// HDF5 只读子集：MATLAB 7.3 `.mat` 解析用（见 MatH5.hpp 的说明）。
#include "MatH5.hpp"

#include <algorithm>
#include <cstdio>
#include <cstring>

namespace jdi
{

	namespace
	{
		constexpr std::uint16_t kMessageDataspace = 0x0001;
		constexpr std::uint16_t kMessageLink = 0x0006;
		constexpr std::uint16_t kMessageDatatype = 0x0003;
		constexpr std::uint16_t kMessageLayout = 0x0008;
		constexpr std::uint16_t kMessageFilter = 0x000B;
		constexpr std::uint16_t kMessageAttribute = 0x000C;
		constexpr std::uint16_t kMessageContinuation = 0x0010;
		constexpr std::uint16_t kMessageSymbolTable = 0x0011;

		std::uint64_t ReadLE(const std::uint8_t* data, std::size_t size)
		{
			std::uint64_t value = 0;
			for (std::size_t index = 0; index < size; ++index)
			{
				value |= static_cast<std::uint64_t>(data[index]) << (8 * index);
			}
			return value;
		}

		// ==================================================================
		// 内置 inflate（RFC 1950 / 1951）
		//
		// ⚠️ **不能用 zlib（`-lz`）**：链接 libz.a 会让 Release 产物多出约 46KB，
		// 而这里只需要「解压 HDF5 分块」这一个功能，自带的紧凑实现只要几 KB。
		// ⚠️ 更不能为了取文件大小包含 `<filesystem>`：它会连带拉进 locale / codecvt /
		// iostream / cp-demangle 等一整套 libstdc++ 代码，实测让产物从 787KB 涨到 1.7MB。
		// ==================================================================

		/** @brief 位流读取器（LSB 优先，与 DEFLATE 一致）。 */
		struct BitReader
		{
			const std::uint8_t* data = nullptr;
			std::size_t size = 0;
			std::size_t position = 0;
			std::uint32_t buffer = 0;
			int bits = 0;
			bool error = false;

			bool Fill(int count)
			{
				while (bits < count)
				{
					if (position >= size)
					{
						error = true;
						return false;
					}
					buffer |= static_cast<std::uint32_t>(data[position++]) << bits;
					bits += 8;
				}
				return true;
			}

			std::uint32_t Read(int count)
			{
				if (count <= 0)
				{
					return 0;
				}
				if (!Fill(count))
				{
					return 0;
				}
				const std::uint32_t value = buffer & ((1u << count) - 1u);
				buffer >>= count;
				bits -= count;
				return value;
			}

			void Align() { buffer = 0; bits = 0; }
		};

		/** @brief 规范 Huffman 表（只存计数与按长度排好序的符号）。 */
		struct HuffmanTable
		{
			std::uint16_t count[16] = {};
			std::uint16_t symbol[320] = {};
		};

		bool BuildHuffman(HuffmanTable& table, const std::uint8_t* lengths, int total)
		{
			for (int index = 0; index < 16; ++index)
			{
				table.count[index] = 0;
			}
			for (int index = 0; index < total; ++index)
			{
				++table.count[lengths[index]];
			}
			if (table.count[0] == total)
			{
				return true;	// 全部长度为 0：该表没有任何码（动态块里可能是距离表）
			}
			int left = 1;
			for (int length = 1; length < 16; ++length)
			{
				left <<= 1;
				left -= table.count[length];
				if (left < 0)
				{
					return false;	// 超订（over-subscribed）
				}
			}
			std::uint16_t offsets[16] = {};
			offsets[1] = 0;
			for (int length = 1; length < 15; ++length)
			{
				offsets[length + 1] = static_cast<std::uint16_t>(offsets[length] + table.count[length]);
			}
			for (int index = 0; index < total; ++index)
			{
				if (lengths[index] != 0)
				{
					table.symbol[offsets[lengths[index]]++] = static_cast<std::uint16_t>(index);
				}
			}
			return true;
		}

		int DecodeSymbol(BitReader& reader, const HuffmanTable& table)
		{
			int code = 0;
			int first = 0;
			int index = 0;
			for (int length = 1; length < 16; ++length)
			{
				code |= static_cast<int>(reader.Read(1));
				if (reader.error)
				{
					return -1;
				}
				const int count = table.count[length];
				if (code - first < count)
				{
					return table.symbol[index + (code - first)];
				}
				index += count;
				first = (first + count) << 1;
				code <<= 1;
			}
			return -1;
		}

		const std::uint16_t kLengthBase[29] = {3,  4,  5,  6,  7,  8,  9,  10, 11,  13,
		                                       15, 17, 19, 23, 27, 31, 35, 43, 51,  59,
		                                       67, 83, 99, 115, 131, 163, 195, 227, 258};
		const std::uint8_t kLengthExtra[29] = {0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2,
		                                       2, 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0};
		const std::uint16_t kDistanceBase[30] = {1,    2,    3,    4,    5,    7,     9,     13,
		                                         17,   25,   33,   49,   65,   97,    129,   193,
		                                         257,  385,  513,  769,  1025, 1537,  2049,  3073,
		                                         4097, 6145, 8193, 12289, 16385, 24577};
		const std::uint8_t kDistanceExtra[30] = {0, 0, 0, 0, 1, 1, 2, 2,  3,  3,  4,  4,  5,  5,  6,
		                                         6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13};

