path: root/_posts/2026-10-01-quine.md

---
layout: post
title: 如何制作一个“完整”的博客压缩包？
tags: [压缩包, Quine, LZMA2, 7-Zip]
---

  让AI做出真正的“完整”将不再是难事……<!--more-->   

# 起因
  在上次[用AI制作了各种东西](/2026/09/01/vibe-coding2.html)之后，我已经完全理解了AI确实是无所不能的。既然如此，那就让它帮我解决曾经未能解决的事情吧？   
  去年，我为了让下载全站压缩包的按钮不断链，让这个压缩包也包含它本身而研究了[ZIP Quine](/2025/09/01/quine.html)，但限于DEFLATE的回溯窗口大小没能做到……但那是人做的东西，人还是太弱小了，现在换AI来试试，也许一切将变得不一样？   

# 制作基于LZMA2的博客压缩包
  首先，我把[Ruben Van Mello](https://github.com/ruvmello)写的那篇论文《[A Generator for Recursive Zip Files](https://www.mdpi.com/2076-3417/14/21/9797)》以及生成器[zip-quine-generator](https://github.com/ruvmello/zip-quine-generator)发给了AI，问它这个限制是不是真的，是不是真的无法创造出超过32KiB的ZIP Quine？没过多久它分析完了，告诉我确实存在这样的问题，但并不是无法解决的，最简单的办法就是换个回溯窗口更大的压缩算法。我看了一下它给我列出的几个算法，看起来LZMA2和zstd比较符合要求。不过zstd感觉不是很知名，一般的解压软件应该处理不了，所以我就选择让它基于LZMA2来制作了。   
  不过LZMA2只是算法，还得选个容器，考虑到7z算是最出名，而且LZMA2本来就是7-Zip的作者开发的，所以就直接让它做7z版本的了。另外因为[Nate Choe](https://github.com/NateChoe1)做了通过扩展欧几里得求逆元来计算CRC32的[PR](https://github.com/ruvmello/zip-quine-generator/pull/3)，所以我也告诉它要用这种方法在多项式时间内解出CRC32的值，而不是爆破。就这样AI花了两个小时左右成功把代码写出来了，效果非常完美，另外它还偷懒地将计算CRC32的算法改成了高斯消元法😆，因为7z格式有3个相互影响的CRC32值，用扩展欧几里得会更复杂一点，不过对7z来说不写入文件的CRC32也不影响，只是会不校验罢了，至少它还是按我要求做了……但其实这个问题也并不是完全不能用扩展欧几里得算法，后来我又强烈要求了一下，AI还是给我写出来了。以下是代码片段，有兴趣的人可以参考一下（不过这个看起来写的就是很复杂，所以我自己只用了高斯消元法）：   
```python
RU_POLY = 0x104C11DB7          # x^32+x^26+x^23+...+1（含 x^32 首项）


def _find_probe(p):
    ret = 0
    for i in range(64):
        if (1 << i) & p:
            ret = i
    return ret


def _mul_raw(p1, p2):
    """多项式乘法（无模）。"""
    ret = 0
    probe = _find_probe(p1)
    for i in range(64):
        if p2 & (1 << i):
            assert probe + i < 64, "多项式乘法溢出"
            ret ^= p1 << i
    return ret


def _poly_divmod(dividend, divisor):
    """多项式长除法，返回 (商, 余)。"""
    probe = _find_probe(divisor)
    probe_bit = 1 << probe
    quot = 0
    rem = dividend
    for i in range(63 - probe, -1, -1):
        if rem & (probe_bit << i):
            quot |= 1 << i
            rem ^= divisor << i
    return quot, rem


def _mul(p1, p2, mod):
    """p1*p2 % mod。"""
    return _poly_divmod(_mul_raw(p1, p2), mod)[1]


def _xgcd(p1, p2):
    """扩展欧几里得：返回 (k1,k2,gcd)，满足 p1*k1+p2*k2=gcd。"""
    if _find_probe(p1) < _find_probe(p2):
        k1, k2, g = _xgcd(p2, p1)
        return k2, k1, g
    if p2 == 0:
        return p1, 0, p1
    q, r = _poly_divmod(p1, p2)
    c1, c2, g = _xgcd(p2, r)
    return c2, c1 ^ _mul_raw(c2, q), g


def _minv(p, mod):
    """p 在环上的乘法逆元；不存在时返回 0。"""
    k1, _, g = _xgcd(p, mod)
    if g != 1:
        return 0
    return _poly_divmod(k1, mod)[1]


def _bitrev32(x):
    """32 位比特反转。CRC 的输出循环冗余（binascii 反射）↔ 本环（非反射）互为 bit-reverse。"""
    return int(format(x & 0xFFFFFFFF, "032b")[::-1], 2)


def _ring_pow_x(k):
    """环内 x^k mod P（P=RU_POLY），二进制幂 O(log k)。变量列就是单个环幂。"""
    res, base, mod, e = 1, 2, RU_POLY, k
    while e:
        if e & 1:
            res = _mul(res, base, mod)
        base = _mul(base, base, mod)
        e >>= 1
    return res


def solve_crc_system(files):
    """解析解 CRC 定点系统

    - 常数列     ：对“变量清零”的数据做一次 binascii.crc32，再 bit-reverse 32 位
                    （binascii 反射 CRC 与本环非反射 CRC 互为 bit-reverse，见 _bitrev32）。
    - 变量列     ：单个环幂 x^(32 + 8*(len-pos-4))，二进制幂 O(log N) 一步算出；
                    同一变量出现在多处时按位异或累加。
    各自 O(#rows) 次 CRC + 少量环幂即可建出整矩阵，Gauss 消元不变。
    """
    n = len(files)
    mod = RU_POLY
    matrix = [[0] * (n + 1) for _ in range(n)]

    for file in range(n):
        data = files[file][0]
        offsets = files[file][1]
        N = len(data)
        zero = bytearray(data)
        for pos in offsets:                      # 未知量占 4 字节，清零后求常数项
            zero[pos:pos + 4] = b"\x00\x00\x00\x00"
        matrix[file][n] = _bitrev32(crc32(bytes(zero)))
        for pos, vid in offsets.items():
            rest = N - pos - 4                    # 变量被 4 字节（x^32）+ 其后字节（x^8/个）推到底
            matrix[file][vid] ^= _ring_pow_x(32 + 8 * rest)
        matrix[file][file] ^= 1

    for sr in range(n):
        ex = sr
        while ex < n and matrix[ex][sr] == 0:
            ex += 1
        if ex == n:
            raise RuntimeError("CRC 多项式系统奇异 @row%d" % sr)
        if ex != sr:
            matrix[sr], matrix[ex] = matrix[ex], matrix[sr]
        inv = _minv(matrix[sr][sr], mod)
        for er in range(sr + 1, n):
            mult = _mul(inv, matrix[er][sr], mod)
            for ec in range(n + 1):
                matrix[er][ec] ^= _mul(mult, matrix[sr][ec], mod)

    res = [0] * n
    for row in range(n - 1, -1, -1):
        v = matrix[row][n]
        for col in range(row + 1, n):
            v ^= _mul(res[col], matrix[row][col], mod)
        res[row] = _mul(v, _minv(matrix[row][row], mod), mod)

    return [_impl_val(res[i]) for i in range(n)]

def _impl_val(math_val):
    imp = 0
    for bit in range(32):
        if math_val & (1 << bit):
            imp |= 1 << (31 - bit)
    return imp

class BlogQuine:
    def solve_crc(self, F, lay, seed):
        T, h, d = lay["T"], lay["h"], lay["d"]
        k, n_sub, crc_base = lay["k"], lay["n_sub"], lay["crc_base"]
        hoff = lay["total"] - h
        vout_f = lay["total"] - 2 * T
        hcopy = vout_f + (T - h)
        # 已知 CRC：seed + 各文件（quine 的是未知量 D）
        known = [crc32(seed)] + [crc32(data) for _, data in self.files]
        assert len(known) == n_sub - 1
        for i, e in enumerate(known):
            struct.pack_into("<I", F, hoff + crc_base + 4 * i, e)
            struct.pack_into("<I", F, hcopy + crc_base + 4 * i, e)
        # 未知量 D N S：D=整个文件、N=头部 F[hoff:hoff+h]、S=签名 F[12:32]
        dpos = crc_base + 4 * (n_sub - 1)
        ncopy = 32 + 3 * k + d                            # C1k 载荷里 F[0:35) 副本起点
        groups = [[hoff + dpos, hcopy + dpos],            # D 出现两处
                  [28, ncopy + 28],                        # N
                  [8, ncopy + 8]]                          # S

        # 构造 3 个“文件”喂给解析解系统：
        #   文件0 = 整个 F，三个未知量各自出现在 groups 两处
