Skip to content

feat(standards): promote LWS-86 1.0, ship the data bundle, verify with an independent implementation - #354

Merged
wangzhizhou merged 2 commits into
mainfrom
dev
Nov 20, 2026
Merged

wangzhizhou merged 2 commits into
mainfrom
dev

Conversation

@wangzhizhou

Copy link
Copy Markdown
Contributor

这一版做什么

1.4.0 = LWS-86 1.0 转正(S-1~S-4)+ 可靠性体检(R-1)+ 文档对账(D-1)。范围依据 2026-11-20 全量待办对账:
P1 段清零、P2 四项触发线均未触发、残余根序与二级扩字卡证据,所以本版不含排序/体验改动。

S-1 数据包随 Release 发布

  • make wubi86-package → LWS-86-1.0.zip + .sha256(5 文件:roots / chars / shortcodes / meta / SHA256SUMS)。
  • Release 工作流新增打包步骤,并在缺失该资产时拒绝发布。
  • 可复现:generatedAt 取提交时间(SOURCE_DATE_EPOCH 可覆盖)、打包时间戳固定、条目按名排序 →
    同一提交两次生成得到同一个 SHA-256(契约测试覆盖:bundle 与 zip 各一条)。

S-3 独立实现验证(暴露了两个真问题)

tools/lws86-reference-impl.py:只读 spec-1.0.md 的规则、不 import 判定器,从「字根序列 + 末笔 + 结构」
重算全码与识别码 → 3,746 / 3,755 逐字一致、0 失败。它抓出:

  1. 多键位字根无法消歧:⺋ 凵 丆 𠂇 丷 儿 𠂉 各挂两个键,静态字根表判不出某字用哪个(755 字受影响)。
    新增 R2-note:以逐字 keys 列为准,拿不到时必须标 partial,不许按"常见键"猜。
  2. 3 个成字字根分类错了:方 yygn / 车 lgnh / 止 hhhg 按各自笔顺套 R9 应得 yygt / lgng / hhgg,
    数据却标着可推导的 radical-char。生成器改为实算分类,R9e 特例从 1 条扩到 4 条(加 母)。
    没有改任何码、没有改笔顺——数据是事实(ADR-388)。

剩 9 条未覆盖:4 个已列明特例 + 5 个缺笔顺数据的成字字根(夕 戈 巳 臼 皿),标 partial 不猜。

S-2 / S-4

  • site/standards.html:规范摘要、十条规则、数据包内容与校验、独立验证结果、许可,以及**「独立制定、非官方、无背书」声明**;
    首页导航加入口。站点契约测试从 7 条扩到 10 条(自包含、可发现、边界写清)。
  • spec-1.0.md 转 1.0 正式,CHANGELOG 记三条修正,覆盖边界写明(GB2312 一级;一级外 27 字 partial)。

D-1 文档对账

ROADMAP 关掉两处早已被 B0 运行时/结构库取代的过期行,修好被断成两截的冷首键条目,corpus 与拼音 FSA 挂到 P2 触发线。

验证

  • make verify(1,164 产物,0 网络符号)
  • make macos-offline-tests:73 文件 / 368 例(新增 7 例:可复现 bundle/zip、特例分类、独立实现 L4、站点标准页 ×3)
  • make wubi86-package:自检 L4(3,755 字 0 失败)+ 独立实现 L4(3,746 字 0 失败)

还没做

R-1 可靠性体检(30 min soak、睡眠唤醒、多屏、dict.extended 档内存)要在本版候选包上跑,
跑完把结论写进报告 331 再发版。

合并方式

merge commit。

Scope decided after a full backlog reconciliation: P1/P1-U/P1-M/P1-W
are empty, all four P2 triggers are unmet (first 83.9%, avgPos 0.15,
cold first key p95 24.4 ms, sentence hit 70.6%), and the remaining
root-order work is blocked on a glyph reference.

1.4.0 therefore does only what needs no further user evidence:
- S-1..S-4: publish the LWS-86 1.0 data bundle with the release, add a
  standards page to the site, verify the spec with an independent minimal
  parser, and promote spec-1.0 to final (ADR-382).
- R-1: 30-minute soak, sleep/wake, multi-display and extended-dictionary
  memory, none of which ran in the 1.3.2-1.3.4 candidates.
- D-1: close two stale ROADMAP lines that described work already done by
  the B0 runtime and the structure library, repair the split cold-first-key
  item, and tie the corpus and pinyin-FSA items to their P2 triggers.

Signed-off-by: wangzhizhou <824219521@qq.com>
…nt implementation found

An independent implementation (tools/lws86-reference-impl.py) recomputes
codes and recognition letters from the spec text alone, without importing
the checker. It reproduced 3,746 of 3,755 in-scope characters with zero
failures and surfaced two real problems:

1. Seven glyphs live on two keys each (⺋ 凵 丆 𠂇 丷 儿 𠂉), so the static
   root table cannot disambiguate a character's key (755 characters are
   affected). New R2-note: the per-character keys column is normative, and
   an implementation without it must mark the character partial.
2. 方 yygn, 车 lgnh and 止 hhhg cannot be derived by R9 from their own
   stroke order (which would give yygt / lgng / hhgg), yet the data tagged
   them radical-char. The generator now derives the class instead of
   guessing from letter shape, and R9e lists all four exceptions.

No code was changed and no stroke order was bent to make the rules pass;
the data is the fact, ADR-388 records the correction.

Also in this commit:
- The standard bundle ships as a release asset (LWS-86-1.0.zip + .sha256);
  the release workflow refuses to publish without it.
- The bundle is reproducible: generatedAt comes from the commit, archive
  timestamps are fixed and entries sorted, so the same commit yields the
  same SHA-256 (contract test asserts it).
- site/standards.html: spec summary, rules, bundle contents, independent
  verification result, licence, and an explicit "independent, not official,
  no endorsement" statement.
- spec-1.0.md promoted to 1.0 final with its coverage boundaries spelled
  out (GB2312 level 1 only; 27 characters outside it stay partial).

Signed-off-by: wangzhizhou <824219521@qq.com>
@wangzhizhou
wangzhizhou merged commit ff14a6b into main Nov 20, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant