fix: validate unfinished UI paste and bound replay cost (#53)

AGENTS.md

@@ -1,6 +1,11 @@
 # Working on TermForge
 
 - Read README.md, docs/PLAN.md, docs/FEATURES.md and the active milestone before making architecture decisions.
+- Follow docs/adr/0006-confirmed-project-direction.md: fully recreate WindTerm 2.7.0, with Windows 11 x64 as the first complete-version target and Windows/Linux/macOS as the final platform goal. Prioritize terminal foundations and performance before SSH/SFTP and the remaining features.
+- Preserve Alacritty as a candidate, not an approved product-core selection. Keep investigating the original code and mechanisms; internal implementation need not be identical.
+- Default to observed WindTerm 2.7.0 behavior when it conflicts with standards, third-party behavior or existing project policy. Unknown behavior requires investigation; important deviations and original-bug fixes require user confirmation.
+- The user confirms scope changes, major architecture/dependency choices, acceptance criteria and important behavior differences. Agents may autonomously research, experiment, implement, test and fix within confirmed scope. Record proposal/confirmed status and confirmation evidence in ADRs and issues; merged code or closed issues do not imply user approval.
+- Performance tolerances have no approved numeric values yet. Propose them from reliable baseline evidence and explicitly agree them with the user before using them for acceptance; keep existing diagnostic guards separate.
 - The project has a limited Windows local-shell prototype. Do not claim a daily-use release, Linux/macOS client, complete WindTerm parity or measured performance equivalence.
 - Do not lock the stack or add a large application skeleton before benchmark evidence supports it.
 - Every performance claim needs workload identity, environment, correctness checks, raw samples and limitations.
@@ -17,5 +22,5 @@
 - Keep build configurations separate in claims: default core/system backend, main candidate/system backend, and an explicit experimental backend need their own source and validation identity. Do not combine their independent successes into a joint acceptance result.
 - Source comments explain responsibilities, invariants and policy. Put changing status, batch names and future work in the owning issue; retain fixed experiment/protocol identities when they define behavior or evidence. A TODO must point to a current task rather than duplicate the roadmap.
 - Start only an issue whose real blockers are resolved. Record source observations separately from TermForge policies, implement independently, and validate the complete user-visible behavior. An unknown target behavior is an investigation item, not a passed comparison.
-- Before expanding terminal modes and common TUI implementation in R2, complete ADR-0003's actual external-core comparison; candidate build and adaptation feasibility can proceed during R1. Before product backend selection, revisit ADR-0004. Do not silently turn experimental choices into permanent architecture.
+- ADR-0003's historical external-core comparison remains evidence for continued candidate work, not user confirmation of a permanent route. Before final product core/backend/GUI selection, revisit ADR-0003/0004 under ADR-0006 and obtain the user's confirmation of a concrete evidence-backed proposal. Do not silently turn experimental choices into permanent architecture.
 - Documentation is primarily Chinese; technical identifiers and concise English summaries are welcome.

CMakeLists.txt

@@ -163,6 +163,9 @@ if(WIN32)
   target_compile_definitions(shell_query_client PRIVATE NOMINMAX WIN32_LEAN_AND_MEAN UNICODE _UNICODE)
   add_dependencies(termforge_shell shell_query_client)
   add_executable(shell_queue_probe experiments/shell/queue_probe.cpp)
+  add_executable(sender_fence_pipe_test experiments/shell/sender_fence_pipe_test.cpp)
+  termforge_warnings(sender_fence_pipe_test)
+  target_compile_definitions(sender_fence_pipe_test PRIVATE NOMINMAX WIN32_LEAN_AND_MEAN)
   add_executable(shell_size_client experiments/shell/size_client.cpp)
   add_executable(shell_input_client experiments/shell/input_client.cpp)
   add_executable(shell_text_client experiments/shell/text_client.cpp)
@@ -244,6 +247,16 @@ if(BUILD_TESTING)
       add_test(NAME candidate_title COMMAND candidate_title_test)
       set_tests_properties(candidate_title PROPERTIES TIMEOUT 35
         ENVIRONMENT "TERMFORGE_TEST_ARTIFACTS=${PROJECT_BINARY_DIR}/test-artifacts")
+      if(TERMFORGE_CONPTY_BACKEND_DIR)
+        add_executable(candidate_reply_close_test experiments/shell/candidate_reply_close_test.cpp experiments/shell/shell_launch.cpp)
+        termforge_warnings(candidate_reply_close_test)
+        target_compile_definitions(candidate_reply_close_test PRIVATE NOMINMAX WIN32_LEAN_AND_MEAN UNICODE _UNICODE)
+        target_link_libraries(candidate_reply_close_test PRIVATE termforge_core d2d1 dwrite gdi32 user32 imm32 ole32 oleaut32 uuid bcrypt)
+        termforge_candidate_shell(candidate_reply_close_test)
+        add_dependencies(candidate_reply_close_test termforge_shell_assets)
+        add_test(NAME candidate_reply_close COMMAND candidate_reply_close_test)
+        set_tests_properties(candidate_reply_close PROPERTIES TIMEOUT 30)
+      endif()
     endif()
     add_dependencies(runtime_observation_test termforge_shell_assets)
     add_test(NAME runtime_observation COMMAND runtime_observation_test)
@@ -299,6 +312,10 @@ if(BUILD_TESTING)
   add_test(NAME paste COMMAND paste_test)
   add_test(NAME sender_fence COMMAND sender_fence_test)
   set_tests_properties(sender_fence PROPERTIES TIMEOUT 10)
+  if(WIN32)
+    add_test(NAME sender_fence_pipe COMMAND sender_fence_pipe_test)
+    set_tests_properties(sender_fence_pipe PROPERTIES TIMEOUT 10)
+  endif()
   add_test(NAME clipboard_policy COMMAND clipboard_policy_test)
   add_test(NAME keyboard COMMAND keyboard_test)
   add_test(NAME text_action_policy COMMAND text_action_policy_test)
@@ -343,6 +360,7 @@ if(BUILD_TESTING)
   termforge_add_python_test(workflow_system_runner -m unittest research.ime.test_workflow_system_runner -v)
   set_tests_properties(workflow_system_runner PROPERTIES TIMEOUT 30)
   termforge_add_python_test(copy_error_frame -m unittest experiments.shell.test_copy_error_frame -v)
+  termforge_add_python_test(paste_fence_evidence -m unittest experiments.shell.test_paste_checks -v)
   termforge_add_python_test(resize_trace -m unittest experiments.shell.test_resize_trace -v)
   if(WIN32)
     termforge_add_python_test(research_guards -m unittest discover -s research/tests -v)
@@ -366,7 +384,7 @@ endif()
 if(WIN32)
   foreach(target termforge_conpty_probe conpty_ctrl_c_probe conpty_margin_probe conpty_backend_probe real_app_capture
                  termforge_shell workflow_fixture_test shell_queue_probe shell_paste_probe
-                 runtime_observation_test candidate_title_test raw_console_receiver_test)
+                 runtime_observation_test candidate_title_test candidate_reply_close_test raw_console_receiver_test)
     if(TARGET ${target})
       termforge_use_conpty_backend(${target})
     endif()

CONTRIBUTING.md

@@ -2,14 +2,16 @@
 
 R1 基础文字交互已按限定合同验收，当前推进 R2 的联合配置、兼容性、输入与交付验收，并行处理性能和环境缺口。
 
-开始前阅读[总体计划](docs/PLAN.md)、[当前能力](docs/FEATURES.md)、[R2 交付路线与验收](docs/milestones/R2-local-terminal.md)和[任务跟踪](docs/TRACKING.md)。已验基础合同见 [R1](docs/milestones/R1-local-text.md)；先核对 issue 的真实阻塞、目标构建配置和工作树，再选取范围明确的任务。PR 说明问题、最终行为、验证方法和限制，并更新对应能力行。大型依赖或技术栈变更需要记录 ADR 和实验依据。
+开始前阅读[总体计划](docs/PLAN.md)、[当前能力](docs/FEATURES.md)、[R2 交付路线与验收](docs/milestones/R2-local-terminal.md)和[任务跟踪](docs/TRACKING.md)。已验基础合同见 [R1](docs/milestones/R1-local-text.md)；先核对 issue 的真实阻塞、目标构建配置和工作树，再选取范围明确的任务。PR 说明问题、最终行为、验证方法和限制，并更新对应能力行。按 [ADR-0006](docs/adr/0006-confirmed-project-direction.md)，目标范围、主要技术路线、验收门槛和重要差异需用户确认；ADR 记录提议、证据和确认依据。研究、候选实验及已确认方向内的实现和修复可自主推进。
 
 规划文件各维护一种事实：能力在 FEATURES、路标范围在 PLAN、交付顺序和验收在当前路标、实时进度和依赖在 GitHub。TRACKING 和 EVIDENCE 分别提供任务与来源导航。源码注释说明机制与不变量，TODO 指向责任 issue；不要把“这一批已完成”“下一波”等动态进度写进实现。固定实验、协议和历史结果的身份应保留，不机械改名。
 
 ## 性能结果
 
 提供具体应用版本/构建、输入 SHA-256、环境、配置、重复样本和正确性检查。说明时间端点是生产完成、消费完成还是画面呈现；不能混为一谈。报告退化和波动，不只挑最佳结果。
 
+关键指标逐项对标 WindTerm 2.7.0，首个完整版本环境为 Windows 11 x64。先测可靠基线，再提出容差并与用户明确；未经确认的数值只作提案，不用于宣布性能验收通过。终端行为默认原版优先；与规范、第三方核心或项目策略冲突时保留原版证据和差异，明确缺陷的修正由用户确认。
+
 ## 来源与隐私
 
 遵守 [PROVENANCE](docs/PROVENANCE.md)。使用自生成输入与隔离测试配置，上传前检查用户名、路径、主机名、地址和凭据。第三方实现单独说明许可证；项目 MIT 许可不覆盖未获授权的第三方材料。

README.md

@@ -1,10 +1,12 @@
 # TermForge
 
-**独立、完整开源地复刻 WindTerm 2.7.0，以可复现证据验证功能、交互和性能。**
+**以完整开源、可独立构建的实现复刻 WindTerm 2.7.0，优先完成 Windows 11 x64，最终覆盖 Windows、Linux、macOS。**
 
-An independent, fully open-source terminal project targeting WindTerm 2.7.0 feature, interaction and performance parity. Not affiliated with WindTerm or KingToolbox.
+An independent, fully open-source terminal project targeting WindTerm 2.7.0 feature, behavior, UI and performance parity. Windows 11 x64 first; Windows, Linux and macOS ultimately. Not affiliated with WindTerm or KingToolbox.
 
-当前是有限 Windows 本地终端原型，**R1 基础文字交互已按限定合同验收，正在推进 R2 可靠本地终端与开发版**。主线已有候选核心路线、首批流式字素、有限多会话和配置恢复；候选核心与实验 ConPTY 后端的联合配置、完整 TUI/输入、真实不同 DPI、性能预算和开发包仍需闭合。现状见[能力矩阵](docs/FEATURES.md)，接续顺序见 [R2 交付路线与验收](docs/milestones/R2-local-terminal.md)。SSH/SFTP 属于 R3，完整 UI 与 WindTerm 性能等价均未验收。`TermForge` 是暂名。
+当前是有限 Windows 本地终端原型，**R1 基础文字交互已按限定合同验收，正在推进 R2 可靠本地终端与开发版**。主线已有 Alacritty 候选实验、首批流式字素、有限多会话和配置恢复；候选核心与实验 ConPTY 后端的联合配置、完整 TUI/输入、真实不同 DPI、性能预算和开发包仍需闭合。现状见[能力矩阵](docs/FEATURES.md)，接续顺序见 [R2 交付路线与验收](docs/milestones/R2-local-terminal.md)。SSH/SFTP 属于 R3，完整 UI 与 WindTerm 性能等价均未验收。`TermForge` 是暂名。
+
+方向已经用户确认，依据见 [ADR-0006](docs/adr/0006-confirmed-project-direction.md)：原版行为优先，内部结构不强求一致，但认真研究、借鉴原版；Alacritty 保留为候选，正式选型待证据和用户确认。关键性能指标分别验收，具体允许波动范围后续与用户明确。
 
 ## 开始使用
 
@@ -24,16 +26,17 @@ An independent, fully open-source terminal project targeting WindTerm 2.7.0 feat
 |---|---|
 | 已经做了什么、还缺什么 | [当前能力与目标矩阵](docs/FEATURES.md) |
 | 最终目标与五个路标（从远到近） | [总体计划](docs/PLAN.md) |
+| 哪些方向已确认、哪些决定仍需确认 | [方向与决策权限](docs/adr/0006-confirmed-project-direction.md) |
 | 官方发行平台、架构与声明的最低系统边界 | [2.7.0 平台范围台账](research/windterm-2.7.0-platform-scope.md) |
 | R2 分几步交付，每步还缺什么 | [R2 交付路线与验收](docs/milestones/R2-local-terminal.md) |
 | 已验收的基础文字交互合同 | [R1 基础文字交互](docs/milestones/R1-local-text.md) |
 | GitHub milestones、issues 和执行顺序 | [任务跟踪](docs/TRACKING.md) |
 | 结论依据、复现命令和全部文档 | [证据索引](docs/EVIDENCE.md) · [文档导航](docs/README.md) |
 
-交付主线：**基础文字交互 → 可靠本地终端 → SSH/SFTP 远程日常版 → 高级功能/跨平台/性能 → 完整 2.7.0 复刻验收**。性能工程和定向逆向贯穿各路标，不把远期整块调查当作近期开发的前置。
+交付主线：**终端基础能力与性能 → SSH/SFTP 远程工作流 → Windows 高级功能与性能收敛 → Windows 11 x64 完整复刻验收 → Linux/macOS 扩展与全平台验收**。保留 R1–R5 编号，R5 分 Windows 和全平台两个出口；性能工程、原版研究和平台可行性调查持续推进。
 
 ## 工程原则与来源
 
-功能先通过正确性验证；性能同时衡量吞吐、交互延迟、CPU、内存、稳定性和多会话公平性。不以丢数据、无界缓冲或积压呈现换取跑分，不预设某种语言或框架必然达标。优先复用许可明确的组件，以测量和可维护性决定产品架构。
+功能先通过正确性验证；性能同时衡量吞吐、交互延迟、CPU、内存、稳定性和多会话公平性。不以丢数据、无界缓冲或积压呈现换取跑分，不预设某种语言或框架必然达标。研究原版公开代码、运行行为与关键内部机制，并比较自主实现和许可明确的组件；主要技术路线、验收门槛和重要差异经用户确认后生效。
 
 原创代码和文档采用 [MIT](LICENSE)，第三方组件保留自身许可。实现不依赖 WindTerm 二进制或专有资源；观察、静态线索和推断分别记录。详见[来源规则](docs/PROVENANCE.md)与[性能规范](benchmarks/README.md)。

benchmarks/PROTOCOL.md

@@ -109,4 +109,4 @@ H01 搜索固定首部/中部/尾部命中及不存在目标各 10 次，冷/重
 
 吞吐和资源报告中位数、最小/最大、IQR、所有样本与失败率；原始批次保留。异常样本只能按预先记录的外部故障理由排除，报告排除前后结果。校准运行不能混入正式样本。
 
-M0 不给出“追平”结论。完成基线后，按每个关键场景制定绝对/相对预算和可接受差异，冻结后用于 M1/M2 对照。吞吐提高但尾延迟或正确性退化不能视为整体达标；没有证据的组件成本（如黑盒布局/解析耗时）必须保持未知。
+本协议的历史 M0 基线阶段不给出“追平”结论；当前执行使用 R1–R5 路标。按 [ADR-0006](../docs/adr/0006-confirmed-project-direction.md)，先建立 WindTerm 2.7.0 的可靠基线，提出每个关键场景的绝对/相对预算和允许波动范围，与用户明确并记录确认后冻结，用于后续验收。当前没有已批准的数值容差，采样噪声或自有回归护栏不能替代该确认，也不得根据待验结果反填门槛。首个完整版本对照环境为 Windows 11 x64，其他平台分别留证。吞吐提高但尾延迟或正确性退化不能视为整体达标；没有证据的组件成本（如黑盒布局/解析耗时）必须保持未知。

benchmarks/README.md

@@ -1,5 +1,7 @@
 # 性能基准
 
+当前优先建立 Windows 11 x64 上 WindTerm 2.7.0 的终端基础能力对照，吞吐、响应、CPU、内存及多会话分别留证。按 [ADR-0006](../docs/adr/0006-confirmed-project-direction.md)，具体允许波动范围在可靠基线后与用户明确，当前没有已批准的数值容差；阶段性诊断与性能验收分开。
+
 > M0/M1/M2 是协议 v0.1 的历史阶段名称；场景和测量约束继续有效，当前交付路标与执行顺序以 [R1–R5 计划](../docs/PLAN.md)及 [#3/#4 跟踪](../docs/TRACKING.md)为准。
 
 已定义 [M0 实验协议 v0.1](PROTOCOL.md)、[机器可读场景参数](cases.json) 与 [记录模板](run-template.json)。[工具使用说明](tools/README.md) 包含生成器、播放器、Windows 资源采样与自建窗口视觉诊断。已有[真实 WindTerm 诊断烟测](results/2026-10-21-windterm-smoke/README.md)，#3 的实际终端端点适配及正式开销校准仍未完成，尚无正式 WindTerm 性能基线。执行以 PROTOCOL.md 为准，本页为导航摘要。