pos: usize = (base + len + 16) as usize; // 272
+
+        // The buffer must be large enough for the block (block_pos + 96 bytes).
+        let buf_len = block_pos + 256;
+        let mut buf = vec![0u8; buf_len];
+
+        // Build a valid plaintext op stream:
+        //   [4, 0, 4, 0, 4, 0, 4, 0, 195]  (4 ADD-AL ops then RET)
+        // Padded to 10 bytes total; count >= 8.
+        let count: usize = 10;
+        let mut plain = [0u8; 256];
+        plain[0] = 4;
+        plain[1] = 0;
+        plain[2] = 4;
+        plain[3] = 0;
+        plain[4] = 4;
+        plain[5] = 0;
+        plain[6] = 4;
+        plain[7] = 0;
+        plain[8] = 195; // ret
+
+        // Compute the LFSR keystream and XOR the first `count` bytes to get the
+        // encrypted representation that the scanner would decrypt back.
+        let mut ks = [0u8; 256];
+        lfsr_keystream(&mut ks);
+        for i in 0..count {
+            buf[block_pos + i] = plain[i] ^ ks[i];
+        }
+        // Raw count byte at block_pos+95 (outside the XOR range since count=10 < 95).
+        buf[block_pos + 95] = count as u8;
+
+        // Verify our construction: find_bytecode_offset with len should NOT find it.
+        assert_eq!(
+            find_bytecode_offset(&buf, base, len),
+            None,
+            "smaller window should not find the block"
+        );
+
+        // The doubled window should find it at block_pos.
+        assert_eq!(
+            find_bytecode_offset(&buf, base, len.saturating_mul(2)),
+            Some(block_pos as u32),
+            "doubled window should locate the block"
+        );
+    }
+}
@@ -0,0 +1,481 @@
+//! PE import reconstruction and memory-image compaction.
+
+use senbei_crypto::primitives::{get_u16, get_u32, write_u16, write_u32};
+
+use super::super::MAX_IMAGE_SIZE;
+
+/// Read a NUL-terminated byte string starting at `off`, bounded to 512 bytes.
+/// Returns the raw bytes up to the terminator (excluding it).
+fn read_cstr_bounded(data: &[u8], off: u32) -> Vec<u8> {
+    let start = off as usize;
+    if start >= data.len() {
+        return Vec::new();
+    }
+    let limit = (start + 512).min(data.len());
+    let mut end = start;
+    while end < limit && data[end] != 0 {
+        end += 1;
+    }
+    data[start..end].to_vec()
+}
+
+fn align_up_u32(value: u32, alignment: u32) -> u32 {
+    ((value.wrapping_add(alignment - 1)) / alignment).wrapping_mul(alignment)
+}
+
+fn align_up_u64(value: u64, alignment: u64) -> u64 {
+    value.div_ceil(alignment) * alignment
+}
+
+#[derive(Clone)]
+enum ImportFunc {
+    Ordinal(u32),
+    Name(u16, Vec<u8>),
+}
+
+struct ImportDesc {
+    time_date: u32,
+    fwd_chain: u32,
+    dll_name: Vec<u8>,
+    iat_rva: u32,
+    functions: Vec<ImportFunc>,
+}
+
+/// Return true when PE32 imports already sit in the original `.idata` layout
+/// (so no relocation to `.kmiat` is needed). May write the IAT data directory
+/// (pe+0xD8).
+pub fn pe32_imports_already_match_idata_layout(data: &mut [u8], pe_header: u32) -> bool {
+    let opt_hdr_size = get_u16(data, pe_header.wrapping_add(20)) as u32;
+    let sec_table = pe_header.wrapping_add(24).wrapping_add(opt_hdr_size);
+    let num_sections = get_u16(data, pe_header.wrapping_add(6)) as u32;
+    let import_rva = get_u32(data, pe_header.wrapping_add(0x80));
+    let import_size = get_u32(data, pe_header.wrapping_add(0x84));
+    let len = data.len() as u32;
+    if !(import_rva > 0 && import_size > 0) {
+        return false;
+    }
+    for idx in 0..num_sections {
+        let sec_off = sec_table.wrapping_add(idx * 40);
+        if (sec_off as usize + 40) > data.len() {
+            return false;
+        }
+        if &data[sec_off as usize..sec_off as usize + 6] != b".idata" {
+            continue;
+        }
+        let sec_va = get_u32(data, sec_off.wrapping_add(12));
+        let sec_size =
+            get_u32(data, sec_off.wrapping_add(8)).max(get_u32(data, sec_off.wrapping_add(16)));
+        let sec_end = sec_va.wrapping_add(sec_size);
+        if !(sec_va <= import_rva
+            && import_rva < sec_end
+            && import_rva.wrapping_add(import_size) <= sec_end)
+        {
+            continue;
+        }
+        let first_oft = get_u32(data, import_rva);
+        let first_name = get_u32(data, import_rva.wrapping_add(12));
+        let first_iat = get_u32(data, import_rva.wrapping_add(16));
+        if !(sec_va <= first_oft
+            && first_oft < sec_end
+            && sec_va <= first_iat
+            && first_iat < sec_end)
+        {
+            return false;
+        }
+        if !(0x1000 < first_name && first_name < len) {
+            return false;
+        }
+        let dll_name = read_cstr_bounded(data, first_name);
+        let lower: Vec<u8> = dll_name.iter().map(|b| b.to_ascii_lowercase()).collect();
+        if !lower.ends_with(b".dll") {
+            return false;
+        }
+        let mut iat_min = first_iat;
+        let mut iat_max = first_iat;
+        let mut idt_pos = import_rva;
+        while idt_pos.wrapping_add(20) <= len {
+            let oft_rva = get_u32(data, idt_pos);
+            let name_rva = get_u32(data, idt_pos.wrapping_add(12));
+            let iat_rva = get_u32(data, idt_pos.wrapping_add(16));
+            if oft_rva == 0 && name_rva == 0 && iat_rva == 0 {
+                break;
+            }
+            if !(sec_va <= oft_rva && oft_rva < sec_end && sec_va <= iat_rva && iat_rva < sec_end) {
+                return false;
+            }
+            let mut thunk = iat_rva;
+            while thunk.wrapping_add(4) <= sec_end {
+                let tv = get_u32(data, thunk);
+                thunk = thunk.wrapping_add(4);
+                if tv == 0 {
+                    break;
+                }
+            }
+            iat_min = iat_min.min(iat_rva);
+            iat_max = iat_max.max(thunk);
+            idt_pos = idt_pos.wrapping_add(20);
+        }
+        if iat_max > iat_min {
+            write_u32(data, pe_header.wrapping_add(0xD8), iat_min);
+            write_u32(data, pe_header.wrapping_add(0xDC), iat_max - iat_min);
+        }
+        return true;
+    }
+    false
+}
+
+/// Rebuild PE32 import metadata (descriptors, lookup tables, names) into the
+/// last section as `.kmiat`, leaving the loader-written IAT in place. Mutates
+/// `data` (may grow it).
+pub fn move_pe32_imports_to_kmiat(data: &mut Vec<u8>, pe_header: u32) {
+    const SECTION_SIZE: u32 = 0x7000;
+    let opt_hdr_size = get_u16(data, pe_header.wrapping_add(20)) as u32;
+    let opt_hdr = pe_header.wrapping_add(24);
+    let sec_table = opt_hdr.wrapping_add(opt_hdr_size);
+    let num_sections = get_u16(data, pe_header.wrapping_add(6)) as u32;
+    if num_sections == 0 {
+        return;
+    }
+    let import_rva = get_u32(data, pe_header.wrapping_add(0x80));
+    let import_size = get_u32(data, pe_header.wrapping_add(0x84));
+    let len = data.len() as u32;
+    if !(0x1000 < import_rva && import_rva < len && import_size > 0 && import_size < SECTION_SIZE) {
+        return;
+    }
+
+    let mut descriptors: Vec<ImportDesc> = Vec::new();
+    let mut idt_pos = import_rva;
+    while idt_pos.wrapping_add(20) <= len {
+        let oft_rva = get_u32(data, idt_pos);
+        let time_date = get_u32(data, idt_pos.wrapping_add(4));
+        let fwd_chain = get_u32(data, idt_pos.wrapping_add(8));
+        let name_rva = get_u32(data, idt_pos.wrapping_add(12));
+        let iat_rva = get_u32(data, idt_pos.wrapping_add(16));
+        if oft_rva == 0 && name_rva == 0 && iat_rva == 0 {
+            break;
+        }
+        if !(0x1000 < name_rva && name_rva < len) {
+            break;
+        }
+        let dll_name = read_cstr_bounded(data, name_rva);
+        let thunk_rva = if 0x1000 < oft_rva && oft_rva < len {
+            oft_rva
+        } else {
+            iat_rva
+        };
+        let mut functions: Vec<ImportFunc> = Vec::new();
+        let mut thunk_pos = thunk_rva;
+        while 0x1000 < thunk_pos.wrapping_add(4) && thunk_pos.wrapping_add(4) <= len {
+            let thunk_val = get_u32(data, thunk_pos);
+            if thunk_val == 0 {
+                break;
+            }
+            if thunk_val & 0x8000_0000 != 0 {
+                functions.push(ImportFunc::Ordinal(thunk_val & 0xFFFF));
+            } else {
+                let hint = if thunk_val.wrapping_add(2) <= len {
+                    get_u16(data, thunk_val)
+                } else {
+                    0
+                };
+                let func_name = if thunk_val.wrapping_add(2) < len {
+                    read_cstr_bounded(data, thunk_val.wrapping_add(2))
+                } else {
+                    Vec::new()
+                };
+                functions.push(ImportFunc::Name(hint, func_name));
+            }
+            thunk_pos = thunk_pos.wrapping_add(4);
+        }
+        descriptors.push(ImportDesc {
+            time_date,
+            fwd_chain,
+            dll_name,
+            iat_rva,
+            functions,
+        });
+        idt_pos = idt_pos.wrapping_add(20);
+    }
+    if descriptors.is_empty() {
+        return;
+    }
+
+    for desc in &mut descriptors {
+        let lower: Vec<u8> = desc
+            .dll_name
+            .iter()
+            .map(|b| b.to_ascii_lowercase())
+            .collect();
+        if lower.starts_with(b"api-ms-win-crt-") {
+            desc.dll_name = b"ucrtbase.dll".to_vec();
+        } else {
+            desc.dll_name = lower;
+        }
+    }
+    descriptors.sort_by_key(|d| d.iat_rva);
+
+    let last_sec = sec_table.wrapping_add((num_sections - 1) * 40);
+    let kmiat_rva = get_u32(data, last_sec.wrapping_add(12));
+    // A zero last-section VA means a corrupt section table: building .kmiat at
+    // RVA 0 would zero the DOS/PE headers and emit a structurally broken image
+    // with no error. Bail and keep the original import table.
+    if kmiat_rva == 0 {
+        return;
+    }
+    // Grow the image when .kmiat overruns it, but cap the growth: a corrupt VA
+    // could otherwise request a multi-gigabyte allocation, which aborts the
+    // process (uncatchable). Use u64 math so a near-u32::MAX VA cannot wrap the
+    // end calculation the way the previous wrapping/plain-add mix could.
+    let kmiat_end = kmiat_rva as u64 + SECTION_SIZE as u64;
+    if kmiat_end > MAX_IMAGE_SIZE {
+        return;
+    }
+    if kmiat_end > data.len() as u64 {
+        data.resize(kmiat_end as usize, 0);
+    }
+    // Zero the .kmiat region.
+    for b in &mut data[kmiat_rva as usize..kmiat_end as usize] {
+        *b = 0;
+    }
+
+    let idt_size = (descriptors.len() as u32 + 1) * 20;
+    let oft_start = kmiat_rva;
+    let mut idt_rva = oft_start;
+    for desc in &descriptors {
+        idt_rva = idt_rva.wrapping_add((desc.functions.len() as u32 + 1) * 4);
+    }
+    idt_rva = align_up_u32(idt_rva.wrapping_add(0x2C), 4);
+
+    // Size check: compute the final name_pos and bail if it overruns .kmiat.
+    let mut name_pos_check = idt_rva.wrapping_add(idt_size);
+    for desc in &descriptors {
+        name_pos_check = name_pos_check.wrapping_add(desc.dll_name.len() as u32 + 1);
+        for func in &desc.functions {
+            if let ImportFunc::Name(_, fname) = func {
+                name_pos_check = name_pos_check.wrapping_add(2 + fname.len() as u32 + 1);
+            }
+        }
+    }
+    if name_pos_check > kmiat_rva.wrapping_add(SECTION_SIZE) {
+        // Section too small; keep existing import table untouched.
+        return;
+    }
+
+    let mut oft_pos = oft_start;
+    let mut name_pos = idt_rva.wrapping_add(idt_size);
+    for (idx, desc) in descriptors.iter().enumerate() {
+        let idt_entry = idt_rva.wrapping_add(idx as u32 * 20);
+        let current_oft = oft_pos;
+        write_u32(data, idt_entry, current_oft);
+        write_u32(data, idt_entry.wrapping_add(4), desc.time_date);
+        write_u32(data, idt_entry.wrapping_add(8), desc.fwd_chain);
+        let dll_name_pos = name_pos;
+        write_u32(data, idt_entry.wrapping_add(12), dll_name_pos);
+        write_u32(data, idt_entry.wrapping_add(16), desc.iat_rva);
+
+        let dnp = dll_name_pos as usize;
+        data[dnp..dnp + desc.dll_name.len()].copy_from_slice(&desc.dll_name);
+        data[dnp + desc.dll_name.len()] = 0;
+        name_pos = name_pos.wrapping_add(desc.dll_name.len() as u32 + 1);
+
+        for func in &desc.functions {
+            match func {
+                ImportFunc::Ordinal(ord) => {
+                    write_u32(data, oft_pos, 0x8000_0000 | ord);
+                }
+                ImportFunc::Name(hint, fname) => {
+                    let hint_name_rva = name_pos;
+                    write_u32(data, oft_pos, hint_name_rva);
+                    write_u16(data, hint_name_rva, *hint as u32);
+                    let fp = (hint_name_rva + 2) as usize;
+                    data[fp..fp + fname.len()].copy_from_slice(fname);
+                    data[fp + fname.len()] = 0;
+                    name_pos = name_pos.wrapping_add(2 + fname.len() as u32 + 1);
+                }
+            }
+            oft_pos = oft_pos.wrapping_add(4);
+        }
+        write_u32(data, oft_pos, 0);
+        oft_pos = oft_pos.wrapping_add(4);
+    }
+    // Null-terminator IDT entry (20 zero bytes) after the last descriptor.
+    let term = idt_rva.wrapping_add(descriptors.len() as u32 * 20) as usize;
+    for b in &mut data[term..term + 20] {
+        *b = 0;
+    }
+
+    let ls = last_sec as usize;
+    data[ls..ls + 8].copy_from_slice(b".kmiat\x00\x00");
+    write_u32(data, last_sec.wrapping_add(8), SECTION_SIZE);
+    write_u32(data, last_sec.wrapping_add(16), SECTION_SIZE);
+    write_u32(data, last_sec.wrapping_add(36), 0xE000_0060);
+    write_u32(data, pe_header.wrapping_add(0x80), idt_rva);
+    write_u32(data, pe_header.wrapping_add(0x84), idt_size);
+    write_u32(
+        data,
+        pe_header.wrapping_add(80),
+        kmiat_rva.wrapping_add(SECTION_SIZE),
+    );
+}
+
+/// Convert the unpacked RVA-addressed image back to a compact PE file layout
+/// (headers at 0x400, sections packed consecutively, FileAlignment 0x200).
+/// Returns `None` if the accumulated output size wraps or exceeds
+/// [`MAX_IMAGE_SIZE`]: the final allocation is sized from header-derived
+/// section data, and an uncapped `vec![0; n]` from a corrupt header would abort
+/// the process (which `catch_unpack` cannot trap).
+pub fn compact_memory_image_to_pe(data: &[u8], pe_header: u32) -> Option<Vec<u8>> {
+    const FILE_ALIGNMENT: u32 = 0x200;
+    const HEADER_SIZE: u32 = 0x400;
+    let opt_hdr_size = get_u16(data, pe_header.wrapping_add(20)) as u32;
+    let opt_hdr = pe_header.wrapping_add(24);
+    let sec_table = opt_hdr.wrapping_add(opt_hdr_size);
+    let num_sections = get_u16(data, pe_header.wrapping_add(6)) as u32;
+
+    struct SecLayout {
+        sec_off: u32,
+        va: u32,
+        vsize: u32,
+        raw_ptr: u32,
+        raw_size: u32,
+    }
+
+    let mut raw_cursor: u64 = HEADER_SIZE as u64;
+    let mut raw_layout: Vec<SecLayout> = Vec::new();
+    for idx in 0..num_sections {
+        let sec_off = sec_table.wrapping_add(idx * 40);
+        let vsize = get_u32(data, sec_off.wrapping_add(8));
+        let va = get_u32(data, sec_off.wrapping_add(12));
+        let sd_start = va as usize;
+        let sd_end = if (va.wrapping_add(vsize) as usize) <= data.len() {
+            va.wrapping_add(vsize) as usize
+        } else {
+            data.len()
+        };
+        let section_data: &[u8] = if sd_start <= sd_end && sd_start <= data.len() {
+            &data[sd_start..sd_end]
+        } else {
+            &[]
+        };
+
+        let mut last_nonzero: i64 = -1;
+        for pos in (0..section_data.len()).rev() {
+            if section_data[pos] != 0 {
+                last_nonzero = pos as i64;
+                break;
+            }
+        }
+        let meaningful = if last_nonzero >= 0 {
+            (last_nonzero + 1) as u32
+        } else {
+            0
+        };
+        let mut raw_size = if meaningful != 0 {
+            align_up_u32(meaningful, FILE_ALIGNMENT)
+        } else {
+            0
+        };
+        if vsize != 0 && raw_size == 0 {
+            raw_size = FILE_ALIGNMENT;
+        }
+        raw_size = raw_size.min(align_up_u32(section_data.len() as u32, FILE_ALIGNMENT));
+
+        let raw_ptr = if raw_size != 0 { raw_cursor as u32 } else { 0 };
+        raw_layout.push(SecLayout {
+            sec_off,
+            va,
+            vsize,
+            raw_ptr,
+            raw_size,
+        });
+        if raw_size != 0 {
+            // Accumulate in u64 and cap: section sizes are header-derived, and
+            // a corrupt table could otherwise wrap raw_cursor (small alloc,
+            // huge recorded raw_ptrs → OOB panic) or request an abort-sized
+            // allocation.
+            raw_cursor = align_up_u64(raw_cursor + raw_size as u64, FILE_ALIGNMENT as u64);
+            if raw_cursor > MAX_IMAGE_SIZE {
+                return None;
+            }
+        }
+    }
+
+    let mut compact = vec![0u8; raw_cursor as usize];
+    let hdr_copy = (HEADER_SIZE as usize).min(data.len());
+    compact[..hdr_copy].copy_from_slice(&data[..hdr_copy]);
+    write_u32(&mut compact, opt_hdr.wrapping_add(36), FILE_ALIGNMENT);
+    write_u32(&mut compact, opt_hdr.wrapping_add(60), HEADER_SIZE);
+
+    for sl in &raw_layout {
+        write_u32(&mut compact, sl.sec_off.wrapping_add(16), sl.raw_size);
+        write_u32(&mut compact, sl.sec_off.wrapping_add(20), sl.raw_ptr);
+        if sl.raw_size != 0 {
+            let sd_start = sl.va as usize;
+            let sd_end = if (sl.va.wrapping_add(sl.vsize) as usize) <= data.len() {
+                sl.va.wrapping_add(sl.vsize) as usize
+            } else {
+                data.len()
+            };
+            let section_data: &[u8] = if sd_start <= sd_end {
+                &data[sd_start..sd_end]
+            } else {
+                &[]
+            };
+            let copy_size = (sl.raw_size as usize).min(section_data.len());
+            let rp = sl.raw_ptr as usize;
+            compact[rp..rp + copy_size].copy_from_slice(&section_data[..copy_size]);
+        }
+    }
+    Some(compact)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Review regression: a zero last-section VA (corrupt section table) must
+    /// bail instead of building .kmiat at RVA 0 — the old code zeroed
+    /// `[0, 0x7000)`, wiping the DOS/PE headers, and returned the broken image
+    /// as a success. A near-2 GiB VA must likewise refuse to grow the image
+    /// past [`MAX_IMAGE_SIZE`].
+    #[test]
+    fn kmiat_bogus_section_va_bails_without_wiping_headers() {
+        for last_sec_va in [0u32, 0x5000_0000] {
+            let pe: u32 = 0x80;
+            let mut data = vec![0xAAu8; 0x8000];
+            // COFF header: 1 section, optional header size 0xE0 (PE32).
+            write_u16(&mut data, pe + 6, 1);
+            write_u16(&mut data, pe + 20, 0xE0);
+            // Import directory at pe+0x80: one descriptor + null terminator.
+            write_u32(&mut data, pe + 0x80, 0x1100);
+            write_u32(&mut data, pe + 0x84, 0x28);
+            write_u32(&mut data, 0x1100, 0x1200); // OFT rva
+            write_u32(&mut data, 0x1100 + 12, 0x1300); // name rva
+            write_u32(&mut data, 0x1100 + 16, 0x1400); // IAT rva
+            for b in &mut data[0x1100 + 20..0x1100 + 40] {
+                *b = 0; // null terminator descriptor
+            }
+            data[0x1300..0x1300 + 13].copy_from_slice(b"KERNEL32.dll\0");
+            write_u32(&mut data, 0x1200, 0x1500); // thunk -> hint/name
+            write_u32(&mut data, 0x1204, 0); // thunk terminator
+            data[0x1500..0x1502].copy_from_slice(&0u16.to_le_bytes());
+            data[0x1502..0x1502 + 12].copy_from_slice(b"ExitProcess\0");
+            // Section table at pe+24+0xE0 = 0x178; VA field at +12.
+            write_u32(&mut data, 0x178 + 12, last_sec_va);
+
+            let head_before: Vec<u8> = data[..0x400].to_vec();
+            let len_before = data.len();
+            move_pe32_imports_to_kmiat(&mut data, pe);
+            assert_eq!(
+                data.len(),
+                len_before,
+                "VA 0x{last_sec_va:08X}: image must not grow"
+            );
+            assert_eq!(
+                &data[..0x400],
+                &head_before[..],
+                "VA 0x{last_sec_va:08X}: headers must be untouched"
+            );
+        }
+    }
+}
@@ -0,0 +1,391 @@
+//! PE detection, unpacking, and structural validation.
+
+pub mod dll;
+mod error;
+pub mod exe;
+pub mod integrity;
+mod layout;
+pub(crate) mod parallel;
+
+use senbei_crypto::primitives;
+use std::cell::RefCell;
+use std::sync::{Arc, Mutex};
+
+pub use dll::{unpack_dll, unpack_dll_v};
+pub use error::*;
+pub use exe::{unpack as unpack_exe, unpack_v as unpack_exe_v};
+pub use integrity::{IntegrityReport, check as check_integrity};
+pub use parallel::thread_cap;
+
+/// Maximum plausible PE `SizeOfImage` we are willing to allocate a zero buffer
+/// for. Guards against a corrupt/crafted header requesting a multi-gigabyte
+/// (or, as a sign-extended negative `i32`, multi-exabyte) allocation, which
+/// would abort the process — an abort that `catch_unpack` below cannot trap.
+/// Real protected binaries are far below this.
+pub(crate) const MAX_IMAGE_SIZE: u64 = senbei_crypto::MAX_IMAGE_SIZE;
+
+#[derive(Clone)]
+pub(crate) struct PanicCapture(Arc<Mutex<Option<PanicDetails>>>);
+
+#[derive(Clone)]
+struct PanicDetails {
+    message: String,
+    file: String,
+    line: u32,
+    column: u32,
+}
+
+thread_local! {
+    static ACTIVE_PANIC_CAPTURE: RefCell<Option<PanicCapture>> = const { RefCell::new(None) };
+}
+
+struct PanicCaptureGuard(Option<PanicCapture>);
+
+impl Drop for PanicCaptureGuard {
+    fn drop(&mut self) {
+        ACTIVE_PANIC_CAPTURE.with(|slot| {
+            slot.replace(self.0.take());
+        });
+    }
+}
+
+impl PanicCapture {
+    fn new() -> Self {
+        Self(Arc::new(Mutex::new(None)))
+    }
+
+    fn record(&self, info: &std::panic::PanicHookInfo<'_>) {
+        let location = info.location();
+        let details = PanicDetails {
+            message: panic_message(info.payload()),
+            file: location
+                .map(|value| value.file().to_owned())
+                .unwrap_or_else(|| "<unknown>".to_owned()),
+            line: location.map_or(0, std::panic::Location::line),
+            column: location.map_or(0, std::panic::Location::column),
+        };
+        let mut captured = self
+            .0
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if captured.is_none() {
+            *captured = Some(details);
+        }
+    }
+
+    fn into_error(self, payload: &(dyn std::any::Any + Send)) -> UnpackError {
+        let details = self
+            .0
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .clone()
+            .unwrap_or_else(|| PanicDetails {
+                message: panic_message(payload),
+                file: "<unknown>".to_owned(),
