carto-md 2.0.7 → 2.0.8

package/README.md

@@ -17,7 +17,7 @@ carto init
 
 That's it. Carto auto-wires itself into every AI tool you have installed. Restart your AI tool and it now knows your codebase architecturally.
 
-**Works with:** Cursor · Claude Code · Kiro · Claude Desktop · Windsurf · VS Code Copilot · JetBrains · Zed
+**Works with:** Cursor · Claude Code · Codex · Kiro · Claude Desktop · Windsurf · VS Code Copilot · JetBrains · Zed
 
 ---
 
@@ -34,7 +34,7 @@ That's it. Carto auto-wires itself into every AI tool you have installed. Restar
 
 ### Claude Code (CLI)
 
-One command from your project root:
+`carto init` writes `<project>/.mcp.json` for you when Claude Code is detected (`claude` binary on PATH or `~/.claude/` exists). Manual:
 ```bash
 claude mcp add carto -- carto serve
 ```
@@ -43,6 +43,21 @@ Or create `.mcp.json` at the project root:
 { "mcpServers": { "carto": { "command": "carto", "args": ["serve"] } } }
 ```
 
+### Codex (CLI)
+
+`carto init` writes `~/.codex/config.toml` for you when Codex is detected (`codex` binary on PATH or `~/.codex/` exists). Manual:
+```bash
+codex mcp add carto -- carto serve
+```
+Or edit `~/.codex/config.toml`:
+```toml
+[mcp_servers.carto]
+command = "carto"
+args = ["serve"]
+cwd = "/your/project"
+enabled = true
+```
+
 ### Kiro
 
 `carto init` writes this for you. Manual: `~/.kiro/settings/mcp.json`
@@ -52,21 +67,25 @@ Or create `.mcp.json` at the project root:
 
 ### Claude Desktop
 
-`carto init` writes this for you. Manual: `~/Library/Application Support/Claude/claude_desktop_config.json`
+`carto init` writes this for you (cross-platform). Manual paths:
+- macOS: `~/Library/Application Support/Claude/claude_desktop_config.json`
+- Windows: `%APPDATA%\Claude\claude_desktop_config.json`
+- Linux: `~/.config/Claude/claude_desktop_config.json` (community Linux builds)
+
 ```json
 { "mcpServers": { "carto": { "command": "carto", "args": ["serve"], "cwd": "/your/project" } } }
 ```
 
 ### VS Code Copilot
 
-`.vscode/mcp.json` in your project root, or Command Palette → `MCP: Add Server`
+`carto init` writes `<project>/.vscode/mcp.json` for you when the `code` binary is on PATH. Manual: `.vscode/mcp.json` in your project root, or Command Palette → `MCP: Add Server`. Note: VS Code uses `servers` (not `mcpServers`) and requires `"type": "stdio"`.
 ```json
-{ "servers": { "carto": { "command": "carto", "args": ["serve"], "cwd": "/your/project" } } }
+{ "servers": { "carto": { "type": "stdio", "command": "carto", "args": ["serve"] } } }
 ```
 
 ### Windsurf
 
-`~/.windsurf/mcp.json`
+`carto init` writes this for you when Windsurf is detected. Manual: `~/.codeium/windsurf/mcp_config.json`
 ```json
 { "mcpServers": { "carto": { "command": "carto", "args": ["serve"], "cwd": "/your/project" } } }
 ```
@@ -94,6 +113,30 @@ Without carto, your AI greps text and guesses. With carto, it queries structural
 
 Your AI's response goes from "here's a guess" to "here's the right change with the architectural context."
 
+### Catching bad changes before save
+
+The harder problem isn't finding the right file — it's stopping a confident-sounding AI from writing a refactor that breaks the rest of the repo. Carto exposes a diff-shaped query the AI can call before the user accepts a change:
+
+> *AI proposes a 12-line patch to `packages/pg-meta/src/pg-format/index.ts`. Before showing the diff, it calls `validate_diff(patch)`.*
+
+```
+# Diff Validation
+
+**Risk:** 🔴 HIGH
+**Files changed:** 1
+**Union blast radius:** 83 transitive dependents
+
+## Violations (1)
+
+| Severity | Kind        | File                                   | Detail                                                          |
+|----------|-------------|----------------------------------------|-----------------------------------------------------------------|
+| HIGH     | high_blast  | `packages/pg-meta/src/pg-format/index.ts` | Modifying this file affects 83 transitive dependents (>50). |
+```
+
