perf(r54): 把冻结成本批次与 hosted/CPU 诊断重放到 main（合并 #104–#107）

.github/workflows/core-candidate.yml

@@ -56,6 +56,8 @@ jobs:
           python-version: '3.12'
       - name: Install fixed semantic references
         run: python -m pip install --require-hashes -r experiments/differential/requirements.txt
+      - name: Install fixed CPU observer
+        run: python -m pip install -r benchmarks/requirements.txt
       - name: Select fixed Rust toolchain
         run: rustup default 1.92.0
       - name: Build locked candidate

.github/workflows/r54-cost.yml

@@ -0,0 +1,55 @@
+name: Frozen R54 hosted cost batches
+
+on:
+  push:
+    branches: ['perf/grapheme-hosted-r54']
+    paths:
+      - '.github/workflows/r54-cost.yml'
+      - 'experiments/measure/*.py'
+      - 'experiments/candidates/alacritty/Cargo.toml'
+      - 'experiments/candidates/alacritty/src/measure.rs'
+
+permissions:
+  contents: read
+
+concurrency:
+  group: r54-cost-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  fixed-cost:
+    # A code/evidence correction must not silently launch replacement samples.
+    if: github.event.created == true
+    runs-on: windows-2025
+    timeout-minutes: 30
+    env:
+      PYTHONUTF8: '1'
+      CARGO_HOME: ${{ github.workspace }}/local/r54-cargo-home
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 0
+          persist-credentials: false
+      - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0 # v7.1.0
+        with:
+          python-version: '3.12'
+      - name: Install fixed host observer
+        run: python -m pip install -r benchmarks/requirements.txt
+      - name: Select fixed compiler
+        run: rustup default 1.92.0
+      - name: Verify runner and publication contracts before measurement
+        run: python -m unittest experiments.measure.test_r54_contract.CostPlanTests experiments.measure.test_r54_batch experiments.measure.test_r54_hosted -v
+      - name: Build audited pair and execute the two frozen batches
+        id: measure
+        run: python -m experiments.measure.r54_hosted --out local/r54-raw
+      - name: Prepare the reviewed synthetic artifact whitelist
+        if: always() && steps.measure.outcome != 'skipped'
+        run: python -m experiments.measure.r54_publish --raw local/r54-raw --out local/r54-public
+      - name: Preserve complete or failed synthetic evidence
+        if: always() && steps.measure.outcome != 'skipped'
+        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
+        with:
+          name: r54-cost-${{ github.run_id }}-${{ github.run_attempt }}
+          retention-days: 30
+          if-no-files-found: error
+          path: local/r54-public/

.github/workflows/r54-cpu.yml

@@ -0,0 +1,50 @@
+name: Bounded R54 hosted CPU diagnosis
+
+on:
+  push:
+    branches: ['perf/grapheme-cpu-hosted-r54', 'perf/grapheme-cpu-hosted-r54-setup2', 'perf/grapheme-cpu-hosted-r54-entry3']
+
+permissions:
+  contents: read
+
+concurrency:
+  group: r54-cpu-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  cpu-boundary:
+    # Evidence or exporter fixes must not silently create replacement observations.
+    if: github.event.created == true
+    runs-on: windows-2025
+    timeout-minutes: 30
+    env:
+      PYTHONUTF8: '1'
+      CARGO_HOME: ${{ github.workspace }}/local/r54-cargo-home
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 0
+          persist-credentials: false
+      - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0 # v7.1.0
+        with:
+          python-version: '3.12'
+      - name: Install fixed CPU observer
+        run: python -m pip install -r benchmarks/requirements.txt
+      - name: Select fixed compiler
+        run: rustup default 1.92.0
+      - name: Verify build, accounting and publication boundaries
+        run: python -m unittest experiments.measure.test_r54_contract.CostPlanTests experiments.measure.test_r54_batch experiments.measure.test_r54_hosted experiments.measure.test_r54_cpu_diagnostic experiments.measure.test_r54_cpu_hosted -v
+      - name: Build audited pair and execute one six-observation diagnosis
+        id: diagnose
+        run: python -m experiments.measure.r54_cpu_hosted --out local/r54-cpu-raw
+      - name: Prepare the synthetic artifact whitelist
+        if: always() && steps.diagnose.outcome != 'skipped'
+        run: python -m experiments.measure.r54_publish --experiment cpu-boundary --raw local/r54-cpu-raw --out local/r54-cpu-public
+      - name: Preserve complete or failed synthetic evidence
+        if: always() && steps.diagnose.outcome != 'skipped'
+        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
+        with:
+          name: r54-cpu-${{ github.run_id }}-${{ github.run_attempt }}
+          retention-days: 30
+          if-no-files-found: error
+          path: local/r54-cpu-public/

