| Files | Last commit | Last update |
|---|---|---|
feat(perf): 10.27.1 — compositor-audit 脚本 + 静态预算门(#123),并纠正评审里三处未经实测的前提 #123 的定位对(性能维度确实没有门),但它给出的静态证据是把注释也算进去的计数。 实测(npm run compositor:audit,注释与代码分开计):backdrop-filter 是 6 条标准声明 + 6 条 -webkit- 孪生(41 次"出现"里 26 次在注释);blur( 17 处;rAF 2 个**调用点** (另 2 次是 typeof 能力检测)。它写的"低功耗只降透明度、不减 blur 层"方向也是反的: 低功耗的规则是 html[data-dsh-dream-skin-performance] * / *::before / *::after 上的 backdrop-filter: none !important——关掉的正是全部模糊。 新增(都是仓库侧工具,不进 npm 包): - scripts/compositor-audit.cjs:逐处输出 行号/半径/选择器;半径解析字面量、var() 回退、 calc() 各项之和并标出滑杆驱动;-webkit- 孪生单列;注释不计;跑 --json 供门读。 **一处自我纠错**:第一版只读 CSS 声明,于是报告"0 处 filter: blur()",而壁纸的全屏 模糊是 JS 写在行内样式上的(wallpaperEl.style.filter,模糊填充时还有 bleed 光晕)—— 现在它作为独立类别计入,并要求两处都留在低功耗守卫后面。 - tests/compositor_budget.test.cjs(7 条):6 块面板 / 静态半径 ≤40px / 40 处渐变 / 2 个 rAF / 0 条 CSS filter: / 2 处行内模糊且受守卫;另含"审计必须真读到东西" (解析器坏掉时全 0 不许当通过)、-webkit- 孪生配对、"注释不是声明"。 验证:11 条变异腿在**真 bundle** 上逐条实跑(含"变异必须真的改了字节"这条前置, 第一版有 4 条静默空转被它抓出来重写),每条只红在它自己点名的断言上,控制腿前后全绿、 按字节复原;整套 620→627 全绿;严格 wash:check 六组照绿。 顺带修回一条随环境变红的门:本机官方桌面壳 2026-10-23 更新后重掷了三个 CSS module 的 基名(app.asar mtime 为证),A1 的通道分叉登记表按壳的新字节重推(aSus8q→DXuOLW / v1kfCW→F07MqG / RlGAzG→GRgSLq),"无害"声明在新基名上复核仍成立,npm 那半未动。 这不是本轮的改动引起的(未改动的 HEAD 上同样红,已实测)。 README 8 份同步新增「合成器预算」一节,写明能证明什么/不能证明什么、低功耗是用户唯一 的开关、以及一份有用性能报告该带哪三样;命令注明要在仓库里跑(npm 包只带运行时)。 | 21 days ago | |
fix(skin): #93 字号极值补首次机读采样——headless 直设 10/22px,三读数落档 - 新增 scripts/font-extreme-sample.cjs:独立夹具页 + 无头引擎直接设 --dsh-content-font-size(本机 rc.1 实装零改动),皮肤 token 从发货 bundle 现读;ivory/abyss/midnight × 10/22px × 三读数共 6 格。 - 对比度全格 ≥5.59(WCAG),两档极值读数逐字相同——极值未产生新的 可读性缺口。 - docs/desktop-support.md 未验证条目改为「计算样式已采样,真机活页与 像素观感未验(采样日期 2026-10-09)」;desktop.fontscale 门钉值与 日期正则同步(12/12 绿,含 93.9 日期变异)。 - 如实保留未验项:22px 观感 0 次目视、真机活页复核、弹窗滑杆两端 × 极值组合、次级档 9/20 px。 Co-Authored-By: AtomCode (glm5.3-flash) <noreply@atomgit.com> | 1 month ago | |
feat(host): adapt to DSH 0.2.2-rc.3 — re-measure the census, un-blind the readers, retire the removed fade The 28 red lights on rc.3 were mostly OUR reading layer going blind, not the host moving: rc.3 re-prints bundle literals as backtick templates, and four readers only understood double quotes. hostCss/cssLiterals now decode both and refuse interpolated templates; the composer-slot and queue-dock shape checks accept both spellings without loosening the claim. The shell channel changed its shape too: the dsh-css region comments are gone, module CSS ships as real files, and the rolled class map (<Name>_module_css_default) is now the strongest key we can derive from. The bundle also embeds the original TSX, so byte-presence counted one element twice — stamps are now counted where the renderer puts them. One surface really was removed: dsh-client-ui-workspace no longer draws _fade. Recorded as host-retired (graded both ways) with the old sheet archived from the npm tarball under two sha256 witnesses, legal only while a fresh scan of all 337 bundles proves no live .<base>_fade rule. Also: the sidebar bleed moved from _regionArea to _scrollArea with a 2px right-edge allowance, so the flush claim now compares against the allowance parsed from the host CSS each run; census re-measured to 337 packages / 1526 files / 3055 tokens with 17 new tokens dispositioned; two hardcoded readings replaced by read-from-census. | 3 days ago | |