+The AI sees this *before* it proposes the diff. It revises its plan, splits the change, or asks the user. The bad refactor never makes it to the screen. Sub-millisecond on a 7,000-file repo — see the **Benchmarks** section below.
+
+Every `validate_diff` call is also written to a local SQLite log, so a session that runs five hours later can ask `did_we_discuss_this("snake_case naming")` and get back the prior decision. The AI stops re-deciding settled questions.
+
 ### In the wild
 
 ![Claude Code reviewing the supabase repo through carto's MCP server](docs/screenshots/claude-code-supabase.png)
@@ -192,7 +235,7 @@ Carto supports any LLM provider — configure in your editor:
 
 ---
 
-## The 16 MCP tools
+## The 22 MCP tools
 
 Once carto is wired in, your AI tool can call any of these mid-task. You don't need to memorize them — your AI will pick the right ones.
 
@@ -201,6 +244,8 @@ Once carto is wired in, your AI tool can call any of these mid-task. You don't n
 | `get_architecture()` | 500-word project overview: domains, entry points, stack, key patterns. **Use this first.** |
 | `get_change_plan(intent)` | Given "add rate limiting to /api/users" → files to touch, domains affected, blast radius, similar patterns |
 | `get_blast_radius(file)` | All files affected by changing a given file, with hop distance |
+| `simulate_change_impact(files)` | Union of all files transitively affected by changing **multiple** files at once. Powered by the bitmap engine — sub-millisecond on 7K-file repos. |
+| `validate_diff(diff)` | Given a unified diff: violations (cross-domain imports, high-blast files), blast radius per file, risk level (SAFE/LOW/MEDIUM/HIGH), suggestions. Each call is recorded in the **episodic memory** log so other tools can ask "did we discuss this?". Sub-15ms p99. |
 | `get_context(file)` | Everything about a file: domain, blast radius, neighbors, routes, models |
 | `get_file_summary(file)` | What a file does, its role, key deps and dependents |
 | `get_similar_patterns(file)` | Files with same domain, route shape, or shared deps — find conventions before writing new code |
@@ -214,15 +259,25 @@ Once carto is wired in, your AI tool can call any of these mid-task. You don't n
 | `get_high_impact_files(n)` | Top N files by blast radius |
 | `get_env_vars(domain?)` | All env vars with domain mapping |
 | `get_domains_list()` | All detected domains with file, route, model counts |
+| `get_recent_decisions(time_range, kind?)` | Recent validation decisions and architectural choices the AI has made in this project |
+| `get_session_context(session_id?)` | Full context for an AI session: every decision and intervention, ordered chronologically |
+| `did_we_discuss_this(topic)` | Substring search over the episodic memory log — avoid re-deciding settled questions |
+| `get_intervention_history(file?)` | Past Carto-issued violations and suggestions, optionally filtered by file |
 
 **All MCP queries:** **<5ms** on every benchmarked repo.
 
+## Episodic Memory
+
+Carto remembers every diff it validates. The `validate_diff` tool writes one row per call into a local SQLite log (`ai_sessions`/`decisions`/`interventions` tables) — so a session that runs five hours later can still ask `did_we_discuss_this("snake_case naming")` and get back the prior decision. The log lives next to the index in `.carto/carto.db` — never sent over the network, never shared between projects.
+
 ---
 
 ## Domain detection
 
 Carto uses **Leiden+CPM graph clustering** — files that import each other heavily cluster together. Domain names are inferred from path tokens, with keyword hints for well-known patterns (AUTH, PAYMENTS, DATABASE, etc.).
 
+**Adaptive strategy:** Repos under 100 files use keyword-only clustering (avoids over-fragmentation). Larger repos with dense import graphs get graph-based clustering with a gamma that scales continuously with repo size.
+
 Works on any repo — not just SaaS apps. vscode gets AUTH/EVENTS/DATABASE. zed (Rust) gets DATABASE/AUTH/EVENTS. A game engine would get RENDERER/PHYSICS/AUDIO.
 
 Custom domains via `carto.config.json`:
@@ -235,55 +290,101 @@ Custom domains via `carto.config.json`:
 }
 ```
 