+                line: 0,
+                column: 0,
+            });
+        UnpackError::InternalPanic {
+            message: details.message,
+            file: details.file,
+            line: details.line,
+            column: details.column,
+        }
+    }
+
+    fn merge_from(&self, other: &Self) {
+        let details = other
+            .0
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .clone();
+        let Some(details) = details else { return };
+        let mut captured = self
+            .0
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if captured.is_none() {
+            *captured = Some(details);
+        }
+    }
+}
+
+fn panic_message(payload: &(dyn std::any::Any + Send)) -> String {
+    if let Some(message) = payload.downcast_ref::<&str>() {
+        (*message).to_owned()
+    } else if let Some(message) = payload.downcast_ref::<String>() {
+        message.clone()
+    } else {
+        "non-string panic payload".to_owned()
+    }
+}
+
+fn install_panic_capture_hook() {
+    static INSTALL: std::sync::Once = std::sync::Once::new();
+    INSTALL.call_once(|| {
+        let previous = std::panic::take_hook();
+        std::panic::set_hook(Box::new(move |info| {
+            let capture = ACTIVE_PANIC_CAPTURE
+                .try_with(|slot| slot.borrow().clone())
+                .ok()
+                .flatten();
+            if let Some(capture) = capture {
+                capture.record(info);
+            } else {
+                previous(info);
+            }
+        }));
+    });
+}
+
+pub(crate) fn current_panic_capture() -> Option<PanicCapture> {
+    ACTIVE_PANIC_CAPTURE.with(|slot| slot.borrow().clone())
+}
+
+pub(crate) fn with_panic_capture<R>(capture: Option<PanicCapture>, f: impl FnOnce() -> R) -> R {
+    let previous = ACTIVE_PANIC_CAPTURE.with(|slot| slot.replace(capture));
+    let _guard = PanicCaptureGuard(previous);
+    f()
+}
+
+/// Run an unpack pipeline, converting any internal panic into a clean
+/// [`UnpackError::InternalPanic`] so the public API stays panic-free on any input
+/// (truncated/garbled files chase offsets out of bounds). The panic location and
+/// payload are captured for diagnostics without printing a backtrace to stderr.
+///
+/// Note: allocation *failures* abort the process and are NOT caught here; size
+/// requests are bounds-checked against [`MAX_IMAGE_SIZE`] before allocating.
+pub(crate) fn catch_unpack<F>(f: F) -> Result<Vec<u8>, UnpackError>
+where
+    F: FnOnce() -> Result<Vec<u8>, UnpackError>,
+{
+    install_panic_capture_hook();
+    let capture = PanicCapture::new();
+    let r = with_panic_capture(Some(capture.clone()), || {
+        std::panic::catch_unwind(std::panic::AssertUnwindSafe(f))
+    });
+    match r {
+        Ok(result) => result,
+        Err(payload) => Err(capture.into_error(payload.as_ref())),
+    }
+}
+
+/// Crackproof header magic stored in `keys[1]`/`info[1]`.
+pub(crate) const MAGIC_KONN: u32 = 0x4E4E4F4B; // b"KONN" little-endian (= 1313754955)
+
+/// True if `magic` is the Crackproof magic this unpacker supports.
+pub(crate) fn is_supported_magic(magic: u32) -> bool {
+    magic == MAGIC_KONN
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Kind {
+    NativeExe,
+    ManagedExe,
+    NativeDll,
+    ManagedDll,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct Detected {
+    pub kind: Kind,
+    pub magic: u32,
+}
+
+// ---------------------------------------------------------------------------
+// Content-based detection
+// ---------------------------------------------------------------------------
+
+/// Derive the 8-element Crackproof key table from the header at offset 4096.
+/// Returns `None` if the input is too short or doesn't have a valid PE signature.
+fn key_table(input: &[u8]) -> Option<[u32; 8]> {
+    // Need at least 4128 bytes: the key-table loop below reads dwords up to
+    // offset 4124 (bytes 4124..4127). Guarding only `< 4096` would let a
+    // 4096..4127-byte PE (e.g. a 4 KiB stub) panic in `get_u32`.
+    if input.len() < 4128 {
+        return None;
+    }
+    // Validate the PE signature with checked arithmetic so a crafted offset
+    // cannot wrap the bounds check on a narrower target.
+    let e_lfanew = primitives::get_u32(input, 0x3C);
+    let pe_start = e_lfanew as usize;
+    if pe_start.checked_add(4).is_none_or(|end| end > input.len()) {
+        return None;
+    }
+    if &input[pe_start..pe_start + 4] != b"PE\0\0" {
+        return None;
+    }
+    // Derive 8 keys per the Crackproof header-key formula.
+    let mut keys = [0u32; 8];
+    keys[0] = primitives::get_u32(input, 4096);
+    let mut k = keys[0];
+    for i in 0u32..7 {
+        let cell = primitives::get_u32(input, 4100u32.wrapping_add(i.wrapping_mul(4)));
+        keys[(i + 1) as usize] = k ^ cell;
+        k = i.wrapping_mul(i) ^ (k.wrapping_add(cell).wrapping_sub(i));
+    }
+    Some(keys)
+}
+
+/// Detect whether `input` is a Crackproof-protected binary and classify it.
+/// Returns `None` if the magic doesn't match.
+///
+/// Routing: `keys[1]` must be the Crackproof magic (`KONN`).
+/// The PE IMAGE_FILE_DLL characteristic distinguishes EXE vs DLL;
+/// the CLR data-directory RVA distinguishes managed from native for both.
+pub fn detect(input: &[u8]) -> Option<Detected> {
+    let keys = key_table(input)?;
+    let magic = keys[1];
+    // Anything whose magic doesn't match is left untouched rather than
+    // detected-then-errored, honoring the "anything that doesn't match is
+    // left untouched" contract.
+    if !is_supported_magic(magic) {
+        return None;
+    }
+    // Use the PE DLL characteristic to distinguish EXE from DLL.
+    // IMAGE_FILE_HEADER.Characteristics is at peOff+4+18; bit 0x2000 = IMAGE_FILE_DLL.
+    let pe_off = primitives::get_u32(input, 0x3C);
+    let chars_offset = pe_off.wrapping_add(4).wrapping_add(18);
+    if (chars_offset as usize)
+        .checked_add(2)
+        .is_none_or(|end| end > input.len())
+    {
+        return None;
+    }
+    let chars =
+        (input[chars_offset as usize] as u16) | ((input[chars_offset as usize + 1] as u16) << 8);
+    let is_dll = (chars & 0x2000) != 0;
+    // Managed vs native via the CLR data-directory RVA.
+    // peOff + 24 = start of optional header. The data directories start at a
+    // magic-dependent offset within it: PE32 (0x10B) at +96, PE32+ (0x20B) at
+    // +112. Using the PE32+ offset on a PE32 image reads the wrong dword and
+    // can mis-flag a native image as managed.
+    //
+    // `get_u16`/`get_u32` index unchecked, so every read past the already-
+    // checked Characteristics word must be bounds-checked first: a truncated
+    // file (e.g. `e_lfanew` pointing at len-24) would otherwise panic here,
+    // and this detector runs on the folder scan threads where a panic aborts
+    // the whole run.
+    let opt_magic_off = pe_off.wrapping_add(24) as usize;
+    let b = input.get(opt_magic_off..opt_magic_off.checked_add(2)?)?;
+    let opt_magic = u16::from_le_bytes([b[0], b[1]]);
+    let dd_off: u32 = if opt_magic == 0x20B { 112 } else { 96 };
+    // + 14*8 = IMAGE_DIRECTORY_ENTRY_COM_DESCRIPTOR
+    let clr_rva_offset = pe_off
+        .wrapping_add(24)
+        .wrapping_add(dd_off)
+        .wrapping_add(14u32.wrapping_mul(8));
+    if (clr_rva_offset as usize)
+        .checked_add(4)
+        .is_none_or(|end| end > input.len())
+    {
+        return None;
+    }
+    let clr_rva = primitives::get_u32(input, clr_rva_offset);
+    let kind = match (is_dll, clr_rva != 0) {
+        (false, false) => Kind::NativeExe,
+        (false, true) => Kind::ManagedExe,
+        (true, false) => Kind::NativeDll,
+        (true, true) => Kind::ManagedDll,
+    };
+    Some(Detected { kind, magic })
+}
+
+/// Detect the file type and dispatch to the matching pipeline.
+/// Returns the detected `Kind` together with the unpacked image bytes.
+pub fn unpack_auto(input: &[u8]) -> Result<(Kind, Vec<u8>), UnpackError> {
+    unpack_auto_v(input, false)
+}
+
+/// Like [`unpack_auto`], but prints detailed `[N/9]` unpack-step progress to
+/// stdout when `verbose` is true. Output bytes are identical regardless.
+pub fn unpack_auto_v(input: &[u8], verbose: bool) -> Result<(Kind, Vec<u8>), UnpackError> {
+    let detected = detect(input).ok_or(UnpackError::NotCrackproof)?;
+    let out = match detected.kind {
+        Kind::NativeExe | Kind::ManagedExe => unpack_exe_v(input, verbose)?,
+        Kind::NativeDll | Kind::ManagedDll => {
+            // Two Crackproof DLL layouts exist. The older one (the byte-identical
+            // DLL goldens) follows the pipeline in `dll.rs`. Newer builds protect
+            // DLLs with the EXE-style shell layout instead — `dll::unpack_dll`
+            // cannot parse them and errors. Try the DLL pipeline first; on
+            // failure, fall back to the EXE pipeline, which handles the new
+            // layout (including managed-DLL CLR metadata restore). The DLL-first
+            // order keeps the old-layout goldens byte-identical (the EXE
+            // pipeline "succeeds" on them but with different bytes).
+            match dll::unpack_dll_v(input, verbose) {
+                Ok(out) => out,
+                Err(dll_err) => match exe::unpack_v(input, verbose) {
+                    Ok(out) => out,
+                    Err(exe_err) => {
+                        return Err(UnpackError::PipelineFallbackFailed {
+                            dll: Box::new(dll_err),
+                            exe: Box::new(exe_err),
+                        });
+                    }
+                },
+            }
+        }
+    };
+    Ok((detected.kind, out))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn caught_panic_reports_location_and_message() {
+        let error = catch_unpack(|| -> Result<Vec<u8>, UnpackError> {
+            panic!("test panic");
+        })
+        .expect_err("panic must become an error");
+        let UnpackError::InternalPanic {
+            message,
+            file,
+            line,
+            column,
+        } = error
+        else {
+            panic!("unexpected error: {error}");
+        };
+        assert_eq!(message, "test panic");
+        assert!(
+            file.ends_with("senbei-pe/src/engine/mod.rs")
+                || file.ends_with("senbei-pe\\src\\engine\\mod.rs")
+        );
+        assert!(line > 0);
+        assert!(column > 0);
+    }
+
+    #[test]
+    fn worker_panic_keeps_the_worker_source_location() {
+        let error = catch_unpack(|| -> Result<Vec<u8>, UnpackError> {
+            let capture = current_panic_capture();
+            let result = std::thread::spawn(move || {
+                with_panic_capture(capture, || panic!("worker panic"));
+            })
+            .join();
+            if let Err(payload) = result {
+                std::panic::resume_unwind(payload);
+            }
+            Ok(Vec::new())
+        })
+        .expect_err("worker panic must become an error");
+        let UnpackError::InternalPanic {
+            message,
+            file,
+            line,
+            column,
+        } = error
+        else {
+            panic!("unexpected error: {error}");
+        };
+        assert_eq!(message, "worker panic");
+        assert!(
+            file.ends_with("senbei-pe/src/engine/mod.rs")
+                || file.ends_with("senbei-pe\\src\\engine\\mod.rs")
+        );
+        assert!(line > 0);
+        assert!(column > 0);
+    }
+}
@@ -21,7 +21,7 @@ use std::sync::atomic::{AtomicBool, Ordering};

