[codeasier]

Dependency and scope

Refs #277. This final candidate is tests-only against actual main e786bfc0395a035670cfa12d568022c758ec848a: exactly the original 13 test/checker/CI/documentation/development-dependency paths. Semantic foundation #283 and the separately authorized Windows checkpoint repair #285 are merged. There is no functional dependency on #286.

Earlier head 14404c7ab2364c8798a22c44c1510c7bbb587035 temporarily included unmerged #286 input validation for pre-validation. Normal two-parent commit a35dfb47edee3fb29cb607c6c8e3007cce02ccc6 removes exactly that nine-file input delta (seven paths restored to actual main, two #286-only new test files removed) without rewriting history. No production file is deleted; merged provenance history is unchanged. All 13 #277-owned blobs/strong assertions are preserved, including focus test blob 35a1bc4a61e9b4ec57590394a2bd88cca98a7733. The independent full-tree isolation review and publication checks verified final tree b2891922f9b67cafb1f8432cec879625488e32b8.

No runtime/product code, parser-v3/frame-tree decision, credential/network/runner-permission change, fixture weakening or release action is included in the final delta.

#277 implementation

• Preserve overall branch-enabled coverage floor 90 and existing independent import-backend/allocator/mutator floors 90/88/84.
• Add independent floors: import_metadata_contract 100; import_metadata, split_service, focus_service, skill_service and snapshot_cache 90 each; representation 95.
• Fail closed on missing/corrupt coverage data, line-only data, missing/unmeasured source, no executable statements, or any unrounded percentage below its own floor. No averaging away weak modules.
• Add targeted metadata validation, cache rejection, representation boundaries, staging/replay/publication, focus compensation, short/zero writes, skill backup restoration and cleanup-fault tests. Test-owned resources are explicitly closed.
• Semantic mutations protect exactly one complete checker after full Python 3.13 measurement; missing/conditional/overridden/ignored gates and narrowed source/omit/exclusion settings are rejected. Positive mutations retain legal splits/merges.
• Retain both accepted Hypothesis and explicit coverage[toml]>=7.10.3 development dependencies.

Evidence and final gates

Independent source, wiring and isolation reviews found no blockers. Authored commit identities, ordered parents, all tree entries and the exact 13-path diff against actual main were verified before normal fast-forward publication. No local project unit tests/builds were duplicated.

The intermediate combined head passed all 12 applicable checks in Test 41320127308 / Platform 41320127240, earned 94.56% overall and all ten independent floors, and the unchanged original Windows test_zero_checkpoint_write_fails_before_any_focus_mutation passed in job 124888366129. That is intermediate evidence only, not acceptance of this final isolated tree. The new exact head must pass the complete matrix, required browser, installed-wheel acceptance and every coverage floor; #285 post-merge checks must also succeed before merge.

The final latest-main/head/review/check state will be rechecked before merging. #277 will be closed only after this PR and its post-merge CI succeed. Nightly/manual dataset acceptance is not claimed by an event-skipped PR job; the separate input-boundary #286 retains its stricter final manual acceptance plan.

Snapshot provenance

Snapshot provenance decision: no-update
Snapshot provenance no-update reason: The final diff against actual main contains only tests, CI/checker documentation and a development dependency; no snapshot runtime, attribution/source mapping or provenance history changes. The unmerged #286 record was removed solely with that out-of-scope pre-validation delta, leaving the actual-main provenance blob unchanged.

[codeasier]

Confirmed blocker at head de04a9e7666fa59facb273c1b4fe1e22d479859a: Windows Python 3.13 shard 1 fails only tests/core/test_critical_coverage_focus.py::test_zero_checkpoint_write_fails_before_any_focus_mutation (827 passed). The original focus bytes and descriptor closure are preserved, but an unexpected .focus.pt-snap-*.recovery file remains. FocusService._checkpoint_file, lines 106–114, attempts verified unlink before closing the open descriptor; Windows rejects that unlink. This is a runtime cleanup defect, not a missing browser/runner dependency. The strong assertion remains unchanged.