+Full schema with anchor pinning (forces files into a domain regardless of clustering):
+```json
+{
+  "domains": {
+    "AUTH": {
+      "keywords": ["auth", "login", "session"],
+      "anchor": ["src/auth/session.ts", "src/auth/middleware.ts"]
+    }
+  }
+}
+```
+
+**Stability tracking:** Carto tracks domain assignments across syncs. If >5% of files change domain, `carto check` flags it as unstable. Two consecutive syncs with no code changes always produce 0% drift.
+
 ---
 
 ## CLI commands
 
 | Command | What it does |
 |---------|-------------|
-| `carto init` | Detect project, index codebase, generate AGENTS.md, auto-wire MCP into every AI tool found |
-| `carto sync` | Full re-index (skips unchanged files via mtime+size cache) |
-| `carto watch` | Incremental live re-index on every file save (<50ms) |
-| `carto serve` | Start MCP server (called by your AI tool — usually you don't run this directly) |
+| `carto init` | Detect project, index codebase, generate AGENTS.md, install git hooks (pre-commit, post-checkout, post-merge, post-rewrite), auto-wire MCP into every AI tool found |
+| `carto sync` | Full re-index (skips unchanged files via mtime+size cache). Called automatically by git hooks on commit/checkout/merge/rebase. |
+| `carto serve` | Start MCP server (called by your AI tool — usually you don't run this directly). On every file-aware query the server mtime+size-checks the file and re-parses inline if stale. |
+| `carto watch` | **Optional.** Live re-index on every file save. Not required — git hooks + lazy MCP re-parse keep the index fresh by default. Use only for AI-heavy workflows that write 50+ files between commits. |
 | `carto agent` | Start ACP agent mode (for Zed / JetBrains / VS Code) |
 | `carto impact <file>` | Blast radius: risk level, affected files, routes at risk |
 | `carto check` | Cross-domain violations, high-risk uncommitted changes, domain health |
+| `carto inspect` | Read-only diagnostic: index paths, sizes, freshness, bitmap sidecar shape, top-impact files, schema version, sync timestamps. `--json` for piping into `jq`. Never triggers a rebuild. |
 | `carto remove` | Remove AGENTS.md and .carto/ from project |
 
 ---
 
 ## Benchmarks
 
-Measured on real open-source repos. Apple M-series, 8 CPUs, 8GB RAM. SHAs pinned in `~/carto-test-repos`.
+Measured on real open-source repos. Apple M-series, 8 CPUs, 8GB RAM. SHAs pinned in `~/carto-test-repos`. Reproducible via `~/carto-test-repos/run-bench.sh`.
+
+### Indexing speed
 
 | Repo | Language | Indexed Files | First Run | Second Run | DB Size | Import Edges |
 |------|----------|---------------|-----------|------------|---------|--------------|
-| [prisma/prisma](https://github.com/prisma/prisma) | TypeScript | 961 | **649ms** | **110ms** | 0.7 MB | 1,387 |
-| [supabase/supabase](https://github.com/supabase/supabase) | TypeScript | 6,259 | **4.8s** | **751ms** | 4.0 MB | 5,321 |
-| [microsoft/vscode](https://github.com/microsoft/vscode) | TypeScript | 7,567 | **7.8s** | **387ms** | 6.7 MB | 13,420 |
-| [zed-industries/zed](https://github.com/zed-industries/zed) | Rust | 1,752 | **3.0s** | **182ms** | 4.4 MB | 2,113 |
+| [prisma/prisma](https://github.com/prisma/prisma) | TypeScript | 961 | **961ms** | **431ms** | 0.7 MB | 1,387 |
+| [supabase/supabase](https://github.com/supabase/supabase) | TypeScript | 6,330 | **5.4s** | **1.2s** | 4.0 MB | 5,321 |
+| [microsoft/vscode](https://github.com/microsoft/vscode) | TypeScript | 7,567 | **7.7s** | **1.0s** | 6.7 MB | 13,420 |
+| [zed-industries/zed](https://github.com/zed-industries/zed) | Rust | 1,752 | **3.0s** | **491ms** | 4.4 MB | 2,113 |
 
 **Indexed Files** counts what Carto actually parses — `.ts/.js/.py/.go/.rs/...` after excluding `node_modules`, build output, and per-file `*.test.*` / `*.spec.*` / `*.stories.*` / `test_*.py` patterns. The on-disk file count of the repo is larger.
 
