build(desktop): Tier-1 payload slimming (~600MB; installer 1.14GB -> 810MB)

Prune build-only / duplicate / regenerable dirs from the staged runtime payload
(stage-runtime.cjs runtimePruneTargets) — all verified unused at USER runtime
(electron/service-manager.js spawns only python/java/jcmd/taskkill):
- appcds (230MB): stale build-machine CDS; regenerated per-user on first run
- wheels + bundle/wheels (246MB): offline pip wheels, no runtime/repair pip-installs;
  the two are byte-identical copies
- node (85MB): build/dev only, never spawned at runtime
- maven + maven-extract (11MB): build-time jar build only
- bundle/dist (31MB): web-mode build; desktop loads bundle/dist-file

They remain in local/workspace/runtime/windows for the build; only the payload is
slimmed. Guarded by release_selfcheck.py check_payload_slimmed() (FAILs if any
reappear in the staged payload).

Verified: dist:win selfcheck 6/6; installer 1.14GB->810MB, payload 2.6GB->2.0GB;
clean-machine cold/warm smoke functionally green (cold+warm backendReady+chartReady,
chart birth correct, forbiddenLogMatches empty, ports released). The smoke's exit-1
is only the warm >10s soft threshold (it force-kills, so the graceful AppCDS dump
never runs and there's no shipped CDS; real users regenerate on graceful close).
jlink DEFERRED (Spring reflection module-completeness risk).

NOT released: no version bump; batches with the next fix per user request.
Audit: docs/PACKAGING_SIZE_AUDIT.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Horace-Maxwell

.claude/skills/horosa-dev/SKILL.md

@@ -138,6 +138,9 @@ chart service `astropy/`. On Windows these live under
     its `release_selfcheck.py` sentinels guard `vendor/kinqimen/{jieqi.py(`zhirun_jieqi`),config.py(`dingju_jieqi`),kinqimen.py}`.
     (Mechanism deep-dive: `docs/qimen-calendar-fix-v2.1.6.md`.)
 
+11. **Payload size lever = `stage-runtime.cjs` `runtimePruneTargets`; don't re-ship build-only/duplicate dirs.**
+    The ~1.14GB installer is ~2.6GB unpacked, and **the embedded runtimes+data (`app-runtime`) are ~2.6GB of it — Chromium is only ~280MB.** Tier-1 slimming (~600MB, 2026-05-26) prunes, FROM THE STAGED PAYLOAD ONLY, dirs verified unused at user runtime: `appcds` (regenerated per-user first run), `wheels` + `bundle/wheels` (offline pip, never installed at runtime; the two are byte-identical copies), `node` (build/dev only), `maven`/`maven-extract` (build-time jar build only), `bundle/dist` (desktop loads `bundle/dist-file`). They stay in `local/workspace/runtime/windows` for the build; only the payload is slimmed. **Before pruning anything new, confirm `electron/` never spawns/reads it at runtime** (service-manager spawns only python/java/jcmd/taskkill). Guarded by `release_selfcheck.py` `check_payload_slimmed()` (FAILs if any reappear in the staged payload). **DEFERRED: jlink the JRE** (228→~60MB) — risky for a Spring Boot app (reflection/JNI module completeness); do it as a focused round with per-feature boot validation, not a blind prune. Biggest remaining wins are Mac-side (jar dep slimming, Python streamlit-stack, the 636MB ML models). Full breakdown + plan: `docs/PACKAGING_SIZE_AUDIT.md`.
+
 ## Commands
 
 ```powershell

desktop_installer_bundle/scripts/release_selfcheck.py

@@ -161,6 +161,25 @@ def check_distfile_not_stale():
     detail = "dist-file newer than astrostudyui/src" if ok else f"STALE: {os.path.relpath(src_p, REPO)} newer than staged dist-file (frontend not rebuilt)"
     record("staged dist-file not stale", ok, detail)
 
+def check_payload_slimmed():
+    # Tier-1 payload slimming (docs/PACKAGING_SIZE_AUDIT.md): these build-only / duplicate /
+    # regenerable dirs must NOT be in the staged runtime payload. If a future change re-includes
+    # them (e.g. the prune list in stage-runtime.cjs is reverted), the installer balloons ~600 MB.