feat(host): adapt to DSH 0.2.2-rc.3 — re-measure the census, un-blind the readers, retire the removed fade The 28 red lights on rc.3 were mostly OUR reading layer going blind, not the host moving: rc.3 re-prints bundle literals as backtick templates, and four readers only understood double quotes. hostCss/cssLiterals now decode both and refuse interpolated templates; the composer-slot and queue-dock shape checks accept both spellings without loosening the claim. The shell channel changed its shape too: the dsh-css region comments are gone, module CSS ships as real files, and the rolled class map (<Name>_module_css_default) is now the strongest key we can derive from. The bundle also embeds the original TSX, so byte-presence counted one element twice — stamps are now counted where the renderer puts them. One surface really was removed: dsh-client-ui-workspace no longer draws _fade. Recorded as host-retired (graded both ways) with the old sheet archived from the npm tarball under two sha256 witnesses, legal only while a fresh scan of all 337 bundles proves no live .<base>_fade rule. Also: the sidebar bleed moved from _regionArea to _scrollArea with a 2px right-edge allowance, so the flush claim now compares against the allowance parsed from the host CSS each run; census re-measured to 337 packages / 1526 files / 3055 tokens with 17 new tokens dispositioned; two hardcoded readings replaced by read-from-census. | 3 days ago | |
feat(host): adapt to DSH 0.2.2-rc.3 — re-measure the census, un-blind the readers, retire the removed fade The 28 red lights on rc.3 were mostly OUR reading layer going blind, not the host moving: rc.3 re-prints bundle literals as backtick templates, and four readers only understood double quotes. hostCss/cssLiterals now decode both and refuse interpolated templates; the composer-slot and queue-dock shape checks accept both spellings without loosening the claim. The shell channel changed its shape too: the dsh-css region comments are gone, module CSS ships as real files, and the rolled class map (<Name>_module_css_default) is now the strongest key we can derive from. The bundle also embeds the original TSX, so byte-presence counted one element twice — stamps are now counted where the renderer puts them. One surface really was removed: dsh-client-ui-workspace no longer draws _fade. Recorded as host-retired (graded both ways) with the old sheet archived from the npm tarball under two sha256 witnesses, legal only while a fresh scan of all 337 bundles proves no live .<base>_fade rule. Also: the sidebar bleed moved from _regionArea to _scrollArea with a 2px right-edge allowance, so the flush claim now compares against the allowance parsed from the host CSS each run; census re-measured to 337 packages / 1526 files / 3055 tokens with 17 new tokens dispositioned; two hardcoded readings replaced by read-from-census. | 3 days ago | |
fix(gitcode): 恢复仓库到可交付状态(社区 PR / issue / 三方评审 / 迁移死链) 此前两次 cherry-pick 已把社区 PR #1 / #2 的运行时修复落到本分支;本次提交 收口评审整改、宿主漂移、迁移死链与文档读数。 评审整改: - 恢复 tests/client.plugin_page.test.cjs 被 PR #1 删掉的正向契约断言 - 撤销 PR #2 在 page-sandbox 里把 181 条 smoke 用例默认切到「有图标插槽」的后门, 新增一条带反向变异的集成用例,钉住有 seat 的宿主在 apply() 当场停用 legacy hook - settings-render 夹具记录 inject() 的插槽名;wash.cascade 的变异锚点改为现场解析宿主 inset 宿主 0.2.2-rc.1: - census 重测到 329 包 / 1226 文件,新增 token 记为 not-skinned - 宿主撤掉 --dsw-specific-sidebar-nav-item-hover 后,新增「已退休宿主插槽」类别, 旧读数钉到归档 census 的固定提交,并配四条反向变异禁止该类别变成免死金牌 链接门禁: - 纳入 github: shorthand 禁止形状;档案豁免改为「保留陈述、不暴露可点击死链」, 且 fence 感知(Store 证据的第三方原文块逐字保留,由 sha256 见证守着) - 发布手册 files 示例与 package.json 对齐,并加反向变异防止再次漂移 迁移收尾: - 主 README + 七语译文的安装入口、成长图说明、图片托管口径统一到 GitCode - CHANGELOG 八条旧 release 自动链接降级成无 scheme 行内代码;顶部补「未发布」段 - SECURITY / CONTRIBUTING / issue 模板 / 四个停用 workflow / collect-stats / font script 修正 - 主预览图压到 592,196 B,避开 GitCode 大图占位 SVG - 文档读数同步到 782 条(780 通过 / 2 具名环境跳过),质量审计 229/229 本机验证:npm test 782/780 pass/0 fail/2 skip;skin:check、craft:audit 12/12、 skin:audit 217/217、host:census:check、previews:check、typecheck、 verify-finite-caller-set、npm pack --dry-run 全过。GitHub Actions 已停用, GitCode 流水线尚未接上,以上均为本机读数。 Closes #1, #2, #3 Refs: PR #1 (556ed28425fc1a1e8f6cb61b2b3562947a678ab7), PR #2 (d774c5735337705451bce196795af0d69b96b3a4) | 3 days ago |
| Files | Last commit | Last update |
|---|---|---|
| 21 days ago | ||
| 1 month ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago |