[gcl-coder]

Goal And Delivered Outcome

• Outcome basis: two remaining git ls-files -z readers rebuilt each NUL record through a text-mode decode, so the path they produced named no file and the tracked file silently left a shipped check.
• Goal/source and gap: the ratchet and the public-boundary scan both need the identity Git recorded, not a display rendering of it. text=True applies universal-newline translation (a CR inside a legal name became LF) and errors="replace" collapses an undecodable record to U+FFFD; the subsequent is_file()/suffix filter then dropped the record without any failure.
• Observable before → after: a tracked carriage\rreturn.py was absent from tracked_python_paths() and a tracked venv/carriage\rreturn.py was absent from iter_scan_files(venv); both now appear. The validation rows below are the two new tests, which fail at eba8a7f95 and pass at this head.
• Issue/task and intended base: no issue; this is a self-contained repair found while auditing the remaining ls-files -z consumers after fix(canary): preserve exact changed Git paths #8585 and fix(integration-branch): preserve worktree paths with line breaks #8603. Intended base eba8a7f95f54fe09cb3dcc28bb1f55443898bbed.

Author Declaration

• Written by: model_agent — cursor-model (cursor).

Implemented against

• Specification and revision: no written specification; the request in this PR is the basis. The documented behavior it preserves is the existing one-record-per-path rule already stated in loopx/canary/maintainability_ratchet.py and the existing scan enumeration contract in tests/test_contract_public_boundary_prefilter.py.
• Criteria:

Criterion (spec clause)	Disposition	Symbol / path	Test or command
One NUL-framed record is one pathname, including a CR inside the name	implemented	loopx/canary/maintainability_ratchet.py::tracked_python_paths	tests/canary/test_git_path_framing.py::test_tracked_python_paths_keeps_a_carriage_return_name
A pruned scan root enumerates its tracked files from Git, so none may be dropped	implemented	loopx/contract.py::_tracked_scan_files	tests/test_contract_public_boundary_prefilter.py::test_pruned_scan_root_keeps_a_carriage_return_tracked_name
The bounded Git probe keeps its timeout, session isolation and helper retirement	implemented	loopx/contract.py::_git_query_bytes	tests/test_contract_git_probe_budget.py, tests/test_contract_git_probe_cancel.py
Scan rules, suffix filters, budgets and selection policy are unchanged	implemented	same modules	existing tests/test_contract_*, tests/canary/test_maintainability_ratchet.py

• Self-check before submission: read both readers and the probe budget/cancel tests; ran the two new tests against the base commit to confirm they fail there for the reported reason; ran the surrounding suites, ruff, and canary premerge --from-git-diff. Did not add a fixture for a filesystem that stores an undecodable byte in a name, because macOS APFS refuses to create one; that branch is os.fsdecode's documented round trip and is unexercised here.

Scope And Continuation

• Completed scope and remaining work: complete within this scope. The other ls-files -z consumers already read bytes (loopx/canary/runner.py, loopx/capabilities/change_quality/scope.py, packages/loopx-core/prepare.py, loopx/semantics/development_probe.py); the two changed here were the ones that dropped a record silently instead of failing.
• Slice boundary / successor: N/A — no successor task is required for this repair.

Validation

• Tested revision: a208bc6ac48e46de20bed8d08b2cd69ec1a83c91
• Run state: finished

Check	Command	Result
Focused suites	python3 -m pytest tests/canary/test_git_path_framing.py tests/canary/test_maintainability_ratchet.py tests/test_contract_public_boundary_prefilter.py tests/test_contract_git_probe_budget.py tests/test_contract_git_probe_cancel.py tests/test_contract_scan_missing_roots.py tests/test_contract_scan_unreadable_directories.py tests/test_contract_scan_unreadable_files.py tests/test_contract_credential_pattern_precision.py -q	166 passed
New tests at base	same two test ids at eba8a7f95	both fail: the ratchet set omits the CR module, the scan list is empty
Lint	ruff check on the four changed files	pass
Premerge	canary premerge --from-git-diff --git-diff-base eba8a7f95	diff hygiene, py_compile, maintainability-ratchet smoke and all 8 risk-profile smokes pass; 5 catalog canaries fail identically at the base commit without this diff
Unrelated local failures	tests/canary/test_rfc_bilingual.py, tests/canary/test_telemetry_isolation.py	12 failures at both revisions; local environment, not this diff

The five catalog canaries that fail here (first-connect-contract-smoke, quota-plan-smoke, quota-work-lane-policy-smoke, monitor-scheduler-contract-smoke, semantic-vocabulary-drift-smoke) fail with the same commands on a clean eba8a7f95 checkout in this environment; the hosted lanes are the authority for them.

Risk / 风险

这两个读取器此前把“显示编码”当成了路径身份：文本模式会做通用换行转换（合法名字里的 CR 变成 LF），errors="replace" 又把无法解码的记录折成 U+FFFD，重建出来的路径不指向任何文件，后续 is_file() 直接把该记录丢掉，既不报错也不计数。改动只在两处 Git 读取边界保留字节身份并用 os.fsdecode 还原（POSIX 与 PEP 529 的 Windows 均为原始字节的无损反解），扫描规则、后缀过滤、超时预算与选择策略一律未动。未验证：在文件名里存放无法解码字节的文件系统（APFS 不允许创建），该分支只有 os.fsdecode 的文档往返保证。

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

评审提交 a208bc6。GitHub 目标与 merge base 为 eba8a7f，本轮读取的最新主干 7bb1b01 未改动这组变更路径。

[P2] Linux 文件名进入发现后会破坏 JSON 输出。 在 Linux Git 索引中加入一个名称含不可解码字节的文件，让该文件生成一条检查发现，再请求 JSON 输出。 两个命令均已返回失败退出码，但 stdout 混入原始字节，严格 UTF-8 解码在 JSON 解析前失败，机器调用者无法读取违规位置。 问题位置

动机 / Motivation

用 JSON 消费预合并检查结果并按位置修复发现的维护者和自动化。 被扫描器跳过的目录里仍可能有 Git 跟踪文件。旧版会漏掉某些合法文件名，操作员看到空扫描而错过实际发现。 Linux 上，两条读取路径现在能打开 Git 实际记录的文件。普通名对照保持一致，带回车的诊断可按 UTF-8 读取。 验证范围是本地 Linux Git、文件系统与两个实际命令；Windows 文件系统及已安装分发未单独验证。

改动思路 / Approach

Git 负责提供路径身份，现有扫描器和 ratchet 负责判定，现有 JSON 输出负责把结果交给调用方。 将无损文件路径和显示字段的职责留在各自已有边界，复用现有 JSON 出口。两种 Git probe 可继续共享相同的超时与清理策略。 两个入口按 NUL 切分，再按系统文件名规则还原路径。CR 不再改名，POSIX 的不可解码字节也能用于实际文件读取。 扫描枚举

具体改动 / Changes

关键代码讲解

ratchet 将还原后的路径用于模块发现，合约扫描将其用于诊断定位。两者随后把代理码点交给 ensure_ascii=False。已是 UTF-8 的标准输出保留原错误策略，在复现环境把它写回原始字节。 ratchet 文件身份 合约扫描文件身份 ratchet JSON 出口

按 docs/reference/public-boundary-read-cost.md 在 eba8a7f 的验收依据，Git-owned coverage 已验证。Explicit check reporting 仍需让机器调用者读取发现。 JSON 输出契约 已接受的验收依据

对主干的风险 / Risks

在同一 Linux 容器中，普通违规文件在两版都返回可解析的失败报告。含 CR 文件基线扫描数为 0，候选为 1 且能解析。不可解码文件在候选进入实际扫描和 ratchet 发现，两个失败报告均无法严格解码。 文件查找继续使用无损字节身份。将路径和派生标识投影到现有报告时做可区分的转义，使 JSON 始终可按 UTF-8 解码并定位对应条目。 在真实 Linux Git fixture 上覆盖普通名、含 CR 和含不可解码字节的名称。让违规命中进入两个实际 JSON 入口，分别断言退出码、严格解码、发现内容和位置。

语义与 CI 对齐

166 项定向测试通过，包含 Git 预算与真实 fsmonitor 子进程清理。Standard premerge 六项直接检查、18 项选中检查均通过。 Linux 实际 JSON 负例仍失败。 托管 CI 未查询。

我的整体评价 / Verdict

请求修改。保留字节路径修复，补齐两个现有结构化出口的报告边界。 此改动的用户入口为命令行结构化报告。已直接检查其 stdout 字节和严格解析。 验证范围是本地 Linux Git、文件系统与两个实际命令；Windows 文件系统及已安装分发未单独验证。

English verdict: REQUEST_CHANGES. On a208bc6, a tracked Linux pathname with an undecodable byte reaches both scanners but makes their failing JSON reports invalid UTF-8.

[loopx-agent]

[P2] Linux 文件名进入发现后会破坏 JSON 输出。 在 Linux Git 索引中加入一个名称含不可解码字节的文件，让该文件生成一条检查发现，再请求 JSON 输出。 两个命令均已返回失败退出码，但 stdout 混入原始字节，严格 UTF-8 解码在 JSON 解析前失败，机器调用者无法读取违规位置。 文件查找继续使用无损字节身份。将路径和派生标识投影到现有报告时做可区分的转义，使 JSON 始终可按 UTF-8 解码并定位对应条目。 在真实 Linux Git fixture 上覆盖普通名、含 CR 和含不可解码字节的名称。让违规命中进入两个实际 JSON 入口，分别断言退出码、严格解码、发现内容和位置。