+    stage = os.path.join(REPO, "desktop_installer_bundle", "build", "app-runtime", "runtime", "windows")
+    if not os.path.isdir(stage):
+        record("payload slimmed (Tier-1)", True, "SKIPPED (staged payload not built yet)")
+        return
+    must_be_absent = [
+        "node", "maven", "maven-extract", "wheels",
+        os.path.join("bundle", "wheels"), os.path.join("bundle", "dist"), "appcds",
+    ]
+    present = [p.replace("\\", "/") for p in must_be_absent if os.path.exists(os.path.join(stage, p))]
+    ok = not present
+    detail = "build-only/duplicate/regenerable dirs pruned from payload" if ok else \
+        f"NOT pruned -> ~600MB installer bloat: {', '.join(present)} (see stage-runtime.cjs runtimePruneTargets)"
+    record("payload slimmed (Tier-1)", ok, detail)
+
+
 def sha256(path):
     h = hashlib.sha256()
     with open(path, "rb") as f:
@@ -226,6 +245,7 @@ def main():
     check_sentinels()
     check_jar_not_stale()
     check_distfile_not_stale()
+    check_payload_slimmed()
     check_release_assets(V)
     width = max(len(n) for n, _, _ in results)
     failed = 0

desktop_installer_bundle/scripts/stage-runtime.cjs

@@ -33,6 +33,23 @@ const runtimePruneTargets = [
   'python/tcl',
   'java/jmods',
   'java/include',
+  // --- Tier-1 payload slimming (~600 MB): build-only / duplicate / regenerable artifacts that are
+  // NOT used at user runtime. Verified before adding (docs/PACKAGING_SIZE_AUDIT.md):
+  //   - electron/service-manager.js spawns only python/java/jcmd/taskkill — never node, never pip.
+  //   - AppCDS is regenerated per-user on first run (-XX:+RecordDynamicDumpInfo + jcmd VM.cds
+  //     dynamic_dump on shutdown); the shipped archives are just this build machine's warm cache.
+  //   - bundle/wheels is a byte-for-byte robocopy of wheels/ (Prepare_Runtime), and no runtime/repair
+  //     path pip-installs from either; maven is build-time only; the desktop loads bundle/dist-file
+  //     (file://), not the web-mode bundle/dist.
+  // These remain in local/workspace/runtime/windows for dev/build; only the shipped payload is slimmed.
+  // Guarded against regression by release_selfcheck.py check_payload_slimmed().
+  'node',
+  'maven',
+  'maven-extract',
+  'wheels',
+  'bundle/wheels',
+  'bundle/dist',
+  'appcds',
 ];
 
 function rmrf(targetPath) {

docs/PACKAGING_SIZE_AUDIT.md

@@ -0,0 +1,63 @@
+# Horosa Windows 打包体积审计 + 减重
+
+> 首次审计 / 实施:2026-05-26(基线 = v2.1.7 本地构建)。
+> 本文是「体积都去哪了 + 减了什么 + 还能减什么」的台账。数字为**实测**;省的估算标注「需确认用途」。
+
+## 基线(v2.1.7,减重前)
+
+- 安装包(NSIS,压缩):**~1.14 GB**
+- 解包后(`win-unpacked`):**~2.8 GB** —— 其中 Electron/Chromium 外壳仅 ~280 MB,**`app-runtime`(内置运行时+数据)占 2.6 GB**。Chromium 不是瓶颈。
+
+### 2.6 GB 载荷拆解(`build/app-runtime`,实测)
+| 块 | 大小 | 性质 |
+|---|---|---|
+| `project/astropy/astrostudy/models` | **636 MB** | 11 个 `.joblib` ML 模型(`virtual_body_corr_<行星>_v1`,每个 ~60MB) |
+| `runtime/windows/python` | 625 MB | 内置 Python + site-packages(含 streamlit/pyarrow/pydeck/altair ~150M) |
+| `runtime/windows/bundle/astrostudyboot.jar` | 310 MB | Spring Boot fat jar(opencv 105M + Oracle/Kafka/gRPC-xds 等服务端依赖) |
+| `runtime/windows/appcds` | 230 MB | Java CDS 存档(构建机多次运行累积的 5 个目录) |
+| `runtime/windows/java` | 228 MB | 内置 JDK/JRE(jmods/include 已 prune) |
+| `runtime/windows/wheels` + `bundle/wheels` | 246 MB | 离线 pip wheels,**同 119 文件打了两遍** |
+| Electron/Chromium 外壳 | ~280 MB | Horosa.exe 192M + locales 42M + paks/dll |
+| `project/flatlib-ctrad2` | 116 MB | flatlib + 150 个星历 `.se1` |
+| `runtime/windows/node` | 85 MB | 内置 Node |
+| `project/vendor` | 67 MB | kentang/kin 引擎 |
+| `bundle/dist` + `bundle/dist-file` | 62 MB | 前端构建产物**两份**(web 版 dist + 桌面版 dist-file) |
+| `runtime/windows/maven` | 11 MB | 构建工具 |
+
+## Tier 1 — 已实施(Windows 侧,零/低风险,~600 MB)
+
+**做法**:在 `desktop_installer_bundle/scripts/stage-runtime.cjs` 的 `runtimePruneTargets` 增加以下条目——只从**打包载荷**删除,`local/workspace/runtime/windows` 源保留(不影响构建)。每一项都**先核查确认用户运行时不用**才删:
+
+| 删除项 | 省 | 核查依据 |
+|---|---|---|
+| `appcds` | 230 MB | 用户首启自重建:`service-manager.js` 无存档时走 `-XX:+RecordDynamicDumpInfo`,关闭时 `jcmd VM.cds dynamic_dump`。shipped 的是构建机暖缓存(5 个 stale 目录) |
+| `wheels` + `bundle/wheels` | 246 MB | `electron/` 零引用 node/wheels/pip;无运行时/repair 走 pip 安装;`bundle/wheels` 是 `wheels` 的 robocopy 副本 |
+| `node` | 85 MB | `electron/` 与 `Prepare_Runtime` 均零引用,纯 dev/build |
+| `maven` + `maven-extract` | 11 MB | 仅 `Prepare_Runtime` 构建 jar 时用,用户运行时不用 |
+| `bundle/dist` | 31 MB | 桌面加载 `bundle/dist-file`(file://),web 版 dist 不用 |
+
+**防回弹闸门**:`release_selfcheck.py` 新增 `check_payload_slimmed()` —— 若上述目录重新出现在 staged 载荷里(prune 列表被还原),发布闸门直接 FAIL。
+
+**验证**:`dist:win` 重新打包后量体积 + `release_selfcheck.py` 全绿 + **干净机器冷/热启动 smoke**(证明删除未破坏启动:backendReady + chartReady + `forbiddenLogMatches` 空)。
+
+> 代价:删了预暖的 AppCDS,首次 warm-start 略慢(实测干净机器 smoke:冷 23.5s / 首热 11.3s,均 backendReady+chartReady、出盘正确、forbiddenLogMatches 空、端口干净——功能全绿;热 11.3s>10s 软阈值的 exit 1 是 smoke 强杀进程不触发 CDS dump + 删了预暖 CDS 所致,真实用户优雅关闭后自生成 CDS,后续热启回落)。其余皆无功能影响。
+>
+> **未来精修(可选)**:`appcds` 之所以有 230M,是构建机多次运行**累积了 5 个 stale 存档**。比「全删」更优的做法是构建期为当前 jar **只生成并随包一个新鲜 ~50M 存档**(保留 warm-start 预暖、又省 ~180M),但需要构建流程改造;本轮先全删求稳。
+
+## 已 deferred — jlink 裁 JRE(228M → ~60M,约 −160 MB)
+
+**本轮不做**,原因:jlink 会**替换** JRE,Spring Boot 大量反射 + JNI(opencv)/JDBC(sqlite、Oracle)/gRPC/POI 各需不同模块,`jdeps` 对 fat jar 推断不全,**漏一个模块 = 某功能运行时崩**,而冷/热 smoke 只覆盖出盘、未必触发所有后端路径——风险与刚修完的 issue #2(「我这能跑、用户那崩」)同类。建议作为**单独一轮**:配合后端逐功能验证,或在 Mac 侧连 jar 依赖一起做(确定模块集后再裁)。
+
+## Tier 2 — 待 Mac 侧改了再同步(共享代码,收益更大,需先确认用途)
+- **jar 瘦身**(`astrostudysrv` pom 排除/scope):Oracle `ojdbc8` / `kafka-clients` / `grpc-xds`/`grpc-netty-shaded` 桌面基本不需要(用 sqlite + mongo-fallback);`opencv-3.4.2`(105M)若无图像功能用到则剔。预计 −150~250 MB。
+- **Python 去 streamlit 全家桶**(`requirements`):streamlit/pyarrow/pydeck/altair(vendored 引擎当库用,非 streamlit app)。预计 −~150 MB。
+- **ML 模型 636M**(最大单块):压缩(`joblib compress=`,sklearn 模型常砍半)或改**首次使用按需下载**。预计 −300~636 MB。需产品确认「virtual body correction」定位。
+- 星历去重(astropy vs flatlib-ctrad2 两套 `.se1`)。
+
+## Tier 3 — 架构级
+- 桌面端去掉 Spring/JRE → 一次省 java 228 + jar 310 + appcds 0(已删) ≈ −540 MB,真重构。
+
+## 影响估算
+- **Tier 1（本轮，已实测）**:解包载荷 2.6 GB → **2.0 GB**;安装包 **1.14 GB → 810 MB（−29%）**。`release_selfcheck.py` 6/6 全绿。
+- + jlink:再 −160 MB。
+- + Tier 2(Mac 侧):再 −600 MB ~ −1 GB,安装包有望进 ~0.4 GB 区间。

docs/PROJECT_STRUCTURE.md

@@ -66,7 +66,7 @@
 
 - `package.json` / `package-lock.json`：桌面打包工程版本与依赖，当前发布版本为 `2.1.7`。
 - `electron/`：Electron 主进程、预加载脚本、加载/错误页（`loading.html`，含启动分步进度、失败态日志结尾透出与“复制诊断信息”）、窗口管理、Python/Java/Web 服务编排（`service-manager.js`，含崩溃归因与 trusted fast-path）、运行时健康检查与日志。
-- `scripts/`：renderer 构建、运行时 staging（`stage-runtime.cjs`，会注入 VC++ 运行时并跑原生依赖自检）、`check_runtime_native_deps.py`（构建期 DLL 完整性闸门）、`generate_brand_assets.py`（从 logo 生成图标/安装器位图）、`release_preflight.py`（发布前版本/品牌资产/运行时一致性闸门，对应 `npm run verify`）、项目路径解析、安装器回归、安装后桌面 smoke、干净机器冷/热启动检查和 kentang 端点验证。
+- `scripts/`：renderer 构建、运行时 staging（`stage-runtime.cjs`，会注入 VC++ 运行时、跑原生依赖自检，并按 `runtimePruneTargets` 从打包载荷裁掉构建期/重复/可重建产物——node、maven、wheels、appcds、web 版 dist 等，详见 `docs/PACKAGING_SIZE_AUDIT.md`）、`check_runtime_native_deps.py`（构建期 DLL 完整性闸门）、`generate_brand_assets.py`（从 logo 生成图标/安装器位图）、`release_preflight.py`（发布前版本/品牌资产/运行时一致性闸门，对应 `npm run verify`）、项目路径解析、安装器回归、安装后桌面 smoke、干净机器冷/热启动检查和 kentang 端点验证。
 - `assets/`：安装器图标与 NSIS 定制资源。唯一品牌源是 `horosa_setup_badge.png`（最新 logo）；`horosa_setup.ico`、`installerHeader.bmp`、`installerSidebar.bmp`、`uninstallerSidebar.bmp` 全部由 `scripts/generate_brand_assets.py` 从它派生，换 logo 后必须重跑该脚本。`installer.nsh` 为 NSIS 定制脚本（含安装目录校验、旧快捷方式清理）。
 - `release/`：本地构建输出目录。v2.1.7 Beta 标准发布资产为：
   - `Horosa-Setup-2.1.7.exe`

docs/SELFCHECK_LOG.md

@@ -2,6 +2,29 @@
 
 最后更新：2026-05-26
 
+## 2026-05-26 Tier-1 打包减重（Windows 侧，~600MB，未发布·待批量）
+
+把打包载荷里**构建期/重复/可重建**的产物从 staged payload 裁掉(只动打包载荷,`local/workspace/runtime/windows` 源保留,不影响构建)。**未升版本、未发布**,等下个修复一起发。完整审计见 `docs/PACKAGING_SIZE_AUDIT.md`。
+
+### 改动
+- `stage-runtime.cjs` `runtimePruneTargets` 增加:`node`、`maven`、`maven-extract`、`wheels`、`bundle/wheels`、`bundle/dist`、`appcds`(共 ~603MB)。
+- `release_selfcheck.py` 新增 `check_payload_slimmed()`:上述目录若重现在 staged 载荷里则 FAIL(防 prune 列表被还原导致体积回弹)。
+
+### 安全核查(删前逐项确认运行时不用)
+- `electron/service-manager.js` 只 spawn python/java/jcmd/taskkill——**零** node/wheels/pip 引用;`Prepare_Runtime` 也不在运行时引用 node。
+- AppCDS 用户首启自重建(`-XX:+RecordDynamicDumpInfo` + 关闭时 `jcmd VM.cds dynamic_dump`);shipped 的是构建机暖缓存(5 个 stale 目录)。
+- `bundle/wheels` 是 `wheels` 的 robocopy 副本;无运行时/repair 走 pip 安装。
+- maven 仅构建期建 jar 用;桌面加载 `bundle/dist-file`(file://)而非 web 版 `bundle/dist`。
+
+### 验证(实测)
+- 安装包:**1.14 GB → 810 MB(−29%)**;解包载荷 **2.6 GB → 2.0 GB**(packed tar 2600→1976 MB);win-unpacked 2.8→2.3 GB。
+- `release_selfcheck.py`:**6/6 全绿**(含新增 `payload slimmed (Tier-1)` PASS)。新 slim exe sha256 `c0312a3fd6ee65e2da7706002599f61432a77db7b2ab719a6c821a878292fd80`(**未发布,仅本地**;批量发布时会升版本重算)。
+- 7 个 prune 目录已确认从载荷消失;python/java/dist-file/jar 等保留项完好。
+- **干净机器冷/热启动 smoke**(`win-unpacked\Horosa.exe`,隔离 LOCALAPPDATA/APPDATA/TEMP):**功能全绿**——冷启 23.5s(extract 13.4s)+ 热启 11.3s,两者 `backendReady`+`chartReady` 均 true、出盘正确(birth 2028-04-06 09:33)、`forbiddenLogMatches` 空、8899/9999/9464 停止后均释放。脚本 exit 1 仅因热启 11.3s > 10s **软阈值**(同 2.1.4 类软失败):smoke 强杀进程→不触发优雅关闭的 AppCDS dump,叠加本轮删了预暖 CDS + 构建机负载,故首次热启略慢;真实用户优雅关闭后会生成自有 CDS,后续热启回落——**非功能回归**。(首次传错 exe 把安装器当 app 的那次失败已排除,纯属调用参数错误。)
+
+### deferred
+- jlink 裁 JRE(228→~60M)本轮不做:Spring 反射/JNI 模块完整性风险高、smoke 难覆盖,留作单独一轮(配合后端逐功能验证)。
+
 ## 2026-05-26 v2.1.7 Beta 奇门/三式 真太阳时定盘修复（纯前端）
 
 把 Mac main（`907e841` / v2.1.7）的 **奇门/三式 真太阳时定盘修复** 同步到 Windows。**纯前端 1 文件，无 Java/Python 改动 → 未重建 `astrostudyboot.jar`。**

local/workspace/runtime/windows/bundle/runtime.manifest.json

@@ -1,7 +1,7 @@
 {
     "product":  "Horosa Windows Stable",
     "channel":  "stable",
-    "generatedAt":  "2026-05-26T19:29:48-07:00",
+    "generatedAt":  "2026-05-26T21:13:58-07:00",
     "root":  "local/workspace",
     "assets":  [
                    {