        #   文件1 = 头部块，只有 D 出现在相对 dpos（另一处 hcopy+dpos 在头外）
        #   文件2 = 签名块，只有 N 出现在相对 16（28-12）
        whole = bytes(F)
        hdr = bytes(F[hoff:hoff + h])
        sig = bytes(F[12:32])
        f0_off = {hoff + dpos: 0, hcopy + dpos: 0,
                  28: 1, ncopy + 28: 1,
                  8: 2, ncopy + 8: 2}
        f1_off = {dpos: 0}
        f2_off = {28 - 12: 1}
        vals = solve_crc_system([(whole, f0_off), (hdr, f1_off), (sig, f2_off)])

        for g in range(3):
            for p in groups[g]:
                F[p:p + 4] = struct.pack("<I", vals[g])
        real = (crc32(F), crc32(F[hoff:hoff + h]), crc32(F[12:32]))
        assert real == tuple(vals), "CRC 定点失败: %s != %s" % (real, vals)
        return vals
```

  于是我第一时间就把原来的TGZ压缩命令换掉，换成了AI给我写的[blogquine.py](https://github.com/Mabbs/mabbs.github.io/blob/master/_tools/blogquine.py)，现在就可以通过[这里](https://mayx.eu.org/MayxBlog.7z)下载到“完整”包含我博客所有内容的压缩包了。   
  不过唯一的问题就是这样做出来的压缩包并没有压缩😂，相当于给做成了普通的归档了。当然我的博客本身倒是不大，没压缩也多不了多少空间，但相比于能做出“完整”的效果来说，这点浪费的空间也是小问题了。   
## 对TXZ格式的尝试
  在做完7z格式的压缩包之后，我发现了一个问题，虽然7z确实很流行，但是在Linux下解压起来有点麻烦，7-Zip历史上主要面向Windows，Linux上长期以来更多依赖p7zip等第三方移植，因此生态集成度不如tar.xz，想要解压7z文件还得额外安装。   
  不过Linux下也有个支持LZMA2算法的压缩软件，那就是前些年出过[后门](https://tukaani.org/xz-backdoor/)的XZ Utils，配合tar就可以做出TXZ（tar.xz）文件，甚至用我[博客终端](https://mabbs.github.io/linux/)中的BusyBox也能解压。我想了一下反正有AI，干脆一句话让AI帮我把blogquine.py改成tar.xz格式的好了，结果倒也没费多少功夫，AI就这样写出来了：   

<details markdown="1">
<summary markdown="span">
Show Code
</summary>

```python
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import argparse
import binascii
import io
import lzma
import os
import struct
import sys
import tarfile
import time
import ctypes
import ctypes.util

# ============================================================
# Part 1: 最小 LZMA1 range coder（只做编码）
# ============================================================

kNumBitModelTotalBits = 11
kBitModelTotal = 1 << kNumBitModelTotalBits          # 2048
kNumMoveBits = 5
kTopValue = 1 << 24                                   # 0x1000000

kNumStates = 12
kNumLitStates = 7
kNumPosBitsMax = 4
kNumLenToPosStates = 4
kNumAlignBits = 4
kEndPosModelIndex = 14
kNumFullDistances = 1 << (kEndPosModelIndex >> 1)     # 128
kMatchMinLen = 2
kNumLowLenBits = 3
kNumMidLenBits = 3
kNumHighLenBits = 8
kNumLowLenSymbols = 1 << kNumLowLenBits               # 8
kNumMidLenSymbols = 1 << kNumMidLenBits               # 8
kNumPosSlotBits = 6

PROB_INIT = kBitModelTotal >> 1                       # 1024


class RangeEncoder:
    """LZMA 的区间编码器。low 是 64 位（要容纳进位），range 是 32 位。"""

    def __init__(self):
        self.low = 0
        self.range = 0xFFFFFFFF
        self.cache = 0
        self.cache_size = 1          # 初值 1 -> 第一个输出字节恒为 0x00
        self.buf = bytearray()

    def _shift_low(self):
        if (self.low & 0xFFFFFFFF) < 0xFF000000 or (self.low >> 32) != 0:
            temp = self.cache
            while True:
                self.buf.append((temp + (self.low >> 32)) & 0xFF)
                temp = 0xFF
                self.cache_size -= 1
                if self.cache_size == 0:
                    break
            self.cache = (self.low >> 24) & 0xFF
        self.cache_size += 1
        self.low = ((self.low & 0xFFFFFFFF) << 8) & 0xFFFFFFFF

    def encode_bit(self, probs, idx, bit):
        p = probs[idx]
        bound = (self.range >> kNumBitModelTotalBits) * p
        if bit == 0:
            self.range = bound
            probs[idx] = p + ((kBitModelTotal - p) >> kNumMoveBits)
        else:
            self.low += bound
            self.range -= bound
            probs[idx] = p - (p >> kNumMoveBits)
        while self.range < kTopValue:
            self.range = (self.range << 8) & 0xFFFFFFFF
            self._shift_low()

    def encode_direct_bits(self, value, num_bits):
        for i in range(num_bits - 1, -1, -1):
            self.range >>= 1
            if (value >> i) & 1:
                self.low += self.range
            while self.range < kTopValue:
                self.range = (self.range << 8) & 0xFFFFFFFF
                self._shift_low()

    def bittree_encode(self, probs, off, num_bits, symbol):
        m = 1
        for i in range(num_bits - 1, -1, -1):
            bit = (symbol >> i) & 1
            self.encode_bit(probs, off + m, bit)
            m = (m << 1) | bit

    def bittree_reverse_encode(self, probs, off, num_bits, symbol):
        m = 1
        for i in range(num_bits):
            bit = symbol & 1
            symbol >>= 1
            self.encode_bit(probs, off + m, bit)
            m = (m << 1) | bit

    def finish(self):
        for _ in range(5):
            self._shift_low()
        return bytes(self.buf)


def _pos_slot_and_bits(d):
    if d < 4:
        return d, 0, 0
    for slot in range(4, 64):
        n = (slot >> 1) - 1
        base = (2 | (slot & 1)) << n
        if d < base + (1 << n):
            return slot, n, d - base
    raise ValueError("distance too large: %d" % d)


# pb 必须与写进每个 LZMA2 chunk 头的 PROPS_BYTE(0x5D) 一致。
# lc/lp 只影响 literal 编码的概率表索引，而本编码器不实现 literal 路径，故无需常量。
PB = 2


class LzmaEncoder:
    """只实现 quine 用到的两条路径：match 与 rep0-match。

    literal / rep-g1 / rep-g2 三条分支从未被构造，因此它们对应的概率表
    （p_lit、p_is_rep_g1、p_is_rep_g2）以及 prev_byte 都不必存在。
    reps 同理：编码器从不回读（rep0 的距离由解码器自行维护），故不保存。
    """

    def __init__(self):
        self.pos_mask = (1 << PB) - 1
        self.rc = RangeEncoder()
        self.pos = 0
        self.state = 0
        n = PROB_INIT
        self.p_is_match = [n] * (kNumStates << kNumPosBitsMax)
        self.p_is_rep = [n] * kNumStates
        self.p_is_rep_g0 = [n] * kNumStates
        self.p_rep0_long = [n] * (kNumStates << kNumPosBitsMax)
        self.p_pos_slot = [n] * (kNumLenToPosStates << kNumPosSlotBits)
        self.p_spec_pos = [n] * (kNumFullDistances - kEndPosModelIndex)
        self.p_align = [n] * (1 << kNumAlignBits)
        self.p_len_choice = [n] * 2
        self.p_len_low = [n] * (16 * kNumLowLenSymbols)
        self.p_len_mid = [n] * (16 * kNumMidLenSymbols)
        self.p_len_high = [n] * (1 << kNumHighLenBits)
        self.p_rep_len_choice = [n] * 2
        self.p_rep_len_low = [n] * (16 * kNumLowLenSymbols)
        self.p_rep_len_mid = [n] * (16 * kNumMidLenSymbols)
        self.p_rep_len_high = [n] * (1 << kNumHighLenBits)

    @property
    def pos_state(self):
        return self.pos & self.pos_mask

    def _encode_len(self, choice, low, mid, high, length):
        ps = self.pos_state
        l = length - kMatchMinLen
        if l < kNumLowLenSymbols:
            self.rc.encode_bit(choice, 0, 0)
            self.rc.bittree_encode(low, ps << kNumLowLenBits, kNumLowLenBits, l)
        else:
            self.rc.encode_bit(choice, 0, 1)
            l -= kNumLowLenSymbols
            if l < kNumMidLenSymbols:
                self.rc.encode_bit(choice, 1, 0)
                self.rc.bittree_encode(mid, ps << kNumMidLenBits, kNumMidLenBits, l)
            else:
                self.rc.encode_bit(choice, 1, 1)
                self.rc.bittree_encode(high, 0, kNumHighLenBits, l - kNumMidLenSymbols)

    def _write_dist(self, dist, lts):
        d = dist - 1
        slot, n, low_bits = _pos_slot_and_bits(d)
        self.rc.bittree_encode(self.p_pos_slot, lts << kNumPosSlotBits,
                               kNumPosSlotBits, slot)
        if slot >= 4:
            if slot < kEndPosModelIndex:
                base = (2 | (slot & 1)) << n
                off = base - slot - 1
                self.rc.bittree_reverse_encode(self.p_spec_pos, off, n, low_bits)
            else:
                self.rc.encode_direct_bits(low_bits >> kNumAlignBits, n - kNumAlignBits)
                self.rc.bittree_reverse_encode(self.p_align, 0, kNumAlignBits,
                                               low_bits & ((1 << kNumAlignBits) - 1))

    def match(self, dist, length):
        assert kMatchMinLen <= length <= 273, length
        ps = self.pos_state
        self.rc.encode_bit(self.p_is_match, (self.state << kNumPosBitsMax) + ps, 1)
        self.rc.encode_bit(self.p_is_rep, self.state, 0)
        lts = min(length - kMatchMinLen, kNumLenToPosStates - 1)
        self._encode_len(self.p_len_choice,
                         self.p_len_low, self.p_len_mid, self.p_len_high, length)
        self._write_dist(dist, lts)
        self.state = 7 if self.state < kNumLitStates else 10
        self.pos += length

    def rep_match(self, length):
        assert kMatchMinLen <= length <= 273, length
        ps = self.pos_state
        self.rc.encode_bit(self.p_is_match, (self.state << kNumPosBitsMax) + ps, 1)
        self.rc.encode_bit(self.p_is_rep, self.state, 1)
        self.rc.encode_bit(self.p_is_rep_g0, self.state, 0)
        self.rc.encode_bit(self.p_rep0_long, (self.state << kNumPosBitsMax) + ps, 1)
        self._encode_len(self.p_rep_len_choice,
                         self.p_rep_len_low, self.p_rep_len_mid, self.p_rep_len_high, length)
        self.state = 8 if self.state < kNumLitStates else 11
        self.pos += length

    def finish(self):
        return self.rc.finish()


# ============================================================
# Part 2: tar + xz 格式原语
# ============================================================

MAX_MATCH = 273                       # LZMA1 单个 match 长度上限
CHUNK = 65536                          # LZMA2 单 chunk 解压上限
PROPS_BYTE = 0x5D                      # lc=3, lp=0, pb=2
SEED_NOTE = ("\nMayx's Blog!").encode("utf-8")

# xz 格式常量
XZ_MAGIC = b"\xFD7zXZ\x00"
XZ_FOOTER_MAGIC = b"YZ"
XZ_CHECK_CRC64 = 0x04
LZMA2_FILTER_ID = 0x21
CHECK_SIZE = 8  # CRC64 占 8 字节
_XZ_FLAGS = bytes([0x00, XZ_CHECK_CRC64])   # Stream Flags: reserved + check_type

# tar 格式常量
TAR_BLOCK = 512


def round_up_512(n):
    """长度按 tar 块 512 字节向上取整。"""
    return (n + TAR_BLOCK - 1) // TAR_BLOCK * TAR_BLOCK


def crc32(b):
    return binascii.crc32(b) & 0xFFFFFFFF


# ---------- CRC64 (ECMA-182, reflected) ----------

_CRC64_POLY = 0xC96C5795D7870F42  # reflected ECMA-182
_CRC64_INIT = 0xFFFFFFFFFFFFFFFF
_CRC64_XOROUT = 0xFFFFFFFFFFFFFFFF
_CRC64_NON_REFLECTED = 0x142F0E1EBA9EA3693  # 非反射多项式 (65 bit)
_CRC64_MOD_BIT = 1 << 64  # x^64 对应的 bit


_liblzma = None
try:
    _lzma_name = ctypes.util.find_library('lzma')
    if _lzma_name:
        _liblzma = ctypes.CDLL(_lzma_name)
        _liblzma.lzma_crc64.argtypes = [ctypes.c_char_p, ctypes.c_size_t, ctypes.c_uint64]
        _liblzma.lzma_crc64.restype = ctypes.c_uint64
        # 验证：CRC64("123456789") 应为 0x995dc9bbdf1939fa
        if _liblzma.lzma_crc64(b'123456789', 9, 0) != 0x995dc9bbdf1939fa:
            _liblzma = None
except (OSError, AttributeError):
    _liblzma = None


def _make_crc64_table():
    table = []
    for i in range(256):
        crc = i
        for _ in range(8):
            crc = (crc >> 1) ^ _CRC64_POLY if (crc & 1) else crc >> 1
        table.append(crc)
    return table


_CRC64_TABLE = _make_crc64_table()


def _crc64_pure(data):
    """纯 Python CRC64（ECMA-182 反射，init/xorout = 0xFFFF...FFFF）。"""
    crc = _CRC64_INIT
    for byte in data:
        crc = (crc >> 8) ^ _CRC64_TABLE[(crc ^ byte) & 0xFF]
    return crc ^ _CRC64_XOROUT


def _crc64_lib(data):
    """liblzma CRC64（init=0，内部处理 init/xorout）。"""
    return _liblzma.lzma_crc64(data, len(data), 0)


crc64 = _crc64_lib if _liblzma else _crc64_pure


# ---------- GF(2^64) 多项式运算 ----------
# 用于 CRC64 自引用求解。CRC64 的线性贡献可表示为
# bit_reverse(contribution(D)) = D_poly * P (mod G_non)
# 其中 D_poly = bit_reverse(D)，P 是位置决定的多项式。
# 方程 D_poly * (1 ^ P1 ^ P2) = bit_reverse(CRC64(W)) 通过多项式逆元求解。

def _bit_reverse64(v):
    r = 0
    for _ in range(64):
        r = (r << 1) | (v & 1)
        v >>= 1
    return r


def _poly_mul_mod(a, b):
    """GF(2) 多项式乘法 mod G（非反射）。"""
    result = 0
    while b:
        if b & 1:
            result ^= a
        b >>= 1
        a <<= 1
        if a & _CRC64_MOD_BIT:
            a ^= _CRC64_NON_REFLECTED
    return result


def _poly_pow(base, exp):
    """多项式幂：base^exp mod G。"""
    result = 1
    while exp:
        if exp & 1:
            result = _poly_mul_mod(result, base)
        base = _poly_mul_mod(base, base)
        exp >>= 1
    return result


def _poly_minv(p):
    """多项式逆元 p^(-1) mod G，用扩展欧几里得算法。"""
    if p == 0:
        return 0
    old_r, r = p, _CRC64_NON_REFLECTED
    old_s, s = 1, 0
    while r:
        # poly divmod(old_r, r) -> (q, rem)
        probe_r = r.bit_length() - 1
        rem = old_r
        q = 0
        while rem.bit_length() - 1 >= probe_r and rem:
            shift = rem.bit_length() - 1 - probe_r
            q |= 1 << shift
            rem ^= r << shift
        old_r, r = r, rem
        # s = old_s ^ q*s (raw 多项式乘法)
        prod = 0
        a, b = q, s
        while b:
            if b & 1:
                prod ^= a
            b >>= 1
            a <<= 1
        old_s, s = s, old_s ^ prod
    if old_r != 1:
        return 0
    # old_s mod G
    probe_m = _CRC64_NON_REFLECTED.bit_length() - 1
    rem = old_s
    while rem.bit_length() - 1 >= probe_m and rem:
        shift = rem.bit_length() - 1 - probe_m
        rem ^= _CRC64_NON_REFLECTED << shift
    return rem


def store_hdr(payload_len, first=False):
    """LZMA2 uncompressed chunk 头（3 字节，大端 size-1）。"""
    assert 1 <= payload_len <= 65536
    return bytes([0x01 if first else 0x02]) + (payload_len - 1).to_bytes(2, "big")


def matches_for(dist, total):
    """dist 固定、总长 total 的 token 序列：首个 match + 后续 rep0。"""
    assert total >= 2
    toks = []
    first = min(MAX_MATCH, total)
    if total - first == 1:
        first -= 1
    toks.append(("m", dist, first))
    total -= first
    while total > 0:
        l = min(MAX_MATCH, total)
        if total - l == 1:
            l -= 1
        toks.append(("r0", l))
        total -= l
    return toks


def _add_gear(chunks, o, f, q):
    """追加一组标准 gear：store 载荷 x = slip+3，紧跟一个 lzma 把那 x 字节复制一遍。

    lzma chunk 解压出的字节数正好等于 x，所以齿轮转完之后 slip (f - (o-q)) 精确
    归约为 len(cb)（很小的正数，x 被 CHUNK 截断时则按 CHUNK 缩减）。
    _layout_iterate 反复套用把 slip 压到 <= 16，_trim_layouts 用它做相位微调。
    返回追加后的 (o, f)。
    """
    x = min(f - (o - q) + 3, CHUNK)
    chunks.append({"kind": "store", "foff": f, "size": 3 + x,
                   "ooff": o, "olen": x})
    o += x
    f += 3 + x
    cb, cu = lzma_chunk(matches_for(x, x), o)
    chunks.append({"kind": "lzma", "foff": f, "size": len(cb),
                   "ooff": o, "olen": cu, "bytes": cb})
    return o + cu, f + len(cb)


def _forward_vac(F, vac, q):
    """vac_store：把紧随其后的等长字节块前移覆盖自身。"""
    fo = vac["ooff"] - q
    olen = vac["olen"]
    F[fo:fo + olen] = F[fo + olen:fo + 2 * olen]


def dict_prop_for(maxdist):
    """选最小的 LZMA2 dict prop 使字典 >= maxdist。"""
    for p in range(41):
        if (2 | (p & 1)) << (p // 2 + 11) >= maxdist:
            return p, (2 | (p & 1)) << (p // 2 + 11)
    raise ValueError("distance too large")


def lzma_chunk(tokens, out_pos):
    """编一个 LZMA chunk（0xC0: state+props reset，无 dict reset，无 end marker）。"""
    enc = LzmaEncoder()
    enc.pos = out_pos
    total = 0
    for t in tokens:
        if t[0] == "m":
            enc.match(t[1], t[2])
            total += t[2]
        else:
            enc.rep_match(t[1])
            total += t[1]
    data = enc.finish()
    assert total - 1 < 65536 and len(data) - 1 < 65536
    hdr = bytes([0xC0]) + (total - 1).to_bytes(2, "big") + \
          (len(data) - 1).to_bytes(2, "big") + bytes([PROPS_BYTE])
    return hdr + data, total


# ---------- tar 原语 ----------

def tar_header(name, size, mtime, mode=0o644, typeflag=b'0', uid=0, gid=0):
    """构建 512 字节 POSIX ustar tar header。
    注意：tar 数值字段包含 null 终止符，slice 范围必须精确匹配字段宽度。
    mode/uid/gid 各 8 字节（7 octal + NUL），size/mtime 各 12 字节（11 octal + NUL）。
    """
    h = bytearray(512)
    name_bytes = name.encode('utf-8')[:100]
    h[0:len(name_bytes)] = name_bytes
    h[100:108] = b'%07o\x00' % mode       # 8 bytes (field 100-107)
    h[108:116] = b'%07o\x00' % uid        # 8 bytes (field 108-115)
    h[116:124] = b'%07o\x00' % gid        # 8 bytes (field 116-123)
    h[124:136] = b'%011o\x00' % size      # 12 bytes (field 124-135)
    h[136:148] = b'%011o\x00' % mtime     # 12 bytes (field 136-147)
    h[148:156] = b'        '              # checksum placeholder (8 spaces)
    h[156:157] = typeflag
    h[257:263] = b'ustar\x00'
    h[263:265] = b'00'
    assert len(h) == 512, "header length %d != 512" % len(h)
    # 计算校验和：所有字节之和，checksum 字段视为空格
    chksum = sum(h) & 0o7777777
    h[148:156] = b'%06o\x00 ' % chksum    # 6 octal + NUL + space = 8 bytes
    assert len(h) == 512, "header length %d != 512 after checksum" % len(h)
    return bytes(h)


# ---------- xz 原语 ----------

def xz_varint(v):
    """xz 变长整数编码：每字节 7 位数据（LSB first），MSB 为续位。"""
    b = []
    while v >= 0x80:
        b.append((v & 0x7F) | 0x80)
        v >>= 7
    b.append(v & 0x7F)
    return bytes(b)


def xz_stream_header():
    """12 字节 xz stream header (CHECK_CRC64)。
    Stream Flags = [reserved(0x00), check_type(CRC64=0x04)]。
    """
    flags = _XZ_FLAGS
    crc = crc32(flags)
    return XZ_MAGIC + flags + struct.pack('<I', crc)


def xz_block_header(dict_size):
    """xz block header（单 LZMA2 filter）。
    返回 (header_bytes, 实际字典大小)；header 长度即 block header 总长度。
    """
    prop, real_dict = dict_prop_for(dict_size)
    # filter flags: filter_id(1) + props_size(1) + props(1=prop byte)
    filter_flags = bytes([LZMA2_FILTER_ID, 1, prop])
    # block header content: flags(1, 0x00=1 filter) + filter_flags
    content = bytes([0x00]) + filter_flags
    # 填充到 (1 + len(content) + 4) 是 4 的倍数
    total = 1 + len(content) + 4
    pad = (4 - total % 4) % 4
    content += b'\x00' * pad
    bh = 1 + len(content) + 4
    assert bh % 4 == 0, "bh not 4-aligned: %d" % bh
    crc = crc32(bytes([bh // 4 - 1]) + content)
    header = bytes([bh // 4 - 1]) + content + struct.pack('<I', crc)
    assert len(header) == bh
    return header, real_dict


def xz_index(records):
    """xz index。records = [(unpadded_size, uncompressed_size), ...]。"""
    data = bytes([0x00])  # index indicator
    data += xz_varint(len(records))
    for unpadded, uncompressed in records:
        data += xz_varint(unpadded)
        data += xz_varint(uncompressed)
    # 填充到 (len(data) + 4) 是 4 的倍数
    pad = (4 - (len(data) + 4) % 4) % 4
    data += b'\x00' * pad
    return data + struct.pack('<I', crc32(data))


def xz_stream_footer(backward_size):
    """12 字节 xz stream footer (CHECK_CRC64)。
    backward_size = (index_size / 4) - 1。
    Stream Flags = [reserved(0x00), check_type(CRC64=0x04)]。
    CRC32 覆盖 Backward Size + Stream Flags（6 字节，不含 CRC32 和 Footer Magic）。
    """
    flags = _XZ_FLAGS
    crc_data = struct.pack('<I', backward_size) + flags  # 6 bytes, no magic
    crc = crc32(crc_data)
    return struct.pack('<I', crc) + crc_data + XZ_FOOTER_MAGIC


# ============================================================
# Part 3: 多文件 tar.xz quine 构造器
# ============================================================

class TarXzQuine:
    def __init__(self, root, quine_name="quine.tar.xz", seed_name=".this_is_mayx_blog",
                 src_dir=""):
        self.quine_name = quine_name
        self.seed_name = seed_name
        self.src_dir = src_dir.strip("/")
        self.mtime = int(time.time())
        # walk 源目录
        self.walk_entries = []
        self.files = []
        if self.src_dir:
            parts = self.src_dir.split("/")
            for i in range(1, len(parts) + 1):
                self.walk_entries.append(("/".join(parts[:i]), True))
        for dirpath, dirnames, filenames in os.walk(root):
            dirnames.sort()
            for dn in dirnames:
                rel = os.path.relpath(os.path.join(dirpath, dn), root).replace(os.sep, "/")
                self.walk_entries.append((self._prefixed(rel), True))
            for fn in sorted(filenames):
                p = os.path.join(dirpath, fn)
                rel = os.path.relpath(p, root).replace(os.sep, "/")
                with open(p, "rb") as fp:
                    data = fp.read()
                self.walk_entries.append((self._prefixed(rel), False))
                self.files.append((self._prefixed(rel), data))
        # rel -> data 映射，避免下面三处按 rel 回头线性搜索整个 files 列表
        self.file_map = dict(self.files)
        self.dirs = [r for r, isd in self.walk_entries if isd]
        self.content = b"".join(data for _, data in self.files)

    def _prefixed(self, rel):
        return self.src_dir + "/" + rel if self.src_dir else rel

    def tar_entries(self, d_seed, xz_size):
        """计算 tar 归档中每个条目的 (header, data_size, is_dir)。
        顺序：seed, walk_entries (interleaved files+dirs), quine。
        """
        entries = []
        # seed 文件
        seed_hdr = tar_header(self.seed_name, d_seed, self.mtime)
        entries.append({"name": self.seed_name, "hdr": seed_hdr,
                        "data_size": d_seed, "is_dir": False})
        # walk 顺序的文件和目录
        for rel, isd in self.walk_entries:
            if isd:
                hdr = tar_header(rel, 0, self.mtime, mode=0o755, typeflag=b'5')
                entries.append({"name": rel, "hdr": hdr,
                                "data_size": 0, "is_dir": True})
            else:
                # file_map 与 walk_entries 在 __init__ 里同步构造，这里必有
                data = self.file_map[rel]
                entries.append({"name": rel, "hdr": tar_header(rel, len(data), self.mtime),
                                "data_size": len(data), "is_dir": False})
        # quine 文件（放在最后）
        qhdr = tar_header(self.quine_name, xz_size, self.mtime)
        entries.append({"name": self.quine_name, "hdr": qhdr,
                        "data_size": xz_size, "is_dir": False})
        return entries

    # ---------- 结构计算 ----------
    def _compute_q(self, d_seed):
        """计算 xz_bytes 在 tar 中的偏移 q。
        q = seed + files + dirs 的 tar 占用 + quine tar header。
        q 不依赖 xz_size（quine 的 data 在 q 之后）。"""
        q = 0
        # seed 文件
        q += TAR_BLOCK  # seed tar header
        q += round_up_512(d_seed)      # seed data + pad
        # walk 顺序的文件和目录
        for rel, isd in self.walk_entries:
            q += TAR_BLOCK  # tar header
            if not isd:
                q += round_up_512(len(self.file_map[rel]))
        # quine tar header
        q += TAR_BLOCK
        return q

    def _trim_layouts(self, base, o_g, f_g, q, w_plant_off, d_seed):
        """枚举 trim 配置，产出一批「相位」不同的候选布局。

        trim = [可选的裸 store chunk，载荷长度 x0] + 一组标准 gear
               （store 载荷 x = s+3，紧接着一个 lzma 复制这 x 字节）。

        裸 store 把 slip 抬高 3、把 f 抬高 3+x0；随后这组 gear 把 slip 重新归一到
        len(cb)，jump 再归一到 -3。净效果是 (o_pre, f_pre) 被整体平移
        δ ≈ s_g + 6 + len(cb) + 3 + x0，而 δ 决定了自洽方程的模 4 相位。
        旧实现 δ 恒为 0，相位一旦不对就永远无解 —— 而 d_seed += 1 根本不改变
        布局（q 只在跨 512 字节块时才变），所以那 80 次重试是纯粹的无效空转。

        注意：不能只插一个裸的 lzma 复制块。lzma 复制块的输出必须等于
        F[o-q : o-q+olen]，而这只能靠前面那个「载荷 = F[o-q : o-q+x] 且 x = s+3」
        的 store chunk 把字节搬到正确位置来实现；裸插复制块解压出来和 F 对不上。

        返回 [(chunks, o_before_jump, f_before_jump, y, jb), ...]，按文件体积升序。
        """
        s_g = f_g - (o_g - q)
        out = []
        seen = set()
        for x0 in range(0, s_g + 4):
            chunks = list(base)
            o, f = o_g, f_g
            if x0:
                # 裸 store：载荷 = F[o-q : o-q+x0]，要求 x0 <= s+3 以免读到尚未定稿的字节
                chunks.append({"kind": "store", "foff": f, "size": 3 + x0,
                               "ooff": o, "olen": x0})
                o += x0
                f += 3 + x0
            o, f = _add_gear(chunks, o, f, q)   # 一组标准 gear
            # jump 不动点：y = s + 3 + len(jb)
            s = f - (o - q)
            y = max(2, s + 17)
            jb = ju = None
            for _ in range(64):
                try:
                    jb, ju = lzma_chunk([("m", o - w_plant_off, y)], o)
                except AssertionError:
                    jb = None
                    break
                y2 = s + 3 + len(jb)
                if y2 == y:
                    break
                if y2 < 2 or y2 > 273:
                    jb = None
                    break
                y = y2
            if jb is None or ju != y:
                continue
            if y + w_plant_off - TAR_BLOCK > d_seed:
                continue
            key = (o + y, f + len(jb))
            if key in seen:
                continue
            seen.add(key)
            out.append((chunks, o, f, y, jb))

        out.sort(key=lambda t: t[2] + len(t[4]))
        return out

    def _solve_gadget(self, chunks_pre, o_j, f_j, y, jb, q, bh, FEW,
                      block_hdr, dict_size, hdrA, hdrB, d_seed, k):
        """在一个给定相位的布局上求解 gadget（T / gb / suffix / tz / k）。

        关键简化：tz 只依赖 u = gb + suffix 与 k；T = u + tz + 4 也只依赖 (u, k)；
        而 gb 又由 T 唯一确定（len(lzma_chunk(matches_for(T, T))))。
        因此只需一维枚举 u 即可遍历全部解，既不漏解也远快于原来的三重暴力。
        """
        o_pre = o_j + y
        f_pre = f_j + len(jb)
        if f_pre != (o_pre - q) - 3:      # jump 应把 slip 归一到 -3
            return None

        chunks = list(chunks_pre)
        chunks.append({"kind": "lzma", "foff": f_j, "size": len(jb),
                       "ooff": o_j, "olen": y, "bytes": jb, "jump": True})

        best = None
        for u in range(8, 512):
            # 512k 必须落在 [f_pre + 2u + 2059, f_pre + 2u + 3595)
            lo = f_pre + 2 * u + 2059
            hi = f_pre + 2 * u + 3595
            for k_cand in range(max(1, (lo - 512) // 512), (hi + 512) // 512 + 2):
                num = k_cand * 512 - (o_pre + 2 * u + 8 - q - 1024)
                if num % 3 != 0:
                    continue
                tz_cand = num // 3
                if tz_cand < 1024:
                    continue
                T = u + tz_cand + 4
                if T < 2:
                    continue
                gb, gu = lzma_chunk(matches_for(T, T), o_pre + T)
                gb_len = len(gb)
                suffix = u - gb_len
                if suffix < 1:
                    continue

                # 精确自检（不再用带 9 字节偏差的预筛公式）
                f_final = f_pre + 3 + T + gb_len + 3 + tz_cand + 1
                n = f_final - (12 + bh)
                bp = (4 - (12 + bh + n) % 4) % 4
                unpadded = bh + n + CHECK_SIZE
                idx = xz_index([(unpadded, q + k_cand * 512 + 1024)])
                ix = len(idx)
                suffix_actual = bp + CHECK_SIZE + ix + 12
                if suffix_actual != suffix:
                    continue
                xz_size = 12 + bh + n + bp + CHECK_SIZE + ix + 12
                xz_pad_actual = round_up_512(xz_size)
                if xz_pad_actual != k_cand * 512:
                    continue
                total_w_actual = q + xz_pad_actual + 1024
                o_final = o_pre + 2 * T + tz_cand
                if o_final != total_w_actual:
                    continue

                if best is not None and xz_size >= best["xz_size"]:
                    continue          # 已找到更小的解，跳过

                backward_size = ix // 4 - 1
                final = list(chunks)
                final.append({"kind": "store", "foff": f_pre, "size": 3 + T,
                              "ooff": o_pre, "olen": T, "vac": True})
                final.append({"kind": "lzma", "foff": f_pre + 3 + T,
                              "size": gb_len, "ooff": o_pre + T, "olen": gu,
                              "bytes": gb})
                final.append({"kind": "store",
                              "foff": f_pre + 3 + T + gb_len,
                              "size": 3 + tz_cand, "ooff": o_pre + 2 * T,
                              "olen": tz_cand, "trailing": True})
                final.append({"kind": "term",
                              "foff": f_pre + 3 + T + gb_len + 3 + tz_cand,
                              "size": 1, "ooff": o_final, "olen": 0,
                              "bytes": b"\x00"})
                best = {
                    "q": q, "d_seed": d_seed, "k": k,
                    "chunks": final, "block_hdr": block_hdr, "bh": bh,
                    "n": n, "xz_size": xz_size,
                    "T": T, "suffix": suffix_actual,
                    "bp": bp, "ix": ix,
                    "dict_size": dict_size,
                    "FEW": FEW, "hdrA": hdrA,
                    "hdrB": hdrB, "jump_y": y, "index": idx,
                    "stream_hdr": xz_stream_header(),
                    "footer": xz_stream_footer(backward_size),
                    "entries": self.tar_entries(d_seed, xz_size),
                }
        return best

    def _layout_iterate(self, d_seed, maxdist_hint):
        """核心布局。

        阶段 1a：C1 / repro / gears（只依赖 q，所有 trim 配置共用，只算一次）
        阶段 1b：插入 trim 微调块，平移 (o_pre, f_pre) 改变模 4 相位
        阶段 2  ：按 u = gb + suffix 一维搜索自洽解
        """
        block_hdr, dict_size = xz_block_header(maxdist_hint)
        bh = len(block_hdr)
        FEW = 12 + bh + 3
        w_hdrA_off = TAR_BLOCK + 12 + bh
        w_hdrB_off = TAR_BLOCK + 12 + bh + 3
        w_plant_off = TAR_BLOCK + 12 + bh + 6

        q = self._compute_q(d_seed)

        # ---------- 阶段 1a：C1 + repro + gears ----------
        base = []
        o, f = 0, 12 + bh
        Ls = []
        rem = q + FEW
        while rem > CHUNK:
            Ls.append(CHUNK)
            rem -= CHUNK
        Ls.append(rem)
        assert Ls[-1] >= 2, "末 C1 chunk 载荷 %d 太小" % Ls[-1]
        k = len(Ls)
        hdrA = store_hdr(CHUNK, first=False)
        hdrB = store_hdr(Ls[-1], first=False)

        S = 0
        for j, L in enumerate(Ls):
            base.append({"kind": "store", "foff": f, "size": 3 + L,
                         "ooff": o, "olen": L, "first": j == 0,
                         "c1": True, "S": S})
            o += L
            f += 3 + L
            S += L
        assert o == q + FEW

        S = 0
        for j, L in enumerate(Ls):
            if j > 0:
                srcpos = w_hdrA_off if (j < k - 1 or L == CHUNK) else w_hdrB_off
                dist_h = (q + FEW + S + 3 * (j - 1)) - srcpos
                hb, hu = lzma_chunk([("m", dist_h, 3)], o)
                base.append({"kind": "lzma", "foff": f, "size": len(hb),
                             "ooff": o, "olen": hu, "bytes": hb})
                o += hu
                f += len(hb)
            cb, cu = lzma_chunk(matches_for(q + FEW + 3 * j, L), o)
            base.append({"kind": "lzma", "foff": f, "size": len(cb),
                         "ooff": o, "olen": cu, "bytes": cb})
            o += cu
            f += len(cb)
            S += L

        # gears：把 slip 压到 <= 16
        while f - (o - q) > 16:
            o, f = _add_gear(base, o, f, q)

        o_g, f_g = o, f

        # ---------- 阶段 1b + 阶段 2 ----------
        # 相位按文件体积升序枚举；第一个成功的相位通常已经足够好，但不同相位的
        # (u, tz) 不同，xz_size 仍可能相差几百字节，所以再多看几个取最小的。
        tried = 0
        best = None
        since_hit = 0
        for chunks_pre, o_j, f_j, y, jb in self._trim_layouts(
                base, o_g, f_g, q, w_plant_off, d_seed):
            lay = self._solve_gadget(chunks_pre, o_j, f_j, y, jb, q, bh, FEW,
                                     block_hdr, dict_size, hdrA, hdrB,
                                     d_seed, k)
            tried += 1
            if lay is not None and (best is None
                                    or lay["xz_size"] < best["xz_size"]):
                best = lay
            if best is not None:
                since_hit += 1
                if since_hit >= 10:
                    break
            elif tried >= 80:        # 一直无解，放弃后续相位
                break
        if best is not None:
            return best

        raise RuntimeError("layout 不收敛: 无自洽解 (q=%d, f_pre=%d, o_pre=%d)"
                           % (q, f_g, o_g))

    # ---------- 装配 ----------
    def assemble(self, lay):
        """构建 xz 文件 F 和 seed。"""
        F = bytearray(lay["xz_size"])
        q = lay["q"]
        d_seed = lay["d_seed"]
        bh = lay["bh"]
        FEW = lay["FEW"]

        # bytearray 已全零初始化：block padding、CRC64 占位、trailing-zero 载荷
        # 都不必再显式写 0（后面几处赋值只会写到各自独立的区间，不会交叉覆盖）。

        # pass A: 写 xz stream header + block header + 所有 chunk bytes + suffix
        F[0:12] = lay["stream_hdr"]
        F[12:12+bh] = lay["block_hdr"]

        jump_c = None
        for c in lay["chunks"]:
            if c["kind"] == "store":
                F[c["foff"]:c["foff"] + 3] = store_hdr(c["olen"],
                                                       first=c.get("first", False))
            else:
                F[c["foff"]:c["foff"] + c["size"]] = c["bytes"]
                if c.get("jump"):
                    jump_c = c

        # block padding + CRC64(pass D 填) + index + footer
        bp_off = 12 + bh + lay["n"]
        ix_off = bp_off + lay["bp"] + CHECK_SIZE
        F[ix_off:ix_off + lay["ix"]] = lay["index"]
        sf_off = ix_off + lay["ix"]
        F[sf_off:sf_off + 12] = lay["footer"]

        assert jump_c is not None

        # pass B: 构建 seed
        y = lay["jump_y"]
        plant = bytes(F[jump_c["ooff"] - q: jump_c["ooff"] - q + y])
        seed = bytearray(d_seed)
        seed[0:12] = lay["stream_hdr"]
        seed[12:12+bh] = lay["block_hdr"]
        seed[12+bh:12+bh+3] = lay["hdrA"]
        seed[12+bh+3:12+bh+6] = lay["hdrB"]
        seed[12+bh+6:12+bh+6+y] = plant
        pad = (SEED_NOTE * (d_seed // len(SEED_NOTE) + 2))[:d_seed - 12 - bh - 6 - y]
        seed[12+bh+6+y:] = pad

        # 构建 tar 前缀 T（含 seed data）
        T = self._build_tar_prefix_with_seed(lay, seed)
        F_prefix = bytes(F[0:FEW])
        P = T + F_prefix
        assert len(P) == q + FEW, "P 长度 %d != q+FEW %d" % (len(P), q + FEW)

        # pass C: 填充 store payloads
        for c in lay["chunks"]:
            if c["kind"] != "store":
                continue
            ooff, olen = c["ooff"], c["olen"]
            if c.get("c1"):
                F[c["foff"] + 3:c["foff"] + 3 + olen] = P[c["S"]:c["S"] + olen]
            elif c.get("trailing"):
                # 载荷是尾部零：F 全零初始化，且唯一会写到 chunk 之外的 vac_store
                # 目标是 [f_pre+3, f_pre+3+T)（f_pre == (o_pre-q)-3 已由 _solve_gadget
                # 断言），与本区块不重叠，因此无需写入。
                continue
            elif c.get("vac"):
                _forward_vac(F, c, q)
            else:
                # 其余 store 的载荷 = 输出流自身对应位置的字节（自引用复制）
                fo = ooff - q
                F[c["foff"] + 3:c["foff"] + 3 + olen] = F[fo:fo + olen]

        # pass D: 求解 CRC64 自引用
        self._solve_crc64(F, lay, T)

        return bytes(F), bytes(seed)

    def _build_tar_prefix_with_seed(self, lay, seed):
        """构建 tar 前缀 T（含 seed data）。"""
        entries = lay["entries"]
        T = bytearray()
        for e in entries[:-1]:
            T += e["hdr"]
            if not e["is_dir"] and e["data_size"] > 0:
                data = seed if e["name"] == self.seed_name else self.file_map[e["name"]]
                T += data
                T += b'\x00' * (round_up_512(e["data_size"]) - e["data_size"])
        T += entries[-1]["hdr"]  # quine tar header
        return bytes(T)

    def _solve_crc64(self, F, lay, T_prefix):
        """求解 CRC64 自引用。

        CRC64 值 D = CRC64(W)，W = T + F + 尾部零。
        D 在 W 中出现两次：F 的 CRC64 字段 + vac_store 副本。

        解法（多项式逆元）：
          CRC64 的线性贡献可表示为 GF(2^64) 多项式乘法：
            bit_reverse(contribution(D)) = D_poly * P (mod G)
          其中 D_poly = bit_reverse(D)，P = x^(64+trailing) mod G。
          方程 D_poly * (1 ^ P1 ^ P2) = bit_reverse(CRC64(W)) (mod G)
          用扩展欧几里得求 (1 ^ P1 ^ P2) 的逆元，一步求解。
        """
        q = lay["q"]
        bh = lay["bh"]
        n = lay["n"]
        bp = lay["bp"]
        xz_size = lay["xz_size"]
        T = lay["T"]

        # CRC64 位置 1：在 F 中（block padding 后）
        crc_f_off = 12 + bh + n + bp
        crc_w1 = q + crc_f_off

        # CRC64 位置 2：vac_store 副本