 /// Worker-thread cap. `SENBEI_THREADS` overrides it (`1` forces the sequential
 /// path); otherwise the host's available parallelism; otherwise 1.
-pub(crate) fn thread_cap() -> usize {
+pub fn thread_cap() -> usize {
    if let Ok(v) = std::env::var("SENBEI_THREADS")
        && let Ok(n) = v.trim().parse::<usize>()
        && n >= 1
@@ -46,7 +46,7 @@ pub(crate) fn thread_cap() -> usize {
 ///
 /// Returns the first `Err` any block produces; re-raises the first block panic
 /// on the calling thread (so the pipeline's existing `catch_unpack` still
-/// converts it to `UnpackError::Corrupt`).
+/// converts it to `UnpackError::InternalPanic`).
 pub(crate) fn parallel_for<E, F>(
    buf: &mut [u8],
    spans: &[(usize, usize)],
@@ -125,6 +125,7 @@ where
    let stop = AtomicBool::new(false);
    let first_err: Mutex<Option<E>> = Mutex::new(None);
    let first_panic: Mutex<Option<Box<dyn std::any::Any + Send>>> = Mutex::new(None);
+    let panic_capture = super::current_panic_capture();

    std::thread::scope(|scope| {
        for _ in 0..workers {
@@ -133,6 +134,7 @@ where
            let first_err = &first_err;
            let first_panic = &first_panic;
            let f = &f;
+            let panic_capture = panic_capture.clone();
            scope.spawn(move || {
                loop {
                    if stop.load(Ordering::Relaxed) {
@@ -141,9 +143,15 @@ where
                    let next = iter.lock().unwrap().next();
                    let Some((i, piece)) = next else { break };
                    let span = piece.unwrap();
-                    let r = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+                    // Keep details local until this panic wins `first_panic`;
+                    // otherwise simultaneous workers could pair one worker's
+                    // location with another worker's propagated payload.
+                    let block_capture = panic_capture.as_ref().map(|_| super::PanicCapture::new());
+                    let r = super::with_panic_capture(block_capture.clone(), || {
+                        std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                            f(i, spans[i].0, span)
-                    }));
+                        }))
+                    });
                    match r {
                        Ok(Ok(())) => {}
                        Ok(Err(e)) => {
@@ -157,6 +165,11 @@ where
                        Err(panic) => {
                            let mut slot = first_panic.lock().unwrap();
                            if slot.is_none() {
+                                if let (Some(parent), Some(block)) =
+                                    (&panic_capture, &block_capture)
+                                {
+                                    parent.merge_from(block);
+                                }
                                *slot = Some(panic);
                            }
                            stop.store(true, Ordering::Relaxed);
@@ -0,0 +1,5 @@
+//! PE detection, unpacking, and structural validation.
+
+mod engine;
+
+pub use engine::*;
@@ -50,21 +50,21 @@ checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"