		/** @brief 解压 zlib 流（含 2 字节头与 4 字节 Adler32 尾）。 */
		bool Inflate(const std::uint8_t* source, std::size_t sourceSize,
		             std::vector<std::uint8_t>& out)
		{
			if (sourceSize < 6)
			{
				return false;
			}
			// zlib 头：CMF/FLG。只接受 deflate（CM = 8），且不接受预设字典。
			const std::uint8_t cmf = source[0];
			const std::uint8_t flg = source[1];
			if ((cmf & 0x0F) != 8 || ((static_cast<unsigned>(cmf) * 256u + flg) % 31u) != 0 ||
			    (flg & 0x20) != 0)
			{
				return false;
			}
			BitReader reader;
			reader.data = source + 2;
			reader.size = sourceSize - 6;	// 末尾 4 字节是 Adler32
			out.clear();
			HuffmanTable fixedLiterals;
			HuffmanTable fixedDistances;
			bool fixedReady = false;

			for (int block = 0; block < 100000; ++block)
			{
				const std::uint32_t final = reader.Read(1);
				const std::uint32_t type = reader.Read(2);
				if (reader.error)
				{
					return false;
				}
				if (type == 0)
				{
					// 存储块（未压缩）：对齐后 4 字节长度 + 反码 + 原始字节。
					reader.Align();
					if (reader.position + 4 > reader.size)
					{
						return false;
					}
					const std::uint16_t length = static_cast<std::uint16_t>(
					    reader.data[reader.position] | (reader.data[reader.position + 1] << 8));
					reader.position += 4;
					if (reader.position + length > reader.size)
					{
						return false;
					}
					out.insert(out.end(), reader.data + reader.position,
					           reader.data + reader.position + length);
					reader.position += length;
				}
				else if (type == 1 || type == 2)
				{
					HuffmanTable literals;
					HuffmanTable distances;
					if (type == 1)
					{
						if (!fixedReady)
						{
							std::uint8_t lengths[288];
							for (int index = 0; index < 144; ++index)
							{
								lengths[index] = 8;
							}
							for (int index = 144; index < 256; ++index)
							{
								lengths[index] = 9;
							}
							for (int index = 256; index < 280; ++index)
							{
								lengths[index] = 7;
							}
							for (int index = 280; index < 288; ++index)
							{
								lengths[index] = 8;
							}
							if (!BuildHuffman(fixedLiterals, lengths, 288))
							{
								return false;
							}
							std::uint8_t distanceLengths[30];
							for (int index = 0; index < 30; ++index)
							{
								distanceLengths[index] = 5;
							}
							if (!BuildHuffman(fixedDistances, distanceLengths, 30))
							{
								return false;
							}
							fixedReady = true;
						}
						literals = fixedLiterals;
						distances = fixedDistances;
					}
					else
					{
						const int literalCount = static_cast<int>(reader.Read(5)) + 257;
						const int distanceCount = static_cast<int>(reader.Read(5)) + 1;
						const int codeCount = static_cast<int>(reader.Read(4)) + 4;
						if (reader.error || literalCount > 286 || distanceCount > 30)
						{
							return false;
						}
						static const std::uint8_t kOrder[19] = {16, 17, 18, 0, 8,  7, 9,  6, 10, 5,
						                                        11, 4,  12, 3, 13, 2, 14, 1, 15};
						std::uint8_t codeLengths[19] = {};
						for (int index = 0; index < codeCount; ++index)
						{
							codeLengths[kOrder[index]] = static_cast<std::uint8_t>(reader.Read(3));
						}
						HuffmanTable codeTable;
						if (reader.error || !BuildHuffman(codeTable, codeLengths, 19))
						{
							return false;
						}
						std::uint8_t lengths[288 + 30] = {};
						int produced = 0;
						while (produced < literalCount + distanceCount)
						{
							const int symbol = DecodeSymbol(reader, codeTable);
							if (symbol < 0)
							{
								return false;
							}
							if (symbol < 16)
							{
								lengths[produced++] = static_cast<std::uint8_t>(symbol);
								continue;
							}
							int repeat = 0;
							std::uint8_t value = 0;
							if (symbol == 16)
							{
								if (produced == 0)
								{
									return false;
								}
								value = lengths[produced - 1];
								repeat = 3 + static_cast<int>(reader.Read(2));
							}
							else if (symbol == 17)
							{
								repeat = 3 + static_cast<int>(reader.Read(3));
							}
							else
							{
								repeat = 11 + static_cast<int>(reader.Read(7));
							}
							if (reader.error || produced + repeat > literalCount + distanceCount)
							{
								return false;
							}
							for (int index = 0; index < repeat; ++index)
							{
								lengths[produced++] = value;
							}
						}
						if (!BuildHuffman(literals, lengths, literalCount) ||
						    !BuildHuffman(distances, lengths + literalCount, distanceCount))
						{
							return false;
						}
					}

