@davideasden/pi-undo 0.2.5 → 0.2.7

package/README.md

@@ -15,7 +15,7 @@ Each completed agent run creates a checkpoint that captures both the Pi session
 - **External concurrency detection** — file fingerprint and inode checks detect external modification. Conflicting changes are never silently overwritten; the system fails closed or enters `recovery required`.
 - **No Git workflow** — snapshots use a private object database. No `git commit`, `git stash`, `git reset`, branches, or forges are required.
 - **Nested repositories** and **initialized submodules** are handled as independent roots. Their `.git` metadata is never modified.
-- **Performance** — batch WAL operations (up to 1,024 files per batch), scoped safety snapshots, prebuilt durable transaction packs, native Rust no-clobber file helper for macOS arm64, indexed WAL record and conflict lookups, and parallel manifest blob reads keep restore fast at scale. Unsupported platforms and failed integrity checks automatically use the TypeScript path.
+- **Performance** — batch WAL operations (up to 1,024 files per batch), scoped safety snapshots, prebuilt durable transaction packs, native Rust no-clobber file helper (macOS/Linux, arm64/x64), indexed WAL record and conflict lookups, and parallel manifest blob reads keep restore fast at scale. Unsupported platforms and failed integrity checks automatically use the TypeScript fallback.
 
 ## Requirements
 
@@ -23,28 +23,47 @@ Each completed agent run creates a checkpoint that captures both the Pi session
 - Node.js `22.19.0` or later.
 - Git available on `PATH` (used internally for content-addressed snapshots).
 
+### Rust 原生加速
+
+pi-undo 包含跨平台 Rust 原生 helper（`pi-undo-fs`），用于加速文件系统操作。发布包会同时包含以下六个预编译二进制：
+
+| 平台 | 架构 | 二进制名称 |
+|---|---|---|
+| macOS | arm64 | `pi-undo-fs-darwin-arm64` |
+| macOS | x64 | `pi-undo-fs-darwin-x64` |
+| Linux | arm64 | `pi-undo-fs-linux-arm64` |
+| Linux | x64 | `pi-undo-fs-linux-x64` |
+| Windows | arm64 | `pi-undo-fs-win32-arm64.exe` |
+| Windows | x64 | `pi-undo-fs-win32-x64.exe` |
+
+扩展会根据当前运行环境自动选择对应二进制，因此用户无需安装 Rust，也无需手动选择平台。Windows 二进制带有 `.exe` 后缀。对于尚未提供预编译二进制的平台，扩展会自动回退到 TypeScript 路径，功能仍然可用。
+
 ## Installation
 
-Install the published package from npm:
+### 从 npm 安装（推荐）
+
+所有受支持的平台使用同一个安装命令；npm 包会包含 macOS/Linux/Windows 的 arm64/x64 预编译 Rust helper，扩展启动时自动选择当前平台的版本：
 
 ```bash
 pi install npm:@davideasden/pi-undo
 ```
 
-Restart Pi if the extension is not already loaded.
+重启 Pi 即可加载扩展。用户不需要安装 Rust，也不需要手动选择或安装平台专用包。
 
-To install from a local checkout:
+### 从本地源码安装
 
 ```bash
 pi install /path/to/pi-undo
 ```
 
-During development, load the extension directly:
+### 开发模式直接加载
 
 ```bash
 pi -e /absolute/path/to/pi-undo/extensions/pi-undo.ts
 ```
 
+> **注意**：`pi install` 会将整个包（包括 `native/bin/` 下的预编译二进制）复制到 Pi 的扩展目录。如果你从源码构建后想要包含新编译的原生二进制，确保运行 `npm run build:native` 后再执行 `pi install`。
+
 ## Usage
 
 Work with Pi normally. `pi-undo` automatically records a boundary after each completed agent run.
@@ -227,6 +246,8 @@ A transaction directory may contain `descriptor.json`, `restore-plan.json`, `sta
 
 ## Development
 
+### 依赖
+
 Clone the repository and install dependencies:
 
 ```bash
@@ -235,6 +256,25 @@ cd pi-undo
 npm install
 ```
 