 [[package]]
 name = "futures-core"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2cd50c473c80f6d7c3670a752354b8e569b1a7cbfdc0419ec88e5edad85e0dc7"
+checksum = "92d699e522242e69e3003b94ecc1f960f3a5e015aa7c5d7486e65ad01dd94f5e"

 [[package]]
 name = "futures-task"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b231ed28831efb4a61a08580c4bc233ec56bc009f4cd8f52da2c3cb97df0c109"
+checksum = "cd417de3d1d015fc3bfd2b1ea46dfc7bab72ef86f1cc7cc9c78e728b34a6d1fd"

 [[package]]
 name = "futures-util"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a77a90a256fce34da66415271e30f94ee91c57b04b8a2c042d9cf3220179deaa"
+checksum = "0d50a92467f8ba5dd6e3ee5d4bd04d73ab2e4e1c44474a0674821dfce14b79bc"
 dependencies = [
 "futures-core",
 "futures-task",
@@ -87,9 +87,9 @@ dependencies = [

 [[package]]
 name = "js-sys"
-version = "0.3.103"
+version = "0.3.104"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "53b44bfcdb3f8d5837a46dae1ca9660a837176eee74a28b229bc626816589102"
+checksum = "0e0c1080212aad755ea003d18543e8768dd432c48819efd73a7bf1e39b7a5a3a"
 dependencies = [
 "cfg-if",
 "futures-util",
@@ -110,9 +110,9 @@ checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"

 [[package]]
 name = "owo-colors"
-version = "4.3.0"
+version = "4.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d211803b9b6b570f68772237e415a029d5a50c65d382910b879fb19d3271f94d"
+checksum = "13c45bb4a6ae1280ec0803b1ef9d3455eb50f01efbbe1447ab020f1d54fba9d8"

 [[package]]
 name = "pin-project-lite"
@@ -122,9 +122,9 @@ checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"

 [[package]]
 name = "portable-atomic"
-version = "1.14.0"
+version = "1.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d20d5497ef88037a52ff98267d066e7f11fcc5e99bbfbd58a42336193aacec3"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"

 [[package]]
 name = "proc-macro2"
@@ -160,24 +160,46 @@ dependencies = [
 ]

 [[package]]
-name = "senbei"
-version = "1.0.0"
+name = "senbei-crypto"
+version = "1.1.0"
+dependencies = [
+ "thiserror",
+]
+
+[[package]]
+name = "senbei-io"
+version = "1.1.0"
 dependencies = [
 "anyhow",
 "indicatif",
 "libc",
 "owo-colors",
- "thiserror",
+ "senbei-metadata",
+ "senbei-pe",
 "walkdir",
 "windows",
 ]

 [[package]]
-name = "senbei-web"
-version = "1.0.0"
+name = "senbei-metadata"
+version = "1.1.0"
+
+[[package]]
+name = "senbei-pe"
+version = "1.1.0"
+dependencies = [
+ "senbei-crypto",
+ "thiserror",
+]
+
+[[package]]
+name = "senbei-wasm"
+version = "1.1.0"
 dependencies = [
 "console_error_panic_hook",
- "senbei",
+ "senbei-io",
+ "senbei-metadata",
+ "senbei-pe",
 "wasm-bindgen",
 ]

@@ -200,9 +222,9 @@ dependencies = [

 [[package]]
 name = "syn"
-version = "3.0.3"
+version = "3.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "53e9bae58849f64dfa4f5d5ae372c8341f7305f82a3868709269343628b659a3"
+checksum = "e6275cddf4610d1775e6d1fe9469b2e77d0f39fd98fb7450901b821e0c53649f"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -211,22 +233,22 @@ dependencies = [

 [[package]]
 name = "thiserror"
-version = "2.0.19"
+version = "2.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09a43598840e33d5b0331f38c5e30d13bb11c11210a4b58f0d9b18a5a5eefcd9"
+checksum = "ec86235f5fcc2a73650310756d2ac5b138a5780bbbdfae3eeccec992c435ba4f"
 dependencies = [
 "thiserror-impl",
 ]