					for (int symbolIndex = 0; symbolIndex < 100000000; ++symbolIndex)
					{
						const int symbol = DecodeSymbol(reader, literals);
						if (symbol < 0)
						{
							return false;
						}
						if (symbol < 256)
						{
							out.push_back(static_cast<std::uint8_t>(symbol));
							continue;
						}
						if (symbol == 256)
						{
							break;	// 块结束
						}
						const int lengthCode = symbol - 257;
						if (lengthCode >= 29)
						{
							return false;
						}
						const std::size_t length =
						    kLengthBase[lengthCode] + reader.Read(kLengthExtra[lengthCode]);
						const int distanceCode = DecodeSymbol(reader, distances);
						if (distanceCode < 0 || distanceCode >= 30 || reader.error)
						{
							return false;
						}
						const std::size_t distance =
						    kDistanceBase[distanceCode] + reader.Read(kDistanceExtra[distanceCode]);
						if (reader.error || distance == 0 || distance > out.size())
						{
							return false;
						}
						const std::size_t start = out.size() - distance;
						for (std::size_t index = 0; index < length; ++index)
						{
							out.push_back(out[start + index]);
						}
					}
				}
				else
				{
					return false;	// 类型 3 非法
				}
				if (final != 0)
				{
					return true;
				}
			}
			return false;
		}
	} // namespace

	std::uint8_t MatHdf5::U8(std::uint64_t offset) const
	{
		return (offset < m_data.size()) ? m_data[static_cast<std::size_t>(offset)] : 0;
	}

	std::uint16_t MatHdf5::U16(std::uint64_t offset) const
	{
		return static_cast<std::uint16_t>(U8(offset)) | (static_cast<std::uint16_t>(U8(offset + 1)) << 8);
	}

	std::uint32_t MatHdf5::U32(std::uint64_t offset) const
	{
		return static_cast<std::uint32_t>(U16(offset)) | (static_cast<std::uint32_t>(U16(offset + 2)) << 16);
	}

	std::uint64_t MatHdf5::U64(std::uint64_t offset) const
	{
		return static_cast<std::uint64_t>(U32(offset)) | (static_cast<std::uint64_t>(U32(offset + 4)) << 32);
	}

	std::uint64_t MatHdf5::Addr(std::uint64_t offset) const
	{
		return (m_offsetSize == 8) ? U64(offset) : U32(offset);
	}

	std::uint64_t MatHdf5::Resolve(std::uint64_t address) const
	{
		if (address == 0 || address == 0xFFFFFFFFFFFFFFFFull)
		{
			return 0;
		}
		return address + m_base;
	}

	bool MatHdf5::Open(const std::wstring& path, std::wstring& error, const OpenProgress& progress)
	{
		error.clear();
		Close();
		// ⚠️ 取文件大小一律用 C 运行库（`_fseeki64` / `_ftelli64`），**不要用 `<filesystem>`**：
		// 后者会把 libstdc++ 的 locale / codecvt / iostream / cp-demangle 整套拉进产物，
		// 实测让 Release 二进制从 787KB 涨到 1.7MB（用户 2026-10-19 报告）。
		FILE* file = nullptr;
		if (_wfopen_s(&file, path.c_str(), L"rb") != 0 || file == nullptr)
		{
			error = L"无法打开文件。";
			return false;
		}
		if (_fseeki64(file, 0, SEEK_END) != 0)
		{
			std::fclose(file);
			error = L"无法读取文件大小。";
			return false;
		}
		const long long size = _ftelli64(file);
		if (size <= 0 || _fseeki64(file, 0, SEEK_SET) != 0)
		{
			std::fclose(file);
			error = L"无法读取文件大小。";
			return false;
		}
		m_data.resize(static_cast<std::size_t>(size));

		// 按块读入并回报进度（加载遮罩层的"已加载大小 / 总大小 / 百分比"由它驱动）。
		constexpr std::size_t kChunk = 4u << 20;
		std::size_t offset = 0;
		while (offset < m_data.size())
		{
			const std::size_t want = std::min(kChunk, m_data.size() - offset);
			const std::size_t got = std::fread(m_data.data() + offset, 1, want, file);
			offset += got;
			if (got != want)
			{
				break;
			}
			if (progress && !progress(offset, m_data.size()))
			{
				std::fclose(file);
				Close();
				error.clear();	// 用户取消：不显示错误
				return false;
			}
		}
		std::fclose(file);
		if (offset != m_data.size())
		{
			error = L"读取文件失败。";
			return false;
		}

		// 超级块：签名可能出现在 0、512、1024…（MATLAB 用 512 字节 user block）。
		std::size_t signature = m_data.size();
		for (std::size_t index = 0; index + 8 <= m_data.size() && index <= 4096; index += 512)
		{
			if (m_data[index] == 0x89 && m_data[index + 1] == 'H' && m_data[index + 2] == 'D' &&
			    m_data[index + 3] == 'F')
			{
				signature = index;
				break;
			}
		}
		if (signature == m_data.size())
		{
			error = L"不是 HDF5（MATLAB 7.3）格式。";
			return false;
		}
		const std::uint8_t version = U8(signature + 8);
		if (version == 0 || version == 1)
		{
			m_offsetSize = U8(signature + 13);
			m_lengthSize = U8(signature + 14);
			m_base = U64(signature + 24);
			const std::uint64_t rootEntry = signature + 24 + 32;
			m_root = Resolve(Addr(rootEntry + 8));
		}
		else
		{
			m_offsetSize = U8(signature + 9);
			m_lengthSize = U8(signature + 10);
			m_base = U64(signature + 12);
			m_root = Resolve(U64(signature + 36));
		}
		if (m_root == 0 || m_root >= m_data.size())
		{
			error = L"HDF5 超级块损坏。";
			return false;
		}
		return true;
	}

	void MatHdf5::Close()
	{
		m_data.clear();
		m_base = 0;
		m_root = 0;
		m_heaps.clear();
	}