CMakeLists.txt

@@ -351,7 +351,7 @@ add_test(NAME edit_csi_core COMMAND edit_csi_test)
   endif()
   termforge_add_python_test(core -m unittest experiments.core.test_core experiments.core.test_unicode -v)
   termforge_add_python_test(recovery -m unittest experiments.core.test_recovery -v)
-  termforge_add_python_test(measure -m unittest experiments.measure.test_measure experiments.measure.test_r54_contract.CostPlanTests -v)
+  termforge_add_python_test(measure -m unittest experiments.measure.test_measure experiments.measure.test_r54_contract.CostPlanTests experiments.measure.test_r54_batch experiments.measure.test_r54_hosted experiments.measure.test_r54_cpu_diagnostic experiments.measure.test_r54_cpu_hosted -v)
   termforge_add_python_test(binary_paths -m unittest experiments.test_build_paths -v)
   termforge_add_python_test(conpty_margin_contract -m unittest experiments.conpty.test_margin_contract -v)
   termforge_add_python_test(benchmark_tools -m unittest discover -s benchmarks/tests -v)

docs/EVIDENCE.md

@@ -83,6 +83,10 @@ R1 固定总出口的源码为 `ce9b24d`，早期 `2dc8d6d` / `9b16acc` 仍分
 | 持续输出中断的原失败记录 | [#36](https://github.com/JavaZeroo/termforge/issues/36)、[ConPTY 输入机制研究](research/conpty-ctrl-c-input.md)、[原 R1 历史证据](../experiments/shell/results/2026-10-23-r1-foundations/validation.json) | 原 Windows 11 回归中的 normal/resize 超时与 ETW/插桩观察；调试扰动及管道读完成缺口保留。不是当前全套结果，后续未复现不能抹去此失败。 |
 | controller B 正常结束后 cmd 退出未验收（#44） | [原失败脱敏记录](../experiments/shell/results/2026-10-26-controller-exit/validation.json)、[新退出诊断及有界重放](../experiments/shell/results/2026-10-26-controller-exit-observation/validation.json)、[关闭合同](https://github.com/JavaZeroo/termforge/issues/44) | 原 child 收完 32B、退出 0/join/Job0；cmd 只留下 close 后 0xC000013A，close 前未知。新观察分列输入、wait、pre-close 与清理；多轮未复现不构成历史修复，继续阻塞 #38。 |
 | 原 10 秒纯夹具超时的阶段保留（#40 未关闭） | [限定合同](research/inventory-timeout-checkpoints.md)、[正常与受控超时对照](../research/ime/results/2026-10-26-inventory-timeout-checkpoints/validation.json) | 保持原 10 秒期限；超时后保留有界阶段、原异常和回收。受控挂起对照已通过，但历史 CI 的阶段/根因仍未知，#40 不因此关闭。 |
+| 固定字素成本的批次协议与执行 | [R54 v1 合同和构建/环境记录](../experiments/measure/R54-COST.md)、[执行器](../experiments/measure/r54_batch.py)、[合成进程回归](../experiments/measure/test_r54_batch.py)、[实际预检及限定验证](../experiments/measure/results/2026-10-27-r54-runner/validation.json) | 固定两个种子、各 112 个位置，字节原样保留、首错停止、完整批次后逐组配对统计。共享工作站十区间活动预检未通过，未启动成本样本；测量相关 25 项测试通过。5% 是环境诊断护栏，并非产品性能容差；实际批次、扰动归因及用户预算确认由 #54 接续。 |