 [[package]]
 name = "thiserror-impl"
-version = "2.0.19"
+version = "2.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "43cbfe0cf76104d42a574802844187e84a305e531ed54455f11fbde0f10541cd"
+checksum = "bc04cd3e1236dd4a98afca4569f2deb3f120e5422a4023be2cb683f8486292af"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 3.0.3",
+ "syn 3.0.4",
 ]

 [[package]]
@@ -259,9 +281,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4b067c0c11094aef6b7a801c1e34a26affafdf3d051dba08456b868789aaf9a4"
+checksum = "1b70935747edd64d89de3efa29d73789b806c15798f8e7dca4d8ac356b50ce70"
 dependencies = [
 "cfg-if",
 "once_cell",
@@ -272,9 +294,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "167ce5e579f6bcf889c4f7175a8a5a585de84e8ff93976ce393efa5f2837aab1"
+checksum = "77775f8f3f7217702089053b94958f8f54061a3f663417df76e19cbdcca29bc1"
 dependencies = [
 "quote",
 "wasm-bindgen-macro-support",
@@ -282,9 +304,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f3997c7839262f4ef12cf90b818d6340c18e80f263f1a94bf157d0ec4420380e"
+checksum = "e11d33f857dc2fb11b8bc75aee111aa9cbeb12cd9f25efd3d4c2a3dd4e235284"
 dependencies = [
 "bumpalo",
 "proc-macro2",
@@ -295,9 +317,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.126"
+version = "0.2.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc1b4cb0cc549fcf58d7dfc081778139b3d283a081644e833e84682ad71cea24"
+checksum = "7ef64dbcc55df09c7e5a46182d181c2cfa3e925f3da937ea764728b4bbb9dcbf"
 dependencies = [
 "unicode-ident",
 ]
@@ -0,0 +1,21 @@
+[package]
+name = "senbei-wasm"
+version = "1.1.0"
+edition = "2024"
+description = "WebAssembly bindings for senbei (browser frontend assets live in web/)"
+license = "AGPL-3.0-only"
+
+[lib]
+crate-type = ["cdylib"]
+
+[dependencies]
+senbei-io = { path = "../senbei-io" }
+senbei-metadata = { path = "../senbei-metadata" }
+senbei-pe = { path = "../senbei-pe" }
+wasm-bindgen = "0.2"
+console_error_panic_hook = "0.1"
+
+[profile.release]
+opt-level = "z"
+lto = true
+codegen-units = 1
@@ -0,0 +1,662 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.
+
@@ -24,7 +24,8 @@ pub struct UnpackResult {

 #[wasm_bindgen]
 impl UnpackResult {
-    /// Detected module kind: `"exe"`, `"native-dll"`, or `"managed-dll"`.
+    /// Detected module kind: `"native-exe"`, `"managed-exe"`, `"native-dll"`,
+    /// or `"managed-dll"`.
    #[wasm_bindgen(getter)]
    pub fn kind(&self) -> String {
        self.kind.clone()
@@ -100,24 +101,26 @@ impl MetadataResult {
    }
 }

-fn kind_str(kind: senbei::unpacker::Kind) -> &'static str {
+fn kind_str(kind: senbei_pe::Kind) -> &'static str {
    match kind {
-        senbei::unpacker::Kind::Exe => "exe",
-        senbei::unpacker::Kind::NativeDll => "native-dll",
-        senbei::unpacker::Kind::ManagedDll => "managed-dll",
+        senbei_pe::Kind::NativeExe => "native-exe",
+        senbei_pe::Kind::ManagedExe => "managed-exe",
+        senbei_pe::Kind::NativeDll => "native-dll",
+        senbei_pe::Kind::ManagedDll => "managed-dll",
    }
 }

 /// Classify a file's bytes without unpacking.
 ///
-/// Returns `"exe"`, `"native-dll"`, `"managed-dll"`, `"metadata"` (an il2cpp
-/// `global-metadata.dat`), or `undefined` for anything unrecognized.
+/// Returns `"native-exe"`, `"managed-exe"`, `"native-dll"`, `"managed-dll"`,
+/// `"metadata"` (an il2cpp `global-metadata.dat`), or `undefined` for
+/// anything unrecognized.
 #[wasm_bindgen]
 pub fn detect(input: &[u8]) -> Option<String> {
-    if senbei::metadata::is_metadata(input) {
+    if senbei_metadata::is_metadata(input) {
        return Some("metadata".to_string());
    }
-    senbei::unpacker::detect(input).map(|d| kind_str(d.kind).to_string())
+    senbei_pe::detect(input).map(|d| kind_str(d.kind).to_string())
 }

 /// Unpack a protected module.
@@ -131,7 +134,7 @@ pub fn unpack_file(
    input: &[u8],
    companion: Option<Vec<u8>>,
 ) -> Result<UnpackResult, JsError> {
-    let r = senbei::job::unpack_bytes(input, companion.as_deref())
+    let r = senbei_io::job::unpack_bytes(input, companion.as_deref())
        .map_err(|e| JsError::new(&e.to_string()))?;
    Ok(UnpackResult {
        kind: kind_str(r.kind).to_string(),
@@ -150,7 +153,7 @@ pub fn unpack_file(
 #[wasm_bindgen]
 pub fn deobfuscate_metadata(data: &[u8]) -> Result<MetadataResult, JsError> {
    let (bytes, report) =
-        senbei::metadata::deobfuscate(data).map_err(|e| JsError::new(&e.to_string()))?;
+        senbei_metadata::deobfuscate(data).map_err(|e| JsError::new(&e.to_string()))?;
    Ok(MetadataResult {
        bytes,
        version: report.version,
@@ -161,14 +164,14 @@ pub fn deobfuscate_metadata(data: &[u8]) -> Result<MetadataResult, JsError> {
 }