	void MatHdf5::ReadMessages(std::uint64_t address, std::vector<Message>& messages) const
	{
		messages.clear();
		if (address + 8 > m_data.size())
		{
			return;
		}
		if (U8(address) == 1)
		{
			const std::uint16_t count = U16(address + 2);
			const std::uint32_t headerSize = U32(address + 8);
			// 先收集消息块：主块（长度 = 对象头里的 headerSize）+ 若干续块（续块里没有对象头前缀）。
			// ⚠️ **主块长度必须用 headerSize 限定**：消息数超过主块容量时，其余消息在续块里，
			// 若在主块里继续按"消息数"往下读，就会读到块外的其它数据（字段全部解析失败）。
			std::vector<std::pair<std::uint64_t, std::uint64_t>> blocks;
			blocks.emplace_back(address + 16, headerSize);
			std::size_t read = 0;
			for (std::size_t blockIndex = 0; blockIndex < blocks.size(); ++blockIndex)
			{
				std::uint64_t offset = blocks[blockIndex].first;
				const std::uint64_t end =
				    (blocks[blockIndex].second > 0) ? (offset + blocks[blockIndex].second) : 0;
				while (read < count && offset + 8 <= m_data.size() && (end == 0 || offset + 8 <= end))
				{
					Message message;
					message.type = U16(offset);
					const std::uint16_t messageSize = U16(offset + 2);
					offset += 8;
					if (offset + messageSize > m_data.size())
					{
						break;
					}
					message.data.assign(m_data.begin() + static_cast<std::ptrdiff_t>(offset),
					                    m_data.begin() + static_cast<std::ptrdiff_t>(offset + messageSize));
					offset += (messageSize + 7u) & ~static_cast<std::uint16_t>(7u);
					++read;
					if (message.type == kMessageContinuation && message.data.size() >= 16)
					{
						const std::uint64_t blockAddress = Resolve(ReadLE(message.data.data(), 8));
						const std::uint64_t blockSize = ReadLE(message.data.data() + 8, 8);
						if (blockAddress != 0)
						{
							blocks.emplace_back(blockAddress, blockSize);
						}
					}
					messages.push_back(std::move(message));
				}
			}
		}
	}

	void MatHdf5::WalkBtree(std::uint64_t node, std::uint64_t heapData,
	                        std::vector<std::pair<std::string, std::uint64_t>>& out) const
	{
		int guard = 0;
		while (node != 0 && node + 24 < m_data.size() && guard++ < 100000)
		{
			if (std::memcmp(&m_data[static_cast<std::size_t>(node)], "TREE", 4) != 0)
			{
				return;
			}
			const std::uint8_t level = U8(node + 5);
			const std::uint16_t entries = U16(node + 6);
			std::uint64_t offset = node + 8 + 8 + 8;
			std::vector<std::uint64_t> children;
			for (std::uint16_t index = 0; index < static_cast<std::uint16_t>(entries + 1u); ++index)
			{
				offset += 8;	// key（分组 B-tree 的 key 是堆偏移）
				if (index < entries)
				{
					children.push_back(Resolve(U64(offset)));
					offset += 8;
				}
			}
			for (const std::uint64_t child : children)
			{
				if (level > 0)
				{
					WalkBtree(child, heapData, out);
				}
				else
				{
					WalkSymbolNode(child, heapData, out);
				}
			}
			node = Resolve(U64(node + 8));	// 右兄弟
		}
	}

	void MatHdf5::WalkSymbolNode(std::uint64_t node, std::uint64_t heapData,
	                             std::vector<std::pair<std::string, std::uint64_t>>& out) const
	{
		if (node == 0 || node + 8 > m_data.size() ||
		    std::memcmp(&m_data[static_cast<std::size_t>(node)], "SNOD", 4) != 0 || heapData == 0)
		{
			return;
		}
		const std::uint16_t count = U16(node + 6);
		std::uint64_t offset = node + 8;
		for (std::uint16_t index = 0; index < count; ++index)
		{
			if (offset + 40 > m_data.size())
			{
				return;
			}
			const std::uint64_t nameOffset = U64(offset);
			const std::uint64_t objectAddress = Resolve(U64(offset + 8));
			offset += 40;
			if (objectAddress == 0 || heapData + nameOffset >= m_data.size())
			{
				continue;
			}
			const char* name = reinterpret_cast<const char*>(m_data.data() + heapData + nameOffset);
			out.emplace_back(std::string(name), objectAddress);
		}
	}