+| 首次 hosted 字素成本批次 | [来源、停止位置与归档](../experiments/measure/results/2026-10-27-r54-hosted/README.md)、[机器可读记录](../experiments/measure/results/2026-10-27-r54-hosted/validation.json) | 两份固定构建与4次sanity通过，空闲预检通过；第24位置的未归属CPU活动估计5.410%触发原5%护栏，保留23个有效记录和1个失败记录，第二批未启动。全部24次输出正确，但批次不完整，不生成配对成本摘要；临时VM逐进程归因、完整受控批次和预算仍未完成。 |
+| CPU计量边界诊断（#54） | [固定六位置及归档](../experiments/measure/results/2026-10-27-r54-cpu-boundary/README.md)、[独立复核](../experiments/measure/results/2026-10-27-r54-cpu-boundary/verify.py) | 同一共享Windows宿主、原已审计构建的六次输出和来源核验通过；分别保留完整子进程、内层区间、runner和全机计数。区间外CPU为0.078125–0.109375s，仍有明显未归属活动，不算空闲成本批次，不反推旧hosted原因或修改原5%护栏。 |
+| 新hosted VM的CPU分量（#54） | [预声明](../experiments/measure/R54-CPU-HOSTED.md)、[六位置/来源/完整归档与只读复核](../experiments/measure/results/2026-10-27-r54-cpu-hosted/README.md) | 新构建和六个输出均通过；完整扣除自身/runner后余额仍为3.333%–5.263%，不归因旧事件、不作成本验收。准备失败与未创建作业的入口分别保留，未重跑或补样。 |
 | Node 纯夹具阶段与失败判定（#144 未关闭） | [工具合同](research/node-contract-deadline.md)、[固定三 case 与身份](../research/ui/results/2026-10-29-node-contract-deadline/README.md) | 固定02aac源码的一次正式批次：正常passed0/23cp、受控挂起timed_out99/6cp、受控异常failed1/11cp，均根回收/Job0；CLI0只表示预定义结果齐全。协议尾、版本漂移及离线出口白名单P2的纯反例分列；出口普通/-O各15项纯测试，原件和旧公开字节保留。旧CI根因/整程硬界仍null，不接原CI执行入口，新增CTest仅注册纯数据回归，不关闭#144/#38。 |
 | 纯 AST 夹具的生命周期记录 | [#40](https://github.com/JavaZeroo/termforge/issues/40)、[计时、进度与回收验证](../research/ime/results/2026-10-26-inventory-lifecycle/validation.json)、[专用 CI 40253727745](https://github.com/JavaZeroo/termforge/actions/runs/40253727745) | 旧 10 秒 AST 夹具及专用 CI 记录启动、AST 和 Job0；原超时根因未确定。诊断增强见 #40 阶段保留，不由一次 CI 成功结项。 |
 | 原生 margin 逐字段合同与修复候选（#56 仍开放） | [合同、源码与路径](research/conpty-native-margin-path.md)、[独立 oracle](../experiments/conpty/margin_contract.py)、[固定运行](../experiments/conpty/results/2026-10-26-native-margin-path/validation.json) | 四阶段文字/格属性/光标独立判据；系统与官方原生 final 仍红，官方80列两次原5秒生命周期失败保留。共享单 margin 源码支持小 MIT fork 候选；patch 仅可应用，未编译、未运行、不计修复。 |
@@ -170,4 +174,4 @@ python -m unittest experiments.shell.test_workflow_sessions -v
 | 证据 | 来源/复现入口 | 可以支持的判断 |
 |---|---|---|
 | 固定功能和发行目标 | [固定 README](https://github.com/kingToolbox/WindTerm/blob/a8336c1d1e3dd1a981500dae8ef9eb42066de95f/README.md)、[2.7.0 release](https://github.com/kingToolbox/WindTerm/releases/tag/2.7.0)、[平台范围台账](../research/windterm-2.7.0-platform-scope.md)、[发行包校验流程](../benchmarks/tools/WINDTERM-LAB.md) | 官方功能声明与包身份；#23 已用固定源码、release API 和资产元数据固定四类发行平台/架构及声明下限，保留安装说明差异和未知系统。声明仍需运行时验收，不跟随未来 master 自动扩大范围 |
-| 已公开/未公开边界 | [源码与发行物调查](research/windterm-audit.md) | 固定公开源码主要是组件/修改，无法据此构建完整 WindTerm；静态依赖与符号不证明运行时实现或授权 |
+| 已公开/未公开边界 | [源码与发行物调查](research/windterm-audit.md) | 固定公开源码主要是组件/修改，无法据此构建完整 WindTerm；静态依赖与符号不证明运行时实现或授权 |