-**Second run** = only changed files re-parsed. mtime+size checked before reading content — if nothing changed, nothing is re-parsed.
+**Second Run** = `carto sync` after no changes. mtime+size checked before reading content — if nothing changed, nothing is re-parsed.
+
+### MCP query latency (bitmap engine vs SQLite)
+
+Carto's MCP query path is bitmap-backed on five tools, plus a sixth (`simulate_change_impact`) that's only feasible with bitmap OR-aggregation. Speedups measured against the same data, same DB, same machine — bitmap path vs the equivalent SQLite path.
+
+| Tool | vscode (7,567 files) | Speedup vs SQLite |
+|------|---------------------:|------------------:|
+| `get_blast_radius` | sub-ms | **10.7×** |
+| `get_cross_domain` | 2.1ms | **6.2×** |
+| `get_high_impact_files` | sub-ms | **559×** |
+| `get_similar_patterns` | sub-ms | **73×** |
+| `simulate_change_impact` | sub-ms | **6.5×** (no SQLite equivalent at this latency) |
+
+Median speedup across all five tools on vscode: **10.7×**. Smaller repos hit higher peaks — laravel-framework `get_high_impact_files` clocks at 155× on the smaller graph. Reproducible via `npm run bench:bitmap -- --repo <path>`.
+
+### `validate_diff` latency
+
+The new diff-shaped query that lets the AI ask "is this patch safe?" before showing it to the user. Profiled with a representative 20-line diff against 20 random mid-blast-radius files per repo, 1000 calls each.
+
+| Repo | Files | p50 | p99 |
+|------|-------|----:|----:|
+| supabase | 6,259 | **0.082ms** | **0.298ms** |
+| vscode | 7,567 | **0.084ms** | **0.489ms** |
+
+Budget was p50 ≤ 5ms, p99 ≤ 15ms. Both targets are cleared by 30-60×. The bitmap engine handles every blast-radius and cross-domain query in microseconds; what's left is diff parsing + result aggregation. Reproducible via `node bench/validation-perf/index.js --repo <path>`.
 
 ### Domains detected
 
 | Repo | Domains |
 |------|---------|
-| prisma | CORE · DATABASE · EVENTS · AUTH |
-| supabase | CORE · AUTH · DATABASE · PAYMENTS · EVENTS · NOTIFICATIONS · TRPC |
-| vscode | CORE · DATABASE · AUTH · EVENTS · NOTIFICATIONS |
-| zed (Rust) | CORE · DATABASE · AUTH · EVENTS · PAYMENTS · NOTIFICATIONS · TRPC |
+| prisma | CORE · DATABASE · AUTH · EVENTS |
+| supabase | CORE · AUTH · DATABASE · PAYMENTS · NOTIFICATIONS · EVENTS · TRPC |
+| vscode | EXTENSIONS · AUTH · EVENTS · DATABASE · EXTENSION · CLI · CORE |
+| zed (Rust) | CORE · DATABASE · AUTH · EVENTS · PAYMENTS · TRPC · NOTIFICATIONS |
 
 vscode at 7,567 indexed files in under 8 seconds. Rust import graph working on zed (2,113 edges from `mod` declarations and `use crate::` paths).
 
+### Accuracy
+
+12 corpus repos pass `node test/accuracy-corpus.js --samples 100` — full parity between the bitmap path and the SQLite path on `blastRadius`, `crossDomain`, `highImpactFiles`, and `simulateChangeImpact`. The bitmap layer is a speedup, never a behavior change.
+
 ---
 
 ## How it works
 
 ```
-carto init / carto sync
+carto init
   ↓
 Discovers all files (no cap — SQLite handles the volume)
 mtime+size check → skip unchanged files
@@ -293,9 +394,22 @@ Leiden+CPM graph clustering → auto-detects domains
 Computes reverse deps → blast radius for every file
 Writes AGENTS.md + .carto/context/*.md (lazy, on-demand)
 Auto-wires MCP into every AI tool found
+Installs 4 git hooks: pre-commit, post-checkout, post-merge, post-rewrite
+  ↓
+[no daemon, no watcher, no background process]
+  ↓
+─── Freshness mechanism 1: git hooks (90% of cases) ───
+You commit / pull / switch branches / rebase
+  → hook runs `carto sync` quietly in <1s
+  → only changed files re-parsed
   ↓
-carto watch (optional)
+─── Freshness mechanism 2: lazy mtime check (the gap) ───
+You edit files between commits, AI asks "blast radius of db.ts?"
+  → MCP server stats db.ts (mtime+size vs DB row)
+  → if stale, re-parses just that file inline (<50ms)
+  → returns fresh answer
   ↓
+─── Optional: carto watch (AI-heavy workflows only) ───
 File saved → debounce 50ms → re-parse 1 file → SQLite write → <50ms
 ```
 