 /// Unpack a protected module, forcing the EXE pipeline (no DLL-pipeline
-/// probe). See [`senbei::job::unpack_bytes_force_exe`] for why the web app
+/// probe). See [`senbei_io::job::unpack_bytes_force_exe`] for why the web app
 /// needs this recovery path.
 #[wasm_bindgen]
 pub fn unpack_file_force_exe(
    input: &[u8],
    companion: Option<Vec<u8>>,
 ) -> Result<UnpackResult, JsError> {
-    let r = senbei::job::unpack_bytes_force_exe(input, companion.as_deref())
+    let r = senbei_io::job::unpack_bytes_force_exe(input, companion.as_deref())
        .map_err(|e| JsError::new(&e.to_string()))?;
    Ok(UnpackResult {
        kind: kind_str(r.kind).to_string(),
@@ -1,210 +0,0 @@
-//! Pure, panic-free Crackproof unpacker core. No file I/O lives here.
-
-mod bytecode;
-mod crc32;
-pub mod dll;
-pub mod exe;
-pub mod integrity;
-pub(crate) mod parallel;
-pub(crate) mod primitives;
-mod tables;
-
-pub use dll::{unpack_dll, unpack_dll_v};
-pub use exe::{UnpackError, unpack as unpack_exe, unpack_v as unpack_exe_v};
-pub use integrity::{IntegrityReport, check as check_integrity};
-
-/// Maximum plausible PE `SizeOfImage` we are willing to allocate a zero buffer
-/// for. Guards against a corrupt/crafted header requesting a multi-gigabyte
-/// (or, as a sign-extended negative `i32`, multi-exabyte) allocation, which
-/// would abort the process — an abort that `catch_unpack` below cannot trap.
-/// Real protected binaries are far below this.
-pub(crate) const MAX_IMAGE_SIZE: u64 = 1 << 30; // 1 GiB
-
-/// Run an unpack pipeline, converting any internal panic into a clean
-/// [`UnpackError::Corrupt`] so the public API stays panic-free on any input
-/// (truncated/garbled files chase offsets out of bounds). The default panic
-/// hook is suppressed transiently so a trapped panic does not spill a
-/// backtrace to stderr.
-///
-/// Note: allocation *failures* abort the process and are NOT caught here; size
-/// requests are bounds-checked against [`MAX_IMAGE_SIZE`] before allocating.
-pub(crate) fn catch_unpack<F>(f: F) -> Result<Vec<u8>, UnpackError>
-where
-    F: FnOnce() -> Result<Vec<u8>, UnpackError>,
-{
-    // Hook suppression is skipped on wasm: the prebuilt std cannot unwind
-    // there, so a panic traps immediately — and the suppressed hook would
-    // hide the panic message, leaving a bare `unreachable` with no clue.
-    #[cfg(not(target_arch = "wasm32"))]
-    let prev = std::panic::take_hook();
-    #[cfg(not(target_arch = "wasm32"))]
-    std::panic::set_hook(Box::new(|_| {}));
-    let r = std::panic::catch_unwind(std::panic::AssertUnwindSafe(f));
-    #[cfg(not(target_arch = "wasm32"))]
-    std::panic::set_hook(prev);
-    r.unwrap_or(Err(UnpackError::Corrupt))
-}
-
-/// Crackproof header magic stored in `keys[1]`/`info[1]`.
-pub(crate) const MAGIC_KONN: u32 = 0x4E4E4F4B; // b"KONN" little-endian (= 1313754955)
-
-/// True if `magic` is the Crackproof magic this unpacker supports.
-pub(crate) fn is_supported_magic(magic: u32) -> bool {
-    magic == MAGIC_KONN
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum Kind {
-    Exe,
-    NativeDll,
-    ManagedDll,
-}
-
-#[derive(Debug, Clone, Copy)]
-pub struct Detected {
-    pub kind: Kind,
-    pub magic: u32,
-}
-
-// ---------------------------------------------------------------------------
-// Content-based detection
-// ---------------------------------------------------------------------------
-
-/// Derive the 8-element Crackproof key table from the header at offset 4096.
-/// Returns `None` if the input is too short or doesn't have a valid PE signature.
-fn key_table(input: &[u8]) -> Option<[u32; 8]> {
-    // Need at least 4128 bytes: the key-table loop below reads dwords up to
-    // offset 4124 (bytes 4124..4127). Guarding only `< 4096` would let a
-    // 4096..4127-byte PE (e.g. a 4 KiB stub) panic in `get_u32`.
-    if input.len() < 4128 {
-        return None;
-    }
-    // Validate PE signature. `checked_add`, not `+`: `usize` is 32-bit on
-    // wasm32, where an `e_lfanew` of 0xFFFF_FFFC..=0xFFFF_FFFF wraps the bound
-    // check, and the slice below then panics with start > end. `detect` runs on
-    // the folder-scan threads and (in the web app) on the main thread outside
-    // the disposable-worker isolation, so it must not panic on any input.
-    let e_lfanew = primitives::get_u32(input, 0x3C);
-    let pe_start = e_lfanew as usize;
-    if pe_start.checked_add(4).is_none_or(|end| end > input.len()) {
-        return None;
-    }
-    if &input[pe_start..pe_start + 4] != b"PE\0\0" {
-        return None;
-    }
-    // Derive 8 keys per the Crackproof header-key formula.
-    let mut keys = [0u32; 8];
-    keys[0] = primitives::get_u32(input, 4096);
-    let mut k = keys[0];
-    for i in 0u32..7 {
-        let cell = primitives::get_u32(input, 4100u32.wrapping_add(i.wrapping_mul(4)));
-        keys[(i + 1) as usize] = k ^ cell;
-        k = i.wrapping_mul(i) ^ (k.wrapping_add(cell).wrapping_sub(i));
-    }
-    Some(keys)
-}
-
-/// Detect whether `input` is a Crackproof-protected binary and classify it.
-/// Returns `None` if the magic doesn't match.
-///
-/// Routing: `keys[1]` must be the Crackproof magic (`KONN`).
-/// The PE IMAGE_FILE_DLL characteristic distinguishes EXE vs DLL;
-/// the CLR data-directory RVA further distinguishes ManagedDll from NativeDll.
-pub fn detect(input: &[u8]) -> Option<Detected> {
-    let keys = key_table(input)?;
-    let magic = keys[1];
-    // Anything whose magic doesn't match is left untouched rather than
-    // detected-then-errored, honoring the "anything that doesn't match is
-    // left untouched" contract.
-    if !is_supported_magic(magic) {
-        return None;
-    }
-    // Use the PE DLL characteristic to distinguish EXE from DLL.
-    // IMAGE_FILE_HEADER.Characteristics is at peOff+4+18; bit 0x2000 = IMAGE_FILE_DLL.
-    let pe_off = primitives::get_u32(input, 0x3C);
-    let chars_offset = pe_off.wrapping_add(4).wrapping_add(18);
-    if (chars_offset as usize)
-        .checked_add(2)
-        .is_none_or(|end| end > input.len())
-    {
-        return None;
-    }
-    let chars =
-        (input[chars_offset as usize] as u16) | ((input[chars_offset as usize + 1] as u16) << 8);
-    let is_dll = (chars & 0x2000) != 0;
-    if !is_dll {
-        return Some(Detected {
-            kind: Kind::Exe,
-            magic,
-        });
-    }
-    // DLL: determine managed vs native via CLR data-directory RVA.
-    // peOff + 24 = start of optional header. The data directories start at a
-    // magic-dependent offset within it: PE32 (0x10B) at +96, PE32+ (0x20B) at
-    // +112. Using the PE32+ offset on a PE32 image reads the wrong dword and
-    // can mis-flag a native DLL as managed.
-    //
-    // `get_u16`/`get_u32` index unchecked, so every read past the already-
-    // checked Characteristics word must be bounds-checked first: a truncated
-    // DLL (e.g. `e_lfanew` pointing at len-24) would otherwise panic here,
-    // and this detector runs on the folder scan threads where a panic aborts
-    // the whole run.
-    let opt_magic_off = pe_off.wrapping_add(24) as usize;
-    let b = input.get(opt_magic_off..opt_magic_off.checked_add(2)?)?;
-    let opt_magic = u16::from_le_bytes([b[0], b[1]]);
-    let dd_off: u32 = if opt_magic == 0x20B { 112 } else { 96 };
-    // + 14*8 = IMAGE_DIRECTORY_ENTRY_COM_DESCRIPTOR
-    let clr_rva_offset = pe_off
-        .wrapping_add(24)
-        .wrapping_add(dd_off)
-        .wrapping_add(14u32.wrapping_mul(8));
-    if (clr_rva_offset as usize)
-        .checked_add(4)
-        .is_none_or(|end| end > input.len())
-    {
-        return None;
-    }
-    let clr_rva = primitives::get_u32(input, clr_rva_offset);
-    let kind = if clr_rva != 0 {
-        Kind::ManagedDll
-    } else {
-        Kind::NativeDll
-    };
-    Some(Detected { kind, magic })
-}
-
-/// Detect the file type and dispatch to the matching pipeline.
-/// Returns the detected `Kind` together with the unpacked image bytes.
-pub fn unpack_auto(input: &[u8]) -> Result<(Kind, Vec<u8>), UnpackError> {
-    unpack_auto_v(input, false)
-}
-
-/// Like [`unpack_auto`], but prints detailed `[N/9]` unpack-step progress to
-/// stdout when `verbose` is true. Output bytes are identical regardless.
-pub fn unpack_auto_v(input: &[u8], verbose: bool) -> Result<(Kind, Vec<u8>), UnpackError> {
-    let detected = detect(input).ok_or(UnpackError::NotCrackproof)?;
-    let out = match detected.kind {
-        Kind::Exe => unpack_exe_v(input, verbose)?,
-        Kind::NativeDll | Kind::ManagedDll => {
-            // Two Crackproof DLL layouts exist. The older one (the byte-identical
-            // DLL goldens) follows the pipeline in `dll.rs`. Newer builds protect
-            // DLLs with the EXE-style shell layout instead — `dll::unpack_dll`
-            // cannot parse them and errors. Try the DLL pipeline first; on
-            // failure, fall back to the EXE pipeline, which handles the new
-            // layout (including managed-DLL CLR metadata restore). The DLL-first
-            // order keeps the old-layout goldens byte-identical (the EXE
-            // pipeline "succeeds" on them but with different bytes).
-            match dll::unpack_dll_v(input, verbose) {
-                Ok(out) => out,
-                Err(dll_err) => match exe::unpack_v(input, verbose) {
-                    Ok(out) => out,
-                    // Surface the DLL-pipeline error, not the EXE one: for a
-                    // genuinely corrupt DLL the DLL error is the more relevant
-                    // diagnostic, and the EXE fallback is best-effort.
-                    Err(_) => return Err(dll_err),
-                },
-            }
-        }
-    };
-    Ok((detected.kind, out))
-}
@@ -1,10 +0,0 @@
-//! Shared test fixtures.
-#![allow(dead_code)]
-
-use std::path::PathBuf;
-
-/// Path to `senbei/samples` — the user-managed corpus dropped in by hand.
-/// Git-ignored except its README; tests here run against whatever is present.
-pub fn samples_dir() -> PathBuf {
-    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("samples")
-}
@@ -1,19 +0,0 @@
-[package]
-name = "senbei-web"
-version = "1.0.0"
-edition = "2024"
-description = "WebAssembly browser frontend for senbei"
-license = "AGPL-3.0-only"
-
-[lib]
-crate-type = ["cdylib"]
-
-[dependencies]
-senbei = { path = ".." }
-wasm-bindgen = "0.2"
-console_error_panic_hook = "0.1"
-
-[profile.release]
-opt-level = "z"
-lto = true
-codegen-units = 1
@@ -46,8 +46,8 @@ including the `wasm32-unknown-unknown` target) and
 [wasm-pack](https://rustwasm.github.io/wasm-pack/installer/).

 ```cmd
-cd web
-wasm-pack build --target web --release
+cd senbei-wasm
+wasm-pack build --target web --release --out-dir ../web/pkg
 ```

 This produces `web/pkg/` (git-ignored). Then serve the `web/` directory with
@@ -63,13 +63,14 @@ python -m http.server -d web 8000
 ## Layout

 ```
+senbei-wasm/   the senbei-wasm cdylib crate (own Cargo.lock, outside the
+               workspace; depends on the senbei-pe/-io/-metadata crates)
+└── src/lib.rs #[wasm_bindgen] bindings: detect / unpack_file /
+               unpack_file_force_exe / deobfuscate_metadata
 web/
-├── Cargo.toml     senbei-web cdylib crate (depends on the senbei lib)
-├── src/lib.rs     #[wasm_bindgen] bindings: detect / unpack_file /
-│                  unpack_file_force_exe / deobfuscate_metadata
 ├── index.html the page
 ├── app.js     dropzone, file list, worker orchestration, downloads
 ├── worker.js  one-shot unpack worker (fresh wasm instance per file)
 ├── style.css
-└── pkg/           wasm-pack output (git-ignored)
+└── pkg/       wasm-pack output (git-ignored; build from senbei-wasm/)
 ```
@@ -1,4 +1,4 @@
-import init, { detect, deobfuscate_metadata } from './pkg/senbei_web.js';
+import init, { detect, deobfuscate_metadata } from './pkg/senbei_wasm.js';

 const dropzone = document.getElementById('dropzone');
 const picker = document.getElementById('picker');
@@ -45,7 +45,8 @@ const files = new Map();
 const rowEls = new Map();

 const KIND_LABEL = {
-  exe: 'protected EXE',
+  'native-exe': 'protected native EXE',
+  'managed-exe': 'protected managed EXE',
  'native-dll': 'protected native DLL',
  'managed-dll': 'protected managed DLL',
  metadata: 'il2cpp metadata',
@@ -323,7 +324,8 @@ async function unpackModule(name, entry) {
  let comp = compEntry ? await read(compEntry.file) : undefined;
  let r = await runUnpack(input, comp, false);

-  if (!r.ok && r.trap && entry.kind !== 'exe') {
+  const isExe = entry.kind === 'native-exe' || entry.kind === 'managed-exe';
+  if (!r.ok && r.trap && !isExe) {
    // The DLL-routing probe trapped (panics can't be caught in wasm). Retry
    // once with the forced-EXE pipeline in a fresh worker — this mirrors the
    // CLI's dll-first/exe-fallback outcome for EXE-shell-layout DLLs.
@@ -10,7 +10,7 @@
 //    instance is never reused). That reproduces the CLI's
 //    dll-first/exe-fallback routing without a catchable panic.

-import init, { unpack_file, unpack_file_force_exe } from './pkg/senbei_web.js';
+import init, { unpack_file, unpack_file_force_exe } from './pkg/senbei_wasm.js';

 let ready = null;
Author	SHA1	Message	Date
Momoko-Ayase	0c3f29f93a	Split web/ into senbei-wasm crate + static assets The Rust bindings move from web/src to a top-level senbei-wasm crate (still outside the workspace, own Cargo.lock), matching the other senbei-* crates. web/ keeps only the static frontend; wasm-pack emits the JS/wasm package into web/pkg/ via --out-dir. JS glue renamed senbei_web -> senbei_wasm with the crate.	2026-08-30 22:44:15 +08:00
Momoko-Ayase	f80749ffd6	Merge bfloat16-senbei workspace restructure, bump to 1.1.0 Adopts the fork's workspace split (senbei-cli / senbei-crypto / senbei-io / senbei-metadata / senbei-pe), its structured error taxonomy, entry-transform and layout validation, PE32 dd8 key-formula selection with a skip floor, the CRT entry-stub dd8 oracle, and the extensionless-file scan skip. Kept from senbei on top of the restructure: - ManagedExe detection/routing and the CLR (COR20 + BSJB) metadata restore in the EXE pipeline. - The RET+int3 padding fingerprint as the primary dd8 padding signal, ahead of the mutated-position 0xCC fallback. - docs/, .github/, samples/, tests/ (moved to senbei-cli/tests), and the web/ wasm frontend (rewired to the split crates), all of which the fork had dropped. - The fork's README compatibility matrix is not taken: it names real games, which the public-repo hygiene rules forbid. - The wasm32 localtime fallback in logfile and unpack_bytes_force_exe (the web app's trap-recovery entry point), both lost in the restructure. Golden corpus: 35/35 byte-identical. clippy -D warnings clean; wasm32 check clean for the full workspace.	2026-08-30 22:31:21 +08:00
Momoko-Ayase	ef13419b93	Bump version to 1.0.1	2026-08-16 05:51:39 +08:00
Momoko-Ayase	dc6e72a8bb	Split the Exe kind into NativeExe / ManagedExe The detector already used the CLR data-directory RVA to split DLLs into NativeDll / ManagedDll; EXEs were a single undifferentiated Exe kind. Apply the same CLR check to EXEs so callers can tell a protected .NET executable from a native one without unpacking. Routing is unchanged: both EXE kinds go to the EXE pipeline. - CLI per-file lines and the run log now print NativeExe / ManagedExe (the kind comes from the same Debug formatting as the DLL variants). - The web API's detect()/unpack_file() kind strings become 'native-exe' / 'managed-exe'; the web UI gains matching labels, and the trap-retry guard (DLL-probe recovery) keys off both EXE kinds. Golden corpus unchanged (35/35 byte-identical); kind is classification only and never affects output bytes.	2026-08-16 05:47:16 +08:00
Momoko-Ayase	fe4f904409	Strengthen the dd8 shift detector with a RET+int3 padding fingerprint The selector replayed each candidate shift over three sample pages and counted 0xCC bytes at dd8-mutated positions. That signal is biased upward for wrong candidates (255 pseudo-random XORs manufacture ~1 spurious 0xCC per page for free) and cannot express 'this candidate destroys real padding', so the decision leaned on a 2x-margin-plus-floor rule tuned around the noise. Score candidates instead by a structural fingerprint of real x64 code: the MSVC function-end padding pattern (a 0xC3 RET followed by a run of >= 4 0xCC bytes), summed over up to four sample pages taken with a head/tail margin. Because dd8 touches only 255 of 4096 bytes per page, an encrypted page keeps most runs under 'no dd8' and restores them only under the correct shift, while an already-plaintext page loses runs under any shift — wrong candidates score below the baseline, which the old count could never say. The gate becomes a positive delta over the baseline (floor 8) instead of the 2x margin. When every candidate's fingerprint is silent (sampled pages with no padded epilogues), fall back to the previous mutated-position count with its 2x-margin-plus-floor rule, so pages without padding still resolve. Across the 35-input golden corpus every decision now comes from the fingerprint with wide, sign-correct margins; all outputs are unchanged (byte-identical goldens).	2026-08-16 05:33:44 +08:00
Momoko-Ayase	7f143b3b27	Support managed (CLR) EXEs in the old-layout EXE pipeline Managed EXE builds differ from their native counterparts in the old layout: the encrypted import-name table pointer is null (their imports are just the CLR bootstrap stub), the config block's entry-point field is 0, and the COR20 header / BSJB metadata stream / CLR resources are stored verbatim in the protected file rather than arriving through the section-block pass. - Skip the import-string walk when the table pointer is null instead of chasing header garbage as a pointer chain (previously a caught out-of-bounds panic reported as corrupt input). - Keep the protected header's entry point when the config block stores 0, instead of overwriting it with 0. - Restore the COR20 header, BSJB metadata stream, and CLR resources verbatim from the protected file after the .text dd8 pass, and suppress the native COR20-directory clearing when the restore ran. Validated by decompiling the unpacked managed EXEs with ilspycmd: full assemblies (types, methods, IL bodies) decompile cleanly. Golden corpus unchanged (35/35 byte-identical).	2026-08-16 05:33:44 +08:00
bfloat16	763bdbb21f	docs(readme): organize compatibility matrix by game	2026-08-13 10:49:44 +08:00
bfloat16	c5982a64b9	fix(unpacker): classify checksum range failures	2026-08-13 10:49:07 +08:00
bfloat16	55a31a2371	doc(none): add README.md	2026-08-12 23:57:00 +08:00
bfloat16	caadbd5325	fix(unpacker): validate executable entry transforms	2026-08-11 19:24:45 +08:00
bfloat16	ab1a14c9d1	perf(scan): skip extensionless files by default	2026-08-11 19:24:32 +08:00
bfloat16	e9ead4dc5f	refactor: init	2026-08-11 14:19:56 +08:00
bfloat16	a89900a812	fix(unpacker): refine layout validation and diagnostics	2026-08-11 11:42:12 +08:00
bfloat16	67178d34af	feat: Optimize error reporting	2026-08-11 10:51:37 +08:00