	std::vector<std::pair<std::string, std::uint64_t>> MatHdf5::Members(std::uint64_t address) const
	{
		std::vector<std::pair<std::string, std::uint64_t>> out;
		std::vector<Message> messages;
		ReadMessages(address, messages);
		for (const Message& message : messages)
		{
			if (message.type == kMessageSymbolTable && message.data.size() >= 16)
			{
				const std::uint64_t btree = Resolve(ReadLE(message.data.data(), 8));
				const std::uint64_t heap = Resolve(ReadLE(message.data.data() + 8, 8));
				std::uint64_t heapData = 0;
				if (heap + 32 <= m_data.size() &&
				    std::memcmp(&m_data[static_cast<std::size_t>(heap)], "HEAP", 4) == 0)
				{
					heapData = Resolve(U64(heap + 24));
					bool known = false;
					for (const auto& entry : m_heaps)
					{
						if (entry.first == heap)
						{
							known = true;
							break;
						}
					}
					if (!known)
					{
						m_heaps.emplace_back(heap, heapData);
					}
				}
				WalkBtree(btree, heapData, out);
			}
			else if (message.type == kMessageLink && message.data.size() >= 8)
			{
				std::size_t offset = 2;
				const std::uint8_t flags = message.data[1];
				if (flags & 0x08)
				{
					offset += 8;
				}
				if (flags & 0x04)
				{
					offset += 8;
				}
				if (flags & 0x10)
				{
					++offset;
				}
				if (offset >= message.data.size())
				{
					continue;
				}
				const std::uint8_t lengthSize = message.data[offset++];
				std::uint64_t nameLength = 0;
				for (std::uint8_t byte = 0; byte < lengthSize && offset < message.data.size(); ++byte)
				{
					nameLength |= static_cast<std::uint64_t>(message.data[offset + byte]) << (8 * byte);
				}
				offset += lengthSize;
				if (offset + nameLength + 8 > message.data.size())
				{
					continue;
				}
				std::string name(reinterpret_cast<const char*>(message.data.data() + offset),
				                 static_cast<std::size_t>(nameLength));
				offset += static_cast<std::size_t>(nameLength);
				const std::uint64_t target = Resolve(ReadLE(message.data.data() + offset, 8));
				out.emplace_back(std::move(name), target);
			}
		}
		return out;
	}

	std::uint64_t MatHdf5::Member(std::uint64_t address, const std::string& name) const
	{
		for (const auto& entry : Members(address))
		{
			if (entry.first == name)
			{
				return entry.second;
			}
		}
		return 0;
	}

	bool MatHdf5::IsDataset(std::uint64_t address) const
	{
		std::vector<Message> messages;
		ReadMessages(address, messages);
		for (const Message& message : messages)
		{
			if (message.type == kMessageDataspace)
			{
				return true;
			}
		}
		return false;
	}

	bool MatHdf5::ParseType(const std::uint8_t* data, std::size_t size, std::size_t offset, Type& out,
	                        std::size_t& consumed) const
	{
		if (offset + 8 > size)
		{
			return false;
		}
		out.version = data[offset] >> 4;
		out.cls = static_cast<TypeClass>(data[offset] & 0x0F);
		const std::uint8_t bits0 = data[offset + 1];
		out.size = ReadLE(data + offset + 4, 4);
		std::size_t cursor = offset + 8;
		if (out.cls == TypeClass::Integer)
		{
			out.bigEndian = (bits0 & 0x01) != 0;
			out.isSigned = (bits0 & 0x08) != 0;
			if (out.version >= 3)
			{
				cursor += 2;	// bit offset
				cursor += 2;	// precision
			}
			else
			{
				cursor += 4;
			}
		}
		else if (out.cls == TypeClass::Float)
		{
			out.bigEndian = (bits0 & 0x01) != 0;
			cursor += 12;
		}
		else if (out.cls == TypeClass::String)
		{
			// 无需额外属性
		}
		else if (out.cls == TypeClass::Reference)
		{
			// 无需额外属性
		}
		else if (out.cls == TypeClass::VariableLength)
		{
			out.isVariableString = (bits0 & 0x0F) == 1;
			std::size_t childConsumed = 0;
			Type child;
			if (ParseType(data, size, cursor, child, childConsumed))
			{
				cursor += childConsumed;
			}
		}
		else if (out.cls == TypeClass::Compound)
		{
			const unsigned members = static_cast<unsigned>(bits0) | (static_cast<unsigned>(data[offset + 2]) << 8);
			for (unsigned index = 0; index < members; ++index)
			{
				if (cursor >= size)
				{
					break;
				}
				const std::size_t nameStart = cursor;
				while (cursor < size && data[cursor] != 0)
				{
					++cursor;
				}
				std::string name(reinterpret_cast<const char*>(data + nameStart), cursor - nameStart);
				++cursor;	// 结束符
				cursor = (cursor + 7) & ~static_cast<std::size_t>(7);
				if (out.version == 3)
				{
					// v3：先维度（1 字节 rank + 4 字节保留），再字节偏移（可变）
				}
				cursor += 4;	// 字节偏移（v1/v2）
				std::size_t childConsumed = 0;
				Type child;
				if (ParseType(data, size, cursor, child, childConsumed))
				{
					cursor += childConsumed;
					out.members.emplace_back(std::move(name), std::move(child));
				}
				else
				{
					break;
				}
			}
		}
		else if (out.cls == TypeClass::Array)
		{
			if (cursor + 4 <= size)
			{
				const std::uint8_t rank = data[cursor];
				cursor += 4;
				cursor += static_cast<std::size_t>(rank) * 4;	// 维度
				cursor += static_cast<std::size_t>(rank) * 4;	// 排列
				std::size_t childConsumed = 0;
				Type child;
				if (ParseType(data, size, cursor, child, childConsumed))
				{
					cursor += childConsumed;
				}
			}
		}
		consumed = cursor - offset;
		return true;
	}