+### 构建 Rust 原生 helper
+
+如果需要构建或更新原生二进制，确保已安装 [Rust 工具链](https://rustup.rs/)：
+
+```bash
+npm run build:native
+```
+
+`npm run build:native` 会在 `native/pi-undo-fs/target/release/` 下生成当前构建平台的 `pi-undo-fs`。发布流程会在 macOS、Linux 和 Windows runner 上分别构建 arm64/x64 版本，统一重命名后放入 `native/bin/`，再打包成包含六个二进制的 npm 包。
+
+本地开发时，如果只需要验证当前平台，可以将生成的文件复制到 `native/bin/` 并按平台重命名，例如：Windows 生成的文件应使用对应的 `.exe` 文件名。
+
+```bash
+cp native/pi-undo-fs/target/release/pi-undo-fs native/bin/pi-undo-fs-darwin-arm64
+chmod +x native/bin/pi-undo-fs-darwin-arm64
+```
+
+发布包必须包含 Requirements 中列出的六个平台二进制；CI 会在打包前检查这一点，并在推送 `v*` tag 时发布该 CI 构建的完整 npm 包。npm 仓库需要为此 GitHub Actions workflow 配置 npm Trusted Publishing（OIDC）。如果当前平台没有对应二进制，扩展会自动使用 TypeScript 回退路径。
+
 ### Project Layout
 
 ```text
@@ -290,7 +330,54 @@ Benchmark tests assert Git call counts and WAL record counts, not wall-clock thr
 | 100-file restore (delete) — batch deletes | ≤12 | 600 | ~0.6s |
 | 4,000-file rollback snapshot — batch Git | ≤24 (4 x `mktree`, 4 x `commit-tree`) | — | ~7s |
 
-The 104-file standalone apply probe (10 warmup iterations + 10 measured) completes in approximately 3.9s post-optimization on the TypeScript path (down from 14.3s before batching), and approximately 2.8s on the native Rust path with durable pack caching and indexed blob reads.
+The 104-file standalone apply probe (10 warmup iterations + 10 measured) completes in approximately 3.9s post-optimization on the TypeScript path, and approximately 2.8s on the native Rust path.
+
+### Performance Optimizations
+
+Several hot paths were profiled and corrected from quadratic (n doubled ≈ 4× cost) to near-linear scaling. Each change was verified against the existing test suite (425+ tests) plus fault-injection and adversarial-diff tests. Optimization order was guided by real profile data rather than static code review alone. After this phase, attention shifted to the native Rust helper path, which dominates production workloads.
+
+#### Wall-clock comparison (n changed files + n ignored files)
+
+**Plan phase** — undo diff derivation and workspace topology check:
+
+| n | Before | After | Improvement |
+|---|---|---|---|
+| 1,000 | ~178ms | ~131ms | 1.4× |
+| 2,000 | ~399ms | ~259ms | 1.5× |
+| 5,000 | ~2.5s | ~699ms | ~3.6× |
+
+**Bidirectional durable pack prepare** — agent-settled caching for undo/redo:
+
+| n | Before (no batch) | After (batched) | Improvement |
+|---|---|---|---|
+| 500 | ~949ms | ~249ms | 3.8× |
+| 1,000 | ~2,940ms | ~386ms | 7.6× |
+| 2,000 | ~10,512ms | ~677ms | 15.5× |
+
+**Native apply** — the production hot path:
+
+| n | TypeScript fallback | Native Rust helper | Improvement |
+|---|---|---|---|
+| 500 | ~2,572ms | ~253ms | 10.2× |
+| 1,000 | ~5,726ms | ~374ms | 15.3× |
+| 2,000 | ~10,521ms | ~646ms | 16.3× |
+
+#### Algorithm microbenchmarks
+
+| Scenario | Before | After | Improvement |
+|---|---|---|---|
+| 8,000×8,000 path overlap check | ~1,352ms | ~7ms | **193×** |
+| 4,000-directory ignored prefix scan (plan) | ~544ms | ~161ms | **3.4×** |
+| 5,000×5,000 manifest integrity verify | ~610ms | ~7ms | **87×** |
+
+#### Key enablers
+
+- **Manifest context batching** (`SnapshotStore.readBlobs`): one manifest read and full revalidation serves all blob fetches per operation instead of one manifest load per blob. 2,000-file pack creation dropped from 10.5s to 796ms.
+- **Blob read batch sizing** (`BLOB_BATCH_MAX_ENTRIES`: 256 → 2,048): 40 `cat-file --batch` processes collapsed to 6 for 5,000 files by letting the 16MiB byte budget drive batch boundaries.
+- **Path overlap indexing** (`pathSetsOverlap` in `path-safety.ts`): sorted binary search plus ancestor enumeration replaces a nested `some()` pattern. 8,000×8,000 path pairs from 1.35s to 7ms.
+- **Ignored-proof prefix index** (`IgnoredProofIndex` in `restore-engine.ts`): lazily sorted binary lower-bound search for directory prefix lookups replaces a full-set spread per candidate. 4,000 directories from 544ms to 161ms.
