ci: balance UI partitions and streamline verified documentation checks - #143

Summary

• Balance four isolated single-worker UI jobs by historical file-duration estimates. Discover the complete current Playwright inventory and assign every file exactly once; new files/tests remain included independently of the timing baseline.
• Cancel obsolete Test runs only within the same PR or main ref. Other PRs and release workflows stay separate.
• Allow browser-free verification only for PRs whose actual tested-tree diff contains exclusively allowlisted regular documentation files. Unknown/mixed changes, code renames, symlinks, dependencies, assets, tests and CI changes retain full UI verification. Main pushes always run full UI, including documentation successors, so cancellation cannot discard preceding runtime coverage.
• Preserve Node Tests, UI Test and Release Package, both dependency audits, immutable tool/image pins and smoke VSIX verification. The stable UI gate rejects failed, cancelled, unknown or unexpected skipped dependencies.
• Restore all nine README editions to one concise v2.5 What's New feature summary and collapse older v2.4 details. Document restrained, tab-scoped plugin design, opt-in unrelated extras and optional habit-changing experiences in bilingual project principles; no production UI or shipped defaults change.

中文审阅

• 四个隔离、单 worker 分片按历史耗时估算均衡；从当前完整清单分配文件，新增文件／用例不会被历史表遗漏。
• 只取消同一 PR 或 main 引用下过时的 Test，不干扰其他 PR／发布工作流。
• 仅已核实纯文档 PR 可省去浏览器；未知、混合、代码重命名、符号链接、依赖、资源、测试及 CI 改动完整测 UI。main 始终完整测 UI，避免新文档提交取消前一轮代码检查后免测。
• 固定必需门禁名称、双锁审计、工具／镜像固定及打包安全校验不变；UI 汇总对异常依赖状态失败关闭。
• 九语 README 的 2.5 改为简短功能概要，旧 2.4 折叠；双语项目原则明确克制插件风格、局部优化、无关新附加功能默认关闭，以及显著改变习惯时提供可选路径。不改变生产界面或已发布默认值。

Verification

• Full local Node/policy suite after the remote-acceptance fix: 2743 passed, 2 existing skips, 0 failures.
• Red/green coverage for documentation classification, actual Git ranges, rename/mode handling, unknown/cancelled/skipped gate results, workflow bypass mutations and main-push cancellation coverage.
• Actual pinned Playwright CLI rediscovery: 876 tests across 40 files, disjoint complete partition coverage. Counts: 208 / 198 / 189 / 281; no removed tests, languages, themes, retries or snapshot tolerances.
• Historical model: approximately 15.3 minutes per partition, compared with a modeled previous maximum of 21.5 minutes using the same file averages. These are estimates from public run 39041847767, not measured speedup; timestamps, fixtures and scheduling introduce uncertainty. Provenance and limitations are retained with the numeric baseline.
• Workflow YAML parsing, focused regression suites and local smoke VSIX packaging/security verification passed. README local links/images/anchors were checked.
• Independent review identified the main-successor cancellation gap; its fix was reproduced red, verified green and rechecked.

Remote acceptance follow-up

The first docs-only probe exposed GitHub's transitive skip behavior: its Node and UI gates passed, but Release Package was also skipped. The workflow now uses an explicit uncancelled/successful-dependency condition so intentional UI omission cannot suppress packaging; failed/cancelled dependencies remain blocked. The regression was reproduced red and verified green, including the condition state table and independent review.

• Corrected candidate: b21c47a4bcf49788c354a6999b2cf9959c34a7de.
• Refreshed documentation-only probe run: Node, scope, UI aggregate and Release Package all succeeded; the browser matrix was intentionally not expanded.
• Temporary verification PR #144 is isolated from main and will not be merged. Its running-source probe was automatically cancelled by the newer run in Test-144; the old UI gate failed closed and packaging did not execute successfully. The restored-docs run then passed Node, scope, UI aggregate, packaging and VSIX verification. Its cumulative diff contains only CONTRIBUTING.md, with the CI script restored to the candidate blob.
• The exact corrected candidate's full run completed successfully on attempt 1: all four UI partitions, Node Tests, UI Test and Release Package succeeded. All 876 UI tests passed exactly once with no skipped, retried, flaky or duplicate cases. Both dependency audits, native compressed-log compatibility and smoke VSIX verification also succeeded.
• Observed end-to-end time (workflow creation to completion, including queue/setup/package) was 18m55s, versus 24m23s for the preceding main run: approximately 22% shorter in this single-run comparison, not a guaranteed benchmark. UI job durations narrowed from 9m01s–23m28s to 15m54s–16m51s. Corrected docs-only probes took 2m02s / 2m04s, including Node, audits and packaging.

中文：首次纯文档实测发现打包因跳过祖先而一并跳过，已修补显式状态条件并补红绿回归。修补后两次文档探针均实际完成打包；旧完整 UI 运行被同一并发组的新提交自动取消，UI 总门禁失败关闭，恢复后的文档运行成功。独立 #143 最终完整对照四片876例及三项必需门禁全绿，无跳过／重试／重复；临时 #144 不合并。单次实测全流程24分23秒降至18分55秒（约22%），分片耗时15分54秒至16分51秒；纯文档约2分钟，保留Node、双锁审计和打包。耗时受排队及运行环境影响，不代表固定提速。

No release/tag/publication or branch-protection changes.

🤖 Generated by OpenAI Codex as an automated first pass — not a maintainer decision.