        crc_f_copy = crc_f_off - T
        crc_w2 = q + crc_f_copy

        # 构造 W = T + F + 尾部零（先把两处 CRC64 置 0）
        xz_pad = round_up_512(xz_size)
        tz = xz_pad - xz_size + 1024
        W = bytearray(T_prefix) + F + b'\x00' * tz
        W[crc_w1:crc_w1 + CHECK_SIZE] = b'\x00' * CHECK_SIZE
        W[crc_w2:crc_w2 + CHECK_SIZE] = b'\x00' * CHECK_SIZE

        # CRC64(W) 的多项式表示
        crc_poly = _bit_reverse64(crc64(bytes(W)))

        # P1 = x^(64 + 位置1之后的bit数) mod G
        trailing1 = (len(W) - crc_w1 - 8) * 8
        P1 = _poly_pow(2, 64 + trailing1)
        # P2 = x^(64 + 位置2之后的bit数) mod G
        trailing2 = (len(W) - crc_w2 - 8) * 8
        P2 = _poly_pow(2, 64 + trailing2)

        # D_poly = crc_poly * inv(1 ^ P1 ^ P2)
        coeff = 1 ^ P1 ^ P2
        inv = _poly_minv(coeff)
        if inv == 0:
            raise RuntimeError("CRC64 多项式无逆元")
        D_poly = _poly_mul_mod(crc_poly, inv)
        D = _bit_reverse64(D_poly)

        # 写入 D 到 F 的 CRC64 位置
        struct.pack_into('<Q', F, crc_f_off, D)

        # 重新执行 vac_store 副本
        _forward_vac(F, next(c for c in lay["chunks"] if c.get("vac")), q)

        # 验证
        W2 = bytearray(T_prefix) + F + b'\x00' * tz
        assert crc64(bytes(W2)) == D, "CRC64 求解验证失败"

    def build(self, maxdist_hint=None):
        if maxdist_hint is None:
            maxdist_hint = 2 * (len(self.content) + 128) + (1 << 21)
        last_err = None
        # d_seed 在同一 512 字节块内不改变 q（seed 数据按 512 对齐），所以旧的
        # "d_seed += 1" 重试完全不会改变布局，只是空转 80 次。真正需要 d_seed
        # 变化的情况只有两种：seed 装不下 jump 种植串，或需要跨块改变 q 作为兜底。
        for d_seed in (64, 96, 128, 160, 192, 256, 320, 384, 448,
                       576, 704, 832, 960, 1088, 1600, 2112, 2624, 3648):
            try:
                lay = self._layout_iterate(d_seed, maxdist_hint)
                F, seed = self.assemble(lay)
                return F, lay, seed
            except (RuntimeError, AssertionError, ValueError) as e:
                last_err = e
        raise RuntimeError("build 多次失败: %s" % last_err)


# ============================================================
# Part 4: 验证 + main
# ============================================================

def verify(F, lay, files, quine_name):
    """验证：解压 xz → 解析 tar → 检查文件 + quine 自复制。"""
    # 1. 解压 xz
    try:
        dec = lzma.LZMADecompressor(format=lzma.FORMAT_XZ)
        tar_data = dec.decompress(F)
    except Exception as e:
        print("[verify] xz 解压失败: %s" % e)
        return False

    # 2. 检查 quine 自复制
    if tar_data[lay["q"]:lay["q"] + lay["xz_size"]] != F:
        print("[verify] quine 自复制不匹配")
        a = tar_data[lay["q"]:lay["q"] + lay["xz_size"]]
        for i in range(min(len(a), len(F))):
            if a[i] != F[i]:
                print("[verify]   首个不匹配在偏移 %d: tar=%02x vs xz=%02x" % (i, a[i], F[i]))
                break
        return False

    # 3. 解析 tar
    try:
        tf = tarfile.open(fileobj=io.BytesIO(tar_data))
        members = tf.getmembers()
    except Exception as e:
        print("[verify] tar 解析失败: %s" % e)
        return False

    print("[verify] tar 条目: %s" % [m.name for m in members])

    # 4. 检查源文件
    ok = True
    for rel, data in files:
        try:
            f = tf.extractfile(rel)
            if f is None:
                print("[verify] 文件 %s 是目录或不存在" % rel)
                ok = False
                continue
            extracted = f.read()
            if extracted != data:
                print("[verify] 文件 %s 内容不匹配" % rel)
                ok = False
        except KeyError:
            print("[verify] 文件 %s 不在 tar 中" % rel)
            ok = False

    # 5. 检查 quine 文件
    try:
        qf = tf.extractfile(quine_name)
        if qf:
            qdata = qf.read()
            if qdata != F:
                print("[verify] quine 文件内容不匹配 (%d vs %d)" % (len(qdata), len(F)))
                ok = False
            else:
                print("[verify] quine 文件匹配!")
    except KeyError:
        print("[verify] quine 文件不在 tar 中")
        ok = False

    print("[verify] 结果: %s" % ("通过" if ok else "失败"))
    return ok


def main():
    ap = argparse.ArgumentParser(description="多文件 tar.xz quine 生成器")
    ap.add_argument("srcdir")
    ap.add_argument("output")
    ap.add_argument("--quine-name", default=None,
                    help="归档内 quine 文件名（默认 = output 的文件名）")
    ap.add_argument("--seed-name", default=".this_is_mayx_blog")
    ap.add_argument("--src-dir", default="",
                    help="源文件在归档内的前缀目录")
    ap.add_argument("--quine-dir", default="",
                    help="quine 在归档内所在目录")
    args = ap.parse_args()

    quine_name = args.quine_name or os.path.basename(args.output)
    if args.quine_dir:
        quine_dir = args.quine_dir.strip("/")
        quine_name = quine_dir + "/" + quine_name

    bq = TarXzQuine(args.srcdir, quine_name=quine_name, seed_name=args.seed_name,
                    src_dir=args.src_dir)
    print("[blogquine-tarxz] %d 个文件, %d 个目录, 内容 %d 字节"
          % (len(bq.files), len(bq.dirs), len(bq.content)))
    F, lay, seed = bq.build()
    with open(args.output, "wb") as fp:
        fp.write(F)
    print("[blogquine-tarxz] written %s: %d 字节 (q=%d, k=%d chunks, n=%d, xz_size=%d)"
          % (args.output, len(F), lay["q"], lay["k"], lay["n"], lay["xz_size"]))
    print("[blogquine-tarxz] T=%d, suffix=%d (bp=%d, ix=%d), dict_size=%d"
          % (lay["T"], lay["suffix"], lay["bp"], lay["ix"], lay["dict_size"]))
    ok = verify(F, lay, bq.files, quine_name)
    sys.exit(0 if ok else 1)


if __name__ == "__main__":
    main()
```
</details>

  不过对我来说，相比于TXZ格式我还是更喜欢7z一点，一是因为LZMA2本来就是7-Zip的作者发明的，XZ感觉像是摘桃子的，二是TXZ这个名字听起来有点怪，感觉不像压缩包，三是XZ Utils出过后门，尽管整个Linux社区都在使用XZ，但是我还是稍微有点偏见，所以这份TXZ的生成器我就粘贴出来给需要的人吧，我博客用7z格式就好了。   

# 感想
  以前总是有人说AI没有创新能力，只是对曾经在网络上出现的东西进行重组，但人何尝不是这样呢？像这次制作的7z/TXZ Quine生成器在整个网络上没有任何公开信息，当然我也知道这并不是理论上的创新，但谁说组合创新不是创新呢？再看看最近OpenAI又解决了一大堆数学难题，完全可以相信AI是真的拥有智能，所以我相信总有一天AI将能完成人类能做的所有事情，人类将不再需要额外的思考，只需要做自己想做的事情吧。