@@ -310,6 +424,16 @@ File saved → debounce 50ms → re-parse 1 file → SQLite write → <50ms
 
 ---
 
+## Origin
+
+I was building [Emfirge](https://www.emfirge.cloud) — a cloud security agent that maps AWS infrastructure into a graph and simulates the blast radius of every change.
+
+To make the AI inside Emfirge understand infrastructure, I wrote a module called `cartography.py`. It mapped AWS resources, built a graph of how they connected, and wrote it into a structured map. The AI stopped hallucinating. It worked with facts, not guesses.
+
+Carto is the same idea, applied to source code. Same insight: AI agents stop guessing once they can query the architecture.
+
+---
+
 ## License
 
 MIT. Free forever.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "carto-md",
-  "version": "2.0.7",
+  "version": "2.0.8",
   "description": "Structural intelligence layer for AI coding tools. Indexes your codebase into SQLite — routes, models, import graph, blast radius, domains — and exposes 16 MCP tools for Kiro, Cursor, and Claude.",
   "bin": {
     "carto": "src/cli/index.js"
@@ -10,7 +10,9 @@
     "test": "node test/test.js",
     "test:correctness": "node test/correctness.js",
     "test:benchmark": "node test/benchmark.js",
-    "test:bench-ci": "node test/bench-ci.js"
+    "test:bench-ci": "node test/bench-ci.js",
+    "postinstall": "node scripts/postinstall.js",
+    "bench:bitmap": "node bench/bitmap-validation/index.js"
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "^0.22.1",
@@ -18,7 +20,9 @@
     "@modelcontextprotocol/sdk": "^1.0.0",
     "better-sqlite3": "11.7.0",
     "chokidar": "^3.6.0",
-    "tree-sitter": "0.25.0",
+    "tree-sitter": "0.25.0"
+  },
+  "optionalDependencies": {
     "tree-sitter-c-sharp": "0.21.3",
     "tree-sitter-cpp": "0.23.4",
     "tree-sitter-go": "0.25.0",
@@ -29,7 +33,7 @@
     "tree-sitter-typescript": "0.23.2"
   },
   "engines": {
-    "node": ">=18.0.0 <26.0.0"
+    "node": ">=18.0.0"
   },
   "author": {
     "name": "Ansh Sonkar",

package/scripts/postinstall.js

@@ -0,0 +1,46 @@
+#!/usr/bin/env node
+'use strict';
+
+// Postinstall script — checks tree-sitter grammar availability and prints
+// actionable guidance if any failed to install. Always exits 0.
+
+if (process.env.CARTO_NO_POSTINSTALL === '1') process.exit(0);
+
+const GRAMMARS = [
+  { pkg: 'tree-sitter-javascript', langs: 'JavaScript' },
+  { pkg: 'tree-sitter-typescript', langs: 'TypeScript' },
+  { pkg: 'tree-sitter-python', langs: 'Python' },
+  { pkg: 'tree-sitter-go', langs: 'Go' },
+  { pkg: 'tree-sitter-rust', langs: 'Rust' },
+  { pkg: 'tree-sitter-java', langs: 'Java' },
+  { pkg: 'tree-sitter-cpp', langs: 'C/C++' },
+  { pkg: 'tree-sitter-c-sharp', langs: 'C#' },
+];
+
+const failed = [];
+for (const g of GRAMMARS) {
+  try { require(g.pkg); } catch { failed.push(g); }
+}
+
+if (failed.length === 0) process.exit(0);
+
+// Print guidance
+const langs = failed.map(g => g.langs).join(', ');
+const isWin = process.platform === 'win32';
+const isMac = process.platform === 'darwin';
+
+let fix;
+if (isWin) {
+  fix = 'Install "Desktop development with C++" from https://aka.ms/vs/17/release/vs_BuildTools.exe then re-run: npm rebuild';
+} else if (isMac) {
+  fix = 'Run: xcode-select --install && npm rebuild';
+} else {
+  fix = 'Run: sudo apt-get install -y build-essential && npm rebuild   (or equivalent for your distro)';
+}
+
+console.log('');
+console.log('[CARTO] ⚠️  Some tree-sitter grammars failed to install.');
+console.log(`[CARTO]    Affected languages: ${langs}`);
+console.log('[CARTO]    These languages will use regex-only extraction (less accurate).');
+console.log(`[CARTO]    To fix: ${fix}`);
+console.log('');

