Jia Sui
3c2f3ae374
此前纠错只有交换（sahng -> shang）与 QWERTY 邻键（shabg -> shang）两类，
生成器在离线阶段就把「一个错串对应多个音节」的变体整个丢掉，运行时也只
取代价最低的一条切分。三类输入因此没有出路：漏字的（shng）、多敲一键的
（shangg、sjhang）、漏声母连基础切分都切不出来的（uanli 想打 quan'li）。
本次按专利 CN 101133411 B 的思路重排整条链路：错串可能对应的读音全部
保留到运行时，先由 k-best 切分带出平行读音，再让词典词频决定谁排前面。
- 纠错类型从两类扩到四类：新增漏字（shng -> shang）与多字（shangg 双打、
sjhang 邻键 -> shang）。两个新位跟随既有 autocorrect_transposition /
autocorrect_neighbor 开关联动（任一打开即生效），定义保持独立，留待
将来的独立配置键。查询与 preedit 重建共用同一映射，两侧纠正空间一致。
- 生成器从两表扩到四表并保留歧义变体（ahan -> {shan, zhan}）；同一条
(错串, 正确) 对被两条规则产出时在两张表各留一份，保证每个类型开关
单独有效。正确音节收敛进 kCorrectSyllables 单表、行内改存 16-bit
下标，一万七千多行不再各自重复音节字符串，--check 拦截过期表头。
- 切分层新增 autocorrect_cut_kbest：按位置自左向右做 top-k 假设传播，
排序键是（纠错边数，边权之和，生成序）——最少纠错优先的契约不变，
权重只在同边数内打破平局，引擎绝不用一条额外纠错换更低权重。边权取
专利 [0052] 允许的估常量（交换 10 / 漏字 11 / 多字 12 / 邻键 13），
标定路径是挖掘用户日志。漏字/多字边不等长，切分段改按 raw 长度推进
位置，preedit 分隔跟着原始字母的区间走。
- 词典层固定 k=9，为离线评测定值：漏声母键 uan 有 10-15 个等权漏字
目标，k=3 按表序截断够不到 quan'li 的正确读音；k=9 提升 R@1/R@3
且 p95 不变。主切分仍是查询键，其余读音交给
merge_alternative_segmentations 与主读法在词频下竞争；备选读音在
缓存命中路径同样发布，corrected_from 标记对全部纠错读音生效；纠错
读音排在音形别名读音之前。结尾单插入（shangg = shang + g）天然是
简拼尾形状，仍被简拼形守卫拦下——多字纠错的可达面在中间/开头插入
与多音节输入。
- 纠错门放宽：!segments.empty() 从门里去掉。BFS 工作在原始字母上，
基础切分为空的丢声母输入恰是最需要纠错的一批，jian -> ian 现在能
走到纠错候选并按词频合并到首位。简拼形守卫保留：合法音节加至多一个
尾字母（zheg）是用户意图，任何纠错表都不得改写。
- preedit 补两道防线：zheg 在显式三位掩码下仍走简拼守卫，不被漏字表
重排为 zheng；只有 BFS 切分的读音与 scheme 别名层完全一致时才按
BFS 位置重画分隔符，否则 sahng 会被切成 sa'hng 而候选实际来自
别名读音 shang。
新增独立评测工具 eval_quanpin_autocorrect：从真实 msime.db 采样高频
词条，按交换/邻键/漏字/多字模型注错，测 R@1/R@3、读音还原率与温态
p50/p95，固定种子可复现。它是评测工具不是测试，永远不进 CI，作为
纠错阈值与门逻辑改动的零回归标尺；注错规则与生成器逐键一致，改任一
侧须同一提交同步另一侧。
表格测试改为按 (错串, 正确) 对断言唯一并加覆盖门（合计 >= 4700 条），
新增歧义保留、漏字/多字变体、k-best 排序与权重定序、preedit 守卫等
用例；engine/tests 的 README 与 AGENTS.md 同步写清 imetest 与评测
工具两个产物的边界。
Input engine (engine/)
The C++ conversion engine behind Metasequoia IME for Windows: input sessions, candidate lookup, helpcodes, the shared contracts and the voice module.
This directory used to be a submodule of metasequoiaime/MSIME-Engine. It is now a component of this repository, a sibling of server/, ui/, windows/ and installer/. Its upstream commit, and exactly what the Windows specialization removed, are recorded in UPSTREAM.md. Treat it as first-party code: reviewed, formatted and tested like anything else here, no longer merged from upstream.
| Directory | Purpose |
|---|---|
core/, quanpin/, shuangpin/, english/, japanese/ |
Input sessions and candidate lookup |
contracts/ |
The authoritative IPC wire format, dictionary format, punctuation policy and WebView message schema |
helpcode/ |
Helpcode (形码) tables and generators |
local_modes/, providers/, user_dictionary/ |
Local modes, online candidate sources, the user dictionary |
voice/ |
Recording, WAV encoding, recognition and text cleanup; builds independently |
googlepinyinime-rev/, utfcpp/, voice/third_party/ |
Vendored third-party code, kept in its upstream formatting and licence |
Integration is through <metasequoia/session.h>. See runtime architecture.
Build
The product build does not build this directory on its own: server/CMakeLists.txt pulls it in with add_subdirectory, so CI compiles and exercises the engine through the server and the TSF frontend. After changing something here, run the server's build and tests.
Building it standalone still works and is much faster while working on the engine. Dependencies are CMake 3.25+, a C++17 compiler, Boost, fmt, spdlog and SQLite3.
vcpkg install --triplet x64-windows-static
cmake -S . -B build -DCMAKE_TOOLCHAIN_FILE="$env:VCPKG_INSTALLATION_ROOT/scripts/buildsystems/vcpkg.cmake" -DVCPKG_TARGET_TRIPLET=x64-windows-static
cmake --build build --config Release --parallel
ctest --test-dir build -C Release --output-on-failure --timeout 20
ctest runs every engine target registered in the root CMakeLists.txt: the segmentation, IPC and punctuation contracts, the platform API, data paths, runtime isolation, dictionary state and the per-mode session tests. tests/ additionally holds a separate Windows-only project that is not part of that build — add new tests to the registered targets so they actually run.
Formatting is driven from the repository root by scripts/format.sh, using the .clang-format in this directory (voice/ has its own with a wider column limit). The vendored third-party trees and the generated headers are excluded.
Contract rules
The IPC wire format, opcodes, voice framing and WebView message definitions live in contracts/ and are the single source of truth. Consumers include the same headers; do not restate them on either side. Dictionary naming is defined by contracts/dictionary/format.json, shared with the repository root's scripts/dictionary_product.py.
Data and licensing
The dictionaries are neither stored nor built here. They arrive as a release pinned by tag and digest in product-lock.json and are downloaded at build time; the pipeline that produced them left with the specialization. Per-source licensing is tracked in NOTICE.md — read it before redistributing the built databases.
Recognition goes through a cloud provider; local Whisper inference was removed. The only committed model is voice/assets/models/silero_vad.onnx, loaded solely by the standalone Windows host that is off by default.
Licence
GPL-3.0. Vendored googlepinyinime-rev, utfcpp and voice/third_party/miniaudio keep their upstream licences.