	bool MatHdf5::DecodeChunked(const Dataset& dataset, std::size_t elementSize,
	                            std::vector<std::uint8_t>& out) const
	{
		if (dataset.chunkAddress == 0 || dataset.chunkDims.empty() || elementSize == 0 ||
		    dataset.dims.empty())
		{
			return false;
		}
		// ⚠️ MATLAB 写出的分块维度可能与数据空间维度**维数不一致**（实测：1 维数据集的
		// 分块维度是 `[N, 1]`）。把维度数组在**右侧补 1**（行优先的最快轴）对齐即可，
		// 二者的元素总数与遍历顺序都不变。
		std::vector<std::size_t> dims = dataset.dims;
		std::vector<std::size_t> chunkDims = dataset.chunkDims;
		while (dims.size() < chunkDims.size())
		{
			dims.push_back(1);
		}
		while (chunkDims.size() < dims.size())
		{
			chunkDims.push_back(1);
		}
		const std::size_t rank = dims.size();
		const std::uint64_t total = [&]() -> std::uint64_t
		{
			std::uint64_t value = 1;
			for (const std::size_t dim : dims)
			{
				value *= dim;
			}
			return value;
		}();
		out.assign(static_cast<std::size_t>(total) * elementSize, 0);
		std::vector<std::uint8_t> chunk;
		std::vector<std::uint8_t> inflated;
		// 遍历分块 B-tree：key = 分块大小 + 过滤器掩码 + 各维偏移。
		std::vector<std::uint64_t> stack{dataset.chunkAddress};
		int guard = 0;
		while (!stack.empty() && guard++ < 1000000)
		{
			const std::uint64_t node = stack.back();
			stack.pop_back();
			if (node == 0 || node + 24 > m_data.size() ||
			    std::memcmp(&m_data[static_cast<std::size_t>(node)], "TREE", 4) != 0)
			{
				continue;
			}
			const std::uint8_t level = U8(node + 5);
			const std::uint16_t entries = U16(node + 6);
			const std::size_t keySize = 8 + 8 * rank;
			std::uint64_t offset = node + 8 + 8 + 8;
			for (std::uint16_t index = 0; index < static_cast<std::uint16_t>(entries + 1u); ++index)
			{
				const std::uint32_t chunkSize = (offset + 4 <= m_data.size()) ? U32(offset) : 0;
				std::vector<std::uint64_t> chunkOffset(rank, 0);
				for (std::size_t axis = 0; axis < rank; ++axis)
				{
					chunkOffset[axis] = U64(offset + 8 + axis * 8);
				}
				offset += keySize;
				if (index < entries)
				{
					const std::uint64_t child = Resolve(U64(offset));
					offset += 8;
					if (level > 0)
					{
						stack.push_back(child);
					}
					else if (child != 0 && chunkSize > 0)
					{
						chunk.assign(m_data.begin() + static_cast<std::ptrdiff_t>(child),
						             m_data.begin() + static_cast<std::ptrdiff_t>(child + chunkSize));
						const std::uint8_t* source = chunk.data();
						std::size_t sourceSize = chunk.size();
						if (dataset.deflate)
						{
							// 分块未压缩时大小 = 分块元素数 × 元素大小：此时不需要解压。
							std::size_t expected = elementSize;
							for (const std::size_t dim : dataset.chunkDims)
							{
								expected *= dim;
							}
							if (chunk.size() != expected)
							{
								// 内置 inflate（不依赖 zlib，见文件头的说明）。
								if (!Inflate(chunk.data(), chunk.size(), inflated))
								{
									continue;
								}
								if (inflated.size() < expected)
								{
									inflated.resize(expected, 0);
								}
								source = inflated.data();
								sourceSize = expected;
							}
						}
						// 把该块拷进输出缓冲区（按行优先展开）。
						const std::size_t rowBytes = chunkDims[rank - 1] * elementSize;
						std::size_t rows = 1;
						for (std::size_t axis = 0; axis + 1 < rank; ++axis)
						{
							rows *= chunkDims[axis];
						}
						for (std::size_t row = 0; row < rows; ++row)
						{
							std::vector<std::size_t> coordinate(rank - 1, 0);
							std::size_t remainder = row;
							for (std::size_t axis = rank - 1; axis-- > 0;)
							{
								coordinate[axis] = remainder % chunkDims[axis];
								remainder /= chunkDims[axis];
							}
							std::uint64_t targetRow = 0;
							bool inside = true;
							std::uint64_t stride = 1;
							for (std::size_t axis = rank - 1; axis-- > 0;)
							{
								const std::uint64_t index = chunkOffset[axis] + coordinate[axis];
								if (index >= dims[axis])
								{
									inside = false;
									break;
								}
								targetRow += index * stride;
								stride *= dims[axis];
							}
							if (!inside)
							{
								continue;
							}
							const std::size_t sourceOffset = row * rowBytes;
							const std::size_t targetOffset =
							    static_cast<std::size_t>(targetRow) * dims[rank - 1] * elementSize;
							if (sourceOffset + rowBytes > sourceSize ||
							    targetOffset + rowBytes > out.size())
							{
								continue;
							}
							std::memcpy(out.data() + targetOffset, source + sourceOffset, rowBytes);
						}
					}
				}
			}
		}
		return true;
	}

	bool MatHdf5::ReadDataset(std::uint64_t address, Dataset& out) const
	{
		out = Dataset();
		std::vector<Message> messages;
		ReadMessages(address, messages);
		bool haveSpace = false;
		bool haveType = false;
		std::size_t compactSize = 0;
		std::uint64_t dataAddress = 0;
		std::uint64_t dataSize = 0;
		for (const Message& message : messages)
		{
			if (message.type == kMessageDataspace && message.data.size() >= 8)
			{
				const std::uint8_t version = message.data[0];
				const std::uint8_t rank = message.data[1];
				std::size_t offset = (version == 1) ? 8 : 4;
				out.dims.assign(rank, 0);
				for (std::uint8_t axis = 0; axis < rank && offset + 8 <= message.data.size(); ++axis)
				{
					out.dims[axis] = static_cast<std::size_t>(ReadLE(&message.data[offset], 8));
					offset += 8;
				}
				haveSpace = true;
			}
			else if (message.type == kMessageDatatype)
			{
				std::size_t consumed = 0;
				haveType = ParseType(message.data.data(), message.data.size(), 0, out.type, consumed);
			}
			else if (message.type == kMessageLayout && message.data.size() >= 8)
			{
				const std::uint8_t version = message.data[0];
				const std::uint8_t klass = message.data[1];
				out.layout = static_cast<LayoutClass>(klass);
				if (klass == 0)
				{
					// 紧凑：数据就放在消息里
					if (version == 3)
					{
						compactSize = static_cast<std::size_t>(message.data[2]) |
						              (static_cast<std::size_t>(message.data[3]) << 8);
					}
					else
					{
						compactSize = static_cast<std::size_t>(ReadLE(&message.data[2], 4));
					}
				}
				else if (klass == 1)
				{
					dataAddress = Resolve(ReadLE(&message.data[2], 8));
					dataSize = ReadLE(&message.data[10], 8);
				}
				else if (klass == 2)
				{
					if (version == 3)
					{
						const std::uint8_t rank = message.data[2];
						out.chunkAddress = Resolve(ReadLE(&message.data[3], 8));
						out.chunkDims.assign(rank, 0);
						for (std::uint8_t axis = 0; axis < rank; ++axis)
						{
							out.chunkDims[axis] = ReadLE(&message.data[11 + axis * 4], 4);
						}
					}
					else
					{
						const std::uint8_t rank = message.data[2];
						out.chunkAddress = Resolve(ReadLE(&message.data[3], 8));
						out.chunkDims.assign(rank, 0);
						for (std::uint8_t axis = 0; axis < rank; ++axis)
						{
							out.chunkDims[axis] = ReadLE(&message.data[11 + axis * 4], 4);
						}
					}
				}
			}
			else if (message.type == kMessageFilter && message.data.size() >= 2)
			{
				const int count = message.data[1];
				std::size_t offset = 8;
				for (int index = 0; index < count && offset + 8 <= message.data.size(); ++index)
				{
					const std::uint16_t id = static_cast<std::uint16_t>(message.data[offset]) |
					                         (static_cast<std::uint16_t>(message.data[offset + 1]) << 8);
					const std::uint16_t nameLength = static_cast<std::uint16_t>(message.data[offset + 2]) |
					                                 (static_cast<std::uint16_t>(message.data[offset + 3]) << 8);
					const std::uint16_t values =
					    static_cast<std::uint16_t>(message.data[offset + 6]) |
					    (static_cast<std::uint16_t>(message.data[offset + 7]) << 8);
					offset += 8;
					offset += (nameLength + 7) & ~static_cast<std::size_t>(7);
					offset += static_cast<std::size_t>(values) * 4;
					if (id == 1)
					{
						out.deflate = true;
					}
				}
			}
		}
		if (!haveSpace || !haveType || out.type.size == 0)
		{
			return false;
		}
		const std::size_t elementSize = out.type.size;
		if (out.layout == LayoutClass::Compact)
		{
			// 紧凑数据在消息里，重新扫描一次取出字节。
			for (const Message& message : messages)
			{
				if (message.type == kMessageLayout)
				{
					const std::size_t offset = (message.data[0] == 3) ? 4 : 8;
					if (offset + compactSize <= message.data.size())
					{
						out.data.assign(message.data.begin() + static_cast<std::ptrdiff_t>(offset),
						                message.data.begin() +
						                    static_cast<std::ptrdiff_t>(offset + compactSize));
					}
				}
			}
		}
		else if (out.layout == LayoutClass::Contiguous)
		{
			if (dataAddress != 0 && dataAddress + dataSize <= m_data.size())
			{
				out.data.assign(m_data.begin() + static_cast<std::ptrdiff_t>(dataAddress),
				                m_data.begin() + static_cast<std::ptrdiff_t>(dataAddress + dataSize));
			}
		}
		else if (out.layout == LayoutClass::Chunked)
		{
			if (!DecodeChunked(out, elementSize, out.data))
			{
				return false;
			}
		}
		out.ok = true;
		return true;
	}

	std::string MatHdf5::AttributeString(std::uint64_t address, const char* name) const
	{
		std::vector<Message> messages;
		ReadMessages(address, messages);
		for (const Message& message : messages)
		{
			if (message.type != kMessageAttribute || message.data.size() < 8)
			{
				continue;
			}
			const std::uint8_t version = message.data[0];
			const std::size_t headerSize = (version == 1) ? 8 : (version == 2 ? 8 : 9);
			const std::uint16_t nameSize = static_cast<std::uint16_t>(message.data[2]) |
			                               (static_cast<std::uint16_t>(message.data[3]) << 8);
			const std::uint16_t typeSize = static_cast<std::uint16_t>(message.data[4]) |
			                               (static_cast<std::uint16_t>(message.data[5]) << 8);
			const std::uint16_t spaceSize = static_cast<std::uint16_t>(message.data[6]) |
			                                (static_cast<std::uint16_t>(message.data[7]) << 8);
			std::size_t offset = headerSize;
			if (offset + nameSize > message.data.size())
			{
				continue;
			}
			std::string attributeName(reinterpret_cast<const char*>(message.data.data() + offset),
			                          nameSize);
			while (!attributeName.empty() && attributeName.back() == '\0')
			{
				attributeName.pop_back();
			}
			offset += (nameSize + 7) & ~static_cast<std::size_t>(7);
			const std::size_t typeOffset = offset;
			offset += (typeSize + 7) & ~static_cast<std::size_t>(7);
			const std::size_t spaceOffset = offset;
			offset += (spaceSize + 7) & ~static_cast<std::size_t>(7);
			if (attributeName != name)
			{
				continue;
			}
			if (offset > message.data.size())
			{
				return std::string();
			}
			Type type;
			std::size_t consumed = 0;
			if (!ParseType(message.data.data(), message.data.size(), typeOffset, type, consumed))
			{
				return std::string();
			}
			std::size_t count = 1;
			if (spaceOffset + 4 <= message.data.size())
			{
				const std::uint8_t spaceVersion = message.data[spaceOffset];
				const std::uint8_t rank = message.data[spaceOffset + 1];
				std::size_t dimOffset = spaceOffset + ((spaceVersion == 1) ? 8 : 4);
				count = 1;
				for (std::uint8_t axis = 0; axis < rank && dimOffset + 8 <= message.data.size(); ++axis)
				{
					count *= static_cast<std::size_t>(ReadLE(&message.data[dimOffset], 8));
					dimOffset += 8;
				}
			}
			std::string text;
			if (type.cls == TypeClass::String || type.cls == TypeClass::Integer)
			{
				text.assign(reinterpret_cast<const char*>(message.data.data() + offset),
				            std::min(message.data.size() - offset, std::max<std::size_t>(count, 1)));
				while (!text.empty() && text.back() == '\0')
				{
					text.pop_back();
				}
			}
			return text;
		}
		return std::string();
	}

	bool MatHdf5::ReadText(std::uint64_t address, std::wstring& out) const
	{
		out.clear();
		Dataset dataset;
		if (!ReadDataset(address, dataset))
		{
			return false;
		}
		if (dataset.type.cls == TypeClass::String)
		{
			out.assign(dataset.data.begin(), dataset.data.end());
			while (!out.empty() && out.back() == L'\0')
			{
				out.pop_back();
			}
			return true;
		}
		if (dataset.type.size == 2)
		{
			for (std::size_t index = 0; index + 1 < dataset.data.size(); index += 2)
			{
				const wchar_t character = static_cast<wchar_t>(
				    static_cast<std::uint16_t>(dataset.data[index]) |
				    (static_cast<std::uint16_t>(dataset.data[index + 1]) << 8));
				if (character != 0)
				{
					out.push_back(character);
				}
			}
			return true;
		}
		if (dataset.type.size == 1)
		{
			for (const std::uint8_t byte : dataset.data)
			{
				if (byte != 0)
				{
					out.push_back(static_cast<wchar_t>(byte));
				}
			}
			return true;
		}
		return false;
	}

	bool MatHdf5::ReadNumbers(std::uint64_t address, std::vector<double>& out) const
	{
		out.clear();
		Dataset dataset;
		if (!ReadDataset(address, dataset) || !dataset.ok)
		{
			return false;
		}
		const std::size_t elementSize = dataset.type.size;
		if (elementSize == 0 || dataset.data.size() % elementSize != 0)
		{
			return false;
		}
		const std::size_t count = dataset.data.size() / elementSize;
		out.resize(count);
		const bool integer = (dataset.type.cls == TypeClass::Integer);
		const bool floating = (dataset.type.cls == TypeClass::Float);
		if (!integer && !floating)
		{
			return false;
		}
		for (std::size_t index = 0; index < count; ++index)
		{
			const std::uint8_t* source = dataset.data.data() + index * elementSize;
			if (floating && elementSize == 8)
			{
				double value = 0.0;
				std::memcpy(&value, source, sizeof(value));
				out[index] = value;
			}
			else if (floating && elementSize == 4)
			{
				float value = 0.0f;
				std::memcpy(&value, source, sizeof(value));
				out[index] = static_cast<double>(value);
			}
			else if (integer && dataset.type.isSigned)
			{
				std::int64_t value = 0;
				for (std::size_t byte = 0; byte < elementSize; ++byte)
				{
					value |= static_cast<std::int64_t>(source[byte]) << (8 * byte);
				}
				// 符号扩展
				const int bits = static_cast<int>(elementSize) * 8;
				if (bits < 64 && (value & (1ll << (bits - 1))) != 0)
				{
					value |= ~((1ll << bits) - 1);
				}
				out[index] = static_cast<double>(value);
			}
			else if (integer)
			{
				std::uint64_t value = 0;
				for (std::size_t byte = 0; byte < elementSize; ++byte)
				{
					value |= static_cast<std::uint64_t>(source[byte]) << (8 * byte);
				}
				out[index] = static_cast<double>(value);
			}
		}
		return true;
	}

	bool MatHdf5::ReadReferences(std::uint64_t address, std::vector<std::uint64_t>& out) const
	{
		out.clear();
		Dataset dataset;		if (!ReadDataset(address, dataset) || !dataset.ok ||
		    dataset.type.cls != TypeClass::Reference)
		{
			return false;
		}
		const std::size_t count = dataset.data.size() / 8;
		out.reserve(count);
		for (std::size_t index = 0; index < count; ++index)
		{
			const std::uint64_t raw = ReadLE(dataset.data.data() + index * 8, 8);
			out.push_back(Resolve(raw));
		}
		return true;
	}

} // namespace jdi