fix(cosh-ng): accept quoted dev null as sink

[jfeng18]

Why

A quoted redirect target such as 2>"/dev/null" was classified as a filesystem write (redirection-write) instead of a null sink, so the approval card showed a misleading reason even though the unquoted form was already classified correctly (#1667).

What changed

• Fully quoted SAFE_OUTPUT_SINKS targets are now recognized as null redirections.
• Every other quoted form stays fail-closed (unchanged).
• apply_null_redirection_policy still resolves to AskUser — noted in a code comment. Layer 2 (auto-approving null redirections) is intentionally out of scope and remains a separate discussion.

Related issue

Refs #1752 (not closing): this PR only fixes the Layer 1 misclassification; the issue's full auto-approve request is not resolved here.

User / Agent impact

The approval-card reason/impact for quoted null redirections now reads as a null redirection instead of redirection-write, aligning with the unquoted behavior from #1667. The execution decision (AskUser) is unchanged.

Risk and compatibility

• Public CLI, API, configuration, or documented behavior changed
• Privileged or security-sensitive behavior changed
• Cross-component contract changed
• Migration or rollback guidance is needed

Execution decision is identical before and after (AskUser); only the approval-card reason/impact changes. Low risk.

Validation

• ECS discriminative two-direction test: misclassification reproduces on pre-fix code, correct classification on fixed code.
• ECS cargo test: 1342 passed (baseline 1340 + 2 new).
• cargo fmt and clippy clean with zero warnings.

Documentation and rollback

Single commit — revert it directly.

[qoderai]

本次审查范围内未发现需要修改的问题。

---

🤖 Generated by Qoder • View workflow run

[SunnyQjm]

PR number: #2710
head_sha: 9bf29dc
reviewed_at: 2026-08-20T08:32:56Z

Findings

• [P1] CI 实测与 PR 验证声明冲突：Test cosh-ng（08:19:25Z）与 Test cosh-ng fast checks（08:20:04Z）在 head_sha 9bf29dc1 上均为 FAILURE，而 PR body 声称 ECS cargo test 1342 passed。静态评审无法判断失败原因（可能是慢 PTY 集成测试抖动，也可能是本 patch 引入的回归），但 Build cosh-ng release 成功、失败恰好落在被改 crate 的测试门上，合并前必须由作者给出失败日志解释或重跑转绿。
• [P2] command_risk_parser.rs 本次新增约 74 行（按 diff 增量估算）。该文件未登记于 large-file-inventory，输入中无其基线行数，无法确认叠加后是否越过 700 行阈值。按规则行数判定以 CI gate 实测为准，此处仅为估算提示：若实测越 700，需补 owner note 或拆分计划。

结构评审结论

未发现 blocking package/module/public API 组织问题。逐项核对：

• Owner 归属：三个文件全部落在 tools/（command risk 分类/解析/测试），符合 standard.md §2 中 tools/ 承担 command risk 的职责定义，无跨 owner 混杂。
• Public API：SAFE_OUTPUT_SINKS 仍为 private const，parse_command 维持 pub(super)，apply_null_redirection_policy 仅加注释；lib.rs 未触碰，inventory 分类计数不受影响。
• Root src / 入口文件：未新增 root implementation 文件，未触碰 lib.rs/main.rs。
• Forbidden dependency：未引入新 import，tools 未新增对 runtime/agent/ui 等方向的依赖。
• Self-crate path：未新增 cosh_shell::...。
• 测试位置：新增用例放入既有 tools/command_risk_tests.rs 决策矩阵，与 [cosh-ng] feat(cosh-ng): distinguish null redirection from filesystem writes #1667 边界测试职责一致。
• 范围：仅 cosh-shell crate 内变更，无跨 crate 混入。

安全语义复核（非 blocking）

本 PR 收窄 fail-closed 规则（整词带引号的 /dev/null 改判 null sink），属安全敏感分类器变更。复核认为方向保守可控：开引号须紧邻重定向目标位置、内容逐字节等于 allowlist 条目、闭引号后必须为词边界（刻意排除 {}）；未闭合引号、转义、后缀拼接（'/dev/null'x）、相邻引号续接、$/通配符、空引号、引号内空白全部落回 RedirectionWrite。正反两向测试齐备（含 &>'/dev/null'、argv 不泄漏断言、AskUser 不变的 V-M10），执行决策未变。map_or + #[allow(clippy::unnecessary_map_or)] 为 MSRV 1.74 兼容写法，已有注释说明。

剩余风险

• 内层内容扫描与 word_ends 各维护一份分隔符集合（{} 差异为刻意保守且有注释），后续修改需两处同步。
• PR body 声称 1342 passed（基线 1340 + 新增 2），与 diff 中两个新 #[test] 相符，但与上述 CI FAILURE 矛盾，以 CI gate 为准。

未跑验证

按评审约束未执行任何工具与构建命令；cargo test、check-layout.sh、inventory-public-api.sh、fmt/clippy 均未本地复验，全部以 CI gate 实测为准——当前 Test cosh-ng 两门为 FAILURE，是合并前必须清零的首要事项。

[jfeng18]

Both findings addressed in 7c77cb7 (single commit, force-pushed):

P1 (CI vs body) — two distinct failures, both explained:

• Test cosh-ng fast checks: the audit gate failed on the test-inventory overlap ratchet — the two new tests had landed in tools/command_risk_tests.rs, which is compiled into both the lib and bin test targets (main.rs re-declares lib modules), pushing the lib/bin overlap to 696 > 694 ceiling. They were moved to tools/command_risk_quoted_tests.rs, declared only from lib.rs via the existing wrap_tests pattern (the parser internals are pub(super), so a tests/ integration file cannot reach the assessment API either), restoring overlap = 694 = ceiling — the audit gate now passes locally.
• Test cosh-ng: provider_lifecycle::cosh_core_sync_drains_child_output_while_writing_large_prompt timed out after its 3s budget (71/72 pass on that job; the same code is green on main; zero interaction with this diff — it does not touch provider lifecycle). Will re-run the failed job off-peak to confirm the environmental attribution.

The quoted-sink behavior itself is unchanged; full ECS re-verification on the amended commit is green (fmt/clippy clean, --lib 1344 passed / 0 failed = baseline 1342 + the 2 relocated tests).

P2 (parser line-count registration) — Deferred with data: command_risk_parser.rs is 517 lines after this change (+99) and command_risk_build.rs is 691 (+5), both under the 700 threshold. Registration belongs to the large-file-inventory maintenance flow and will follow if and when an actual crossing occurs.

[SunnyQjm]

PR number: #2710
head_sha: 7c77cb7
reviewed_at: 2026-08-21T06:20:14Z

Findings

• [P3][advisory] src/cosh-ng/crates/cosh-shell/src/tools/command_risk_build.rs:510 本次 +5 行（纯注释），作者声明当前 691 行，距 700 行 owner-note 阈值仅剩约 9 行余量。行数判定以 CI gate 实测为准（本轮 fast checks 已 SUCCESS，未越线），此处仅提示后续增长需提前登记或拆分。
• 无 blocking findings。

结构评审结论

未发现 blocking package/module/public API 组织问题。逐项核对：

• Owner 归属：全部改动落在 tools/（command risk 解析/构建/测试），符合 standard.md §2 中 tools/ 承担 command classification/risk 的职责定义，无跨 owner 混杂。
• Public API：lib.rs 仅新增 #[cfg(test)] #[path] mod command_risk_quoted_tests;（私有、test-only），不是 pub mod/pub use，inventory 分类计数不受影响；SAFE_OUTPUT_SINKS 仍为 private const，parse_command 维持 pub(super)，未为测试新增 production pub（作者说明因 pub(super) 不可达，测试经 assessment API 断言可观察效果，此边界成立）。
• Root src / 入口文件：新测试文件位于 tools/ 下，未新增 root implementation 文件。
• Forbidden dependency / self-crate path：未引入新 import，未新增 cosh_shell::...。
• 范围：仅 cosh-shell crate 内变更，无跨 crate 混入，单一 commit。

上轮 findings 复核

• 上轮 P1（CI FAILURE 与 body 声明冲突）：当前 head 7c77cb727 上 Test cosh-ng（04:06:02Z）与 Test cosh-ng fast checks（04:04:49Z）均为 SUCCESS；作者对 overlap ratchet（696>694 回落至 694）与 provider_lifecycle 超时的解释合理，本 diff 不触碰 provider lifecycle。关闭。
• 上轮 P2（parser 行数登记）：作者给出实测 parser 517 行（+99）、build 691 行（+5），均低于 700 阈值，且 layout audit 已随 fast checks 转绿。关闭，保留上方 P3 余量提示。

安全语义复核（非 blocking）

收窄 fail-closed 规则的方向保守：开引号须紧邻重定向目标位置、内容逐字节等于 allowlist 条目、闭引号后必须为词边界（刻意排除 {} 且有注释）；未闭合引号、转义、后缀拼接（'/dev/null'x）、$/通配符、空引号、引号内空白均落回 RedirectionWrite。正反两向测试齐备（含 &>'/dev/null'、AskUser 不变断言），执行决策未变。#[allow(clippy::unnecessary_map_or)] 为 MSRV 1.74 兼容写法，已有注释。

剩余风险

• 内层内容扫描与 word_ends 的分隔符集合各维护一份（{} 差异为刻意保守），后续修改需两处同步。
• 测试文件通过 lib.rs #[path] 声明而非 tools/ owner 目录挂载，作者声明为既有 wrap_tests pattern 且服务于 overlap ratchet；测试位置最终口径以测试 SDD 与 CI 的 test-inventory gate 为准。

未跑验证

按评审约束未执行任何工具；cargo test、check-layout.sh、inventory-public-api.sh、fmt/clippy 均未本地复验，全部以 CI gate 实测为准——当前 head 上相关检查已全绿。