The test-system work stays separate. A newly authorized minimal runtime follow-up will close the descriptor before failure cleanup, retain safe-path/device/inode verification before unlink, refuse to delete a substituted path, and preserve the original error. This PR remains blocked until that separate repair and complete combined CI pass. PR283 has now merged; no parser-v3/product decision is included.

Minimal reproduction on a provisioned Windows checkout: python -m pytest tests/core/test_critical_coverage_focus.py::test_zero_checkpoint_write_fails_before_any_focus_mutation -vv --tb=short. The test uses only its own temporary files.

[codeasier]

Pre-validating the planned combined tree at 14404c7ab2364c8798a22c44c1510c7bbb587035 via a normal two-parent fast-forward update. It integrates the separately reviewed #285 checkpoint cleanup and #286 input-boundary repair stack without folding those runtime changes into this PR's own scope. Merge ordering still waits for the prerequisite runtime PRs and final current-main checks.

All original #284 files/strong safety assertions are preserved byte-for-byte, including test_zero_checkpoint_write_fails_before_any_focus_mutation (file blob 35a1bc4a61e9b4ec57590394a2bd88cca98a7733). The only mechanical overlap is pyproject.toml: both hypothesis>=6.100.0,<7 and coverage[toml]>=7.10.3 are retained. The union and every untouched-tree entry were verified; independent integration review found no blocker. All original coverage floors remain, and exact combined-head CI will retest the previously failing Windows cleanup assertion. #277 is not closed until the real merged-main/post-CI path succeeds.

[codeasier]

The original Windows safety regression now has direct passing evidence on combined head 14404c7ab2364c8798a22c44c1510c7bbb587035: Windows shard 1, job 124888366129 logged test_zero_checkpoint_write_fails_before_any_focus_mutation PASSED and finished successfully (884 passed). This is the unchanged original assertion, not merely the new #285 tests. Other required combined-head jobs and the separate runtime-PR merge/post-CI sequence are still pending; #277 remains open.

[codeasier]

Final isolated head a35dfb47edee3fb29cb607c6c8e3007cce02ccc6 passed all 12 applicable checks in Test 41326152889 and Platform 41326152736. This final tests-only tree earned 94.51% overall and passed all ten independent floors; the unchanged original Windows zero-checkpoint assertion passed again in job 124906398495.

After latest main/head/reviews/checks and the exact 13-path no-runtime delta were reverified, normal squash merge is 81af6698675efcc0bd63c8b9be0a331b14241f66, tree b2891922f9b67cafb1f8432cec879625488e32b8 matching the validated head. No #286 input-validation code was included. Post-merge CI is pending; #277 remains open until it succeeds.

[codeasier]

最终合后验收已完成。main 81af6698675efcc0bd63c8b9be0a331b14241f66 的原 push Test run 41329139788, attempt 2 与 Platform run 41329139799, attempt 2 均成功，12 项适用检查全部通过。attempt 1 曾被已合 PR 正文修正触发的并发组取消；取消不算通过，本次恢复后已核 event=push、branch=main、精确 head。

Python 3.13 合后证据：3478 passed，总体覆盖 94.51%，原总体90及三个90/88/84门槛保持，新增七项独立门槛全部通过（10项总计）；真实构建 wheel、来源审计和离线安装 CLI/skills 双格式验收均成功。Windows 原始零写入安全断言 test_zero_checkpoint_write_fails_before_any_focus_mutation 保持不变并通过；最小运行时修复独立在 #285，覆盖测试/CI 变更在 #284。

Python 3.11 最后于 17:54 UTC 成功（3462 passed）；此前较慢不据此宣称设备故障。nightly/manual job 按普通 push 事件策略 skipped，不冒充已执行；本项按已确认的 focus/覆盖门禁范围验收；#286 输入解析变更另需该补充验收。相关 issue 将在本回执后显式完成。#280 仍开放，#286 的精确最终 head 仍需独立 manual dataset/200-seed 验收及后续合后检查。