docs/FEATURES.md

@@ -80,7 +80,7 @@ Qt 构建显式准备并部署[固定微软 CRT 14.51.36247.0](../third_party/ms
 | 自动正确性与构建 | **限定验收**：Windows shell/ConPTY/呈现及 Windows/Linux 核心 CI、差分/恢复/队列/生命周期场景 | 产品打包、干净机器使用、依赖升级、长期回归和更广平台矩阵；#40 的 CI 纯夹具超时、窄窗口选择 #73、四 shell 原事件 #80、hosted 准备 #81 和[复制夹具 OpenClipboard 5 #85](https://github.com/JavaZeroo/termforge/issues/85) 等独立失败仍未闭合；#72 保存判据与 #80 同步条件的限定修正不代替其他配置或未知根因验收 | R2 / R4 |
 | Windows 产品交付 | **实现中，未验收**：首个 [v0.1.0-alpha.1 预发布](releases/v0.1.0-alpha.1.md)提供 Windows x64、静态 MSVC 运行库、附资源许可和构建身份的免安装 ZIP；未签名，已验环境限 Windows 11 本机与 Server 2025 CI，布局仍写入 LOCALAPPDATA | 这是 R1 体验包，尚非 R2 日常开发版；干净机器与更广平台矩阵、签名/安装、配置升级及持续运行仍待完成。Windows 10 未验收，Server 2022 已知资源门槛未过 | R2 / R3 |
 | Linux/macOS 与其他架构 | **未实现**客户端；Linux 核心 CI 已有。[官方平台范围台账](../research/windterm-2.7.0-platform-scope.md)已固定 2.7.0 的 Windows x86/x64、macOS x64、Linux x64 资产及声明下限；这不是 TermForge 支持或实测结论 | 最终需真实 Linux/macOS 客户端；平台调查可提前进行，整合及完整验收排在 Windows 11 x64 完整版本之后。其他系统版本、最低版本和架构由用户后续明确；旧系统及 ARM 等保持调查记录，不自动承诺或排除，也不能由核心 CI 推断产品支持 | 持续调查；R5 全平台出口 |
-| 吞吐、响应、内存与稳定性 | **实现中，未验收**：已有机制实验、连续 CPU 区间校准与合成回归；自主C++ [recovery双构建准备](../experiments/measure/results/2026-10-29-recovery-core-prepare/README.md)已核对包装有无、构建/对象身份及24短正确性样本，旧strict结果分列。候选字素改动的历史受限诊断中 project/interactive 吞吐中位数约下降 16%/9%，由 [#54](https://github.com/JavaZeroo/termforge/issues/54) 跟踪 | 共享宿主、历史allocator包装与缺少真实呈现/输入时延端点使旧诊断不构成正式预算；新双构建仅为准备，未形成正式成本样本。仍需同负载/同正确性、受控对照、观测开销和长期/多会话公平性；可靠原版基线后提出逐项预算及允许波动范围，**由用户明确确认**。与 WindTerm 的实际差距**待取证** | 当前优先；贯穿 R1–R5 |
+| 吞吐、响应、内存与稳定性 | **实现中，未验收**：已有机制实验、连续 CPU 区间校准与合成回归；候选字素改动的受限诊断中 project/interactive 吞吐中位数约下降 16%/9%，由 [#54](https://github.com/JavaZeroo/termforge/issues/54) 跟踪。[冻结批次执行器](../experiments/measure/R54-COST.md#冻结批次执行器windows)核对构建来源、原始输出、停止规则与逐轮配对统计；[独立 hosted 构建与实测](../experiments/measure/results/2026-10-27-r54-hosted/README.md)已验证来源和失败保留，但首批因环境护栏停止，尚无完整成本批次；[共享宿主CPU诊断](../experiments/measure/results/2026-10-27-r54-cpu-boundary/README.md)及[新hosted六位置](../experiments/measure/results/2026-10-27-r54-cpu-hosted/README.md)分别记录完整子进程/执行器分量；后者仍有3.333%–5.263%未归属余额，不改原护栏或判成本通过 | 旧诊断的共享宿主、allocator 包装与缺少真实呈现/输入时延端点使其不构成正式预算；新的不完整批次也不作吞吐结论。仍需同负载/同正确性、无插桩对照、观测开销和长期/多会话公平性；可靠原版基线后提出逐项预算及允许波动范围，**由用户明确确认**。与 WindTerm 的实际差距**待取证** | 当前优先；贯穿 R1–R5 |
 | 大历史与动态内存压缩 | **未实现**产品能力；WindTerm 压缩属于官方性能声明，内部机制尚未证实 | 1/10/100 万行等明确规模下的内存/滚动/搜索/淘汰实验；根据瓶颈决定独立存储/压缩实现 | R4 |
 | 完整目标验收 | **待取证**：已有固定版本和局部观察，没有完整行为/性能对照矩阵 | 先在 Windows 11 x64 完整验收功能/行为/UI/性能及独立构建，再完成 Linux/macOS 全平台出口；重要差异和性能容差须有用户确认，Windows 通过不代表全平台完成；不依赖 WindTerm 二进制或专有资源 | R5 分平台出口 |