+- **WAL ordinal indexing** (`loadOrdinal` in `mutation-journal.ts`): direct array access by contiguous ordinal replaces linear `.find()` throughout quarantine and legacy recovery.
+- **Native Rust no-clobber helper** (`native/pi-undo-fs`): platform native hardlink operations with EEXIST guard. Processes in ~0.3ms per file vs ~5.3ms for the TypeScript fallback, with WAL materialized from the durable pack.
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@davideasden/pi-undo",
-	"version": "0.2.5",
+	"version": "0.2.7",
 	"description": "Persistent workspace undo and redo for Pi",
 	"type": "module",
 	"keywords": [

package/src/native-restore.ts

@@ -84,10 +84,12 @@ export async function createNativeFileBatch(options: {
 function nativeExecutable(): string | undefined {
 	const platform = process.platform === "darwin"
 		? "darwin"
-		: process.platform === "linux" ? "linux" : undefined;
+		: process.platform === "linux" ? "linux"
+		: process.platform === "win32" ? "win32" : undefined;
 	const architecture = process.arch === "arm64" ? "arm64" : process.arch === "x64" ? "x64" : undefined;
 	if (platform === undefined || architecture === undefined) return undefined;
-	return fileURLToPath(new URL(`../native/bin/pi-undo-fs-${platform}-${architecture}`, import.meta.url));
+	const extension = process.platform === "win32" ? ".exe" : "";
+	return fileURLToPath(new URL(`../native/bin/pi-undo-fs-${platform}-${architecture}${extension}`, import.meta.url));
 }
 
 function runNative(executable: string, requestPath: string, expected: number): Promise<void> {

package/src/restore-engine.ts

@@ -169,7 +169,7 @@ export class RestoreEngine {
 			this.readOwnedPaths(current, canonicalScopePaths),
 			this.readOwnedPaths(target, canonicalScopePaths),
 		]);
-		const targetIgnoredPaths = ignoredWorkspacePaths(target);
+		const targetIgnoredProof = new IgnoredProofIndex(ignoredWorkspacePaths(target));
 		const deleteByRoot = new Map<string, string[]>();
 		const writeByRoot = new Map<string, string[]>();
 
@@ -183,7 +183,7 @@ export class RestoreEngine {
 				continue;
 			}
 			if (targetOwned === undefined) {
-				if (isProtectedByIgnoredProof(path, owned.entry.kind, targetIgnoredPaths)) {
+				if (targetIgnoredProof.isProtected(path, owned.entry.kind)) {
 					continue;
 				}
 				appendPath(deleteByRoot, owned.root.relativeRoot, path);
@@ -1728,15 +1728,42 @@ function ignoredWorkspacePaths(manifest: SnapshotManifest): Set<string> {
 	));
 }
 
-function isProtectedByIgnoredProof(
-	path: string,
-	kind: RestorePath["kind"],
-	ignoredPaths: ReadonlySet<string>,
-): boolean {
-	if (kind !== "directory") {
-		return ignoredPaths.has(path);
+/**
+ * ignored 证明的前缀索引。
+ *
+ * 目录判定需要回答"是否存在以 `${path}/` 开头的 ignored 路径"。逐次线性扫描
+ * 整个 set 会让"目录数 × ignored 数"变成平方项，因此这里预排序一次，
+ * 之后每次判定用二分下界定位第一个不小于前缀的元素。
+ *
+ * 语义与线性扫描完全一致：只回答存在性，不改变 fail-closed 行为，也不放宽
+ * 任何 ignored 保护。非目录仍走精确 `has()`。
+ */
+class IgnoredProofIndex {
+	private readonly exact: ReadonlySet<string>;
+	private sortedPaths: readonly string[] | undefined;
+
+	constructor(paths: ReadonlySet<string>) {
+		this.exact = paths;
+	}
+
+	isProtected(path: string, kind: RestorePath["kind"]): boolean {
+		if (kind !== "directory") {
+			return this.exact.has(path);
+		}
+		if (this.exact.size === 0) return false;
+		// 排序成本只在第一次目录判定时付出，纯文件计划完全不触发。
+		this.sortedPaths ??= [...this.exact].sort();
+		const sorted = this.sortedPaths;
+		const prefix = `${path}/`;
+		let low = 0;
+		let high = sorted.length;
+		while (low < high) {
+			const middle = (low + high) >>> 1;
+			if (sorted[middle]! < prefix) low = middle + 1;
+			else high = middle;
+		}
+		return low < sorted.length && sorted[low]!.startsWith(prefix);
 	}
-	return [...ignoredPaths].some((ignoredPath) => ignoredPath.startsWith(`${path}/`));
 }
 
 function rootBoundaryDirectories(root: string): string[] {