package/src/acp/agent.js

@@ -71,11 +71,11 @@ class CartoAgent {
 
   // Provider methods
   // The custom `unstable_*` provider-management methods (list / set / disable)
-  // were removed in 2.0.7 (Spec 7 Bug 3). The ACP SDK only dispatches the
-  // methods declared in its AGENT_METHODS constant, so those custom routes
-  // returned -32601 "Method not found" anyway. Provider config flows
-  // through env vars / editor settings; ProviderRegistry remains as the
-  // internal configuration carrier.
+  // were removed. The ACP SDK only dispatches the methods declared in
+  // its AGENT_METHODS constant, so those custom routes returned -32601
+  // "Method not found" anyway. Provider config flows through env vars /
+  // editor settings; ProviderRegistry remains as the internal
+  // configuration carrier.
 
   // Session list/load stubs
   async listSessions(_params) { return { sessions: [] }; }

package/src/acp/providers/index.js

@@ -30,8 +30,8 @@ class ProviderRegistry {
     this._config = null;
   }
 
-  // list() was removed in 2.0.7 (Spec 7 Bug 3). It was only called by the
-  // dropped custom provider-list ACP method, which the SDK never dispatched.
+  // list() was removed — it was only called by a dropped custom
+  // provider-list ACP method that the SDK never actually dispatched.
   // SUPPORTED_PROVIDERS is still exported for external callers / tests.
 
   /**

package/src/agents/leiden.js

@@ -300,24 +300,18 @@ function clusterByGraph(importGraph, gamma = 0.03, keywordSeeds = {}) {
 
   // Name each community
   const commNames = new Map();
-  const usedNames = new Map(); // name → count (for deduplication)
 
   for (const [commId, members] of communities) {
-    let name = nameCommunity(members, keywordSeeds);
-
-    // Deduplicate: if name already used, append a number
-    if (usedNames.has(name)) {
-      const count = usedNames.get(name) + 1;
-      usedNames.set(name, count);
-      name = `${name}_${count}`;
-    } else {
-      usedNames.set(name, 1);
-    }
-
+    const name = nameCommunity(members, keywordSeeds);
     commNames.set(commId, name);
   }
 
-  // Build final file → domain map
+  // Build final file → domain map.
+  // When multiple communities resolve to the same base name (e.g. with a
+  // dense import graph, several disjoint sub-systems can each rank highest
+  // for AUTH), merge them into a single domain. The keyword that names the
+  // domain is the user-visible truth; preserving "AUTH_2", "AUTH_3", ...
+  // suffixes leaks Leiden's internal community count into the UI.
   const result = new Map();
   for (const [node, commId] of communityMap) {
     result.set(node, commNames.get(commId) || 'CORE');

package/src/bitmap/bitset.js

@@ -0,0 +1,190 @@
+'use strict';
+
+/**
+ * Bitset — Uint32Array-backed dense bitmap.
+ *
+ * 60-line implementation,
+ * zero deps. Used by Carto's bitmap engine (`src/bitmap/`) for sub-millisecond
+ * forward/reverse adjacency queries on the import graph.
+ *
+ * Operations are word-level (32-bit lanes) — `or`, `and`, `andNot`, `popcount`,
+ * `iterate` all run at one cycle per 32 bits on modern CPUs. For Carto's typical
+ * 10K-file repos this is 312 words per bitmap → tens of nanoseconds for a
+ * popcount, microseconds for a 1-hop frontier expansion.
+ *
+ * Layout: bit `i` lives in `words[i >>> 5]` at lane `i & 31`. Out-of-range reads
+ * return false; out-of-range writes are caller's responsibility (the bitset
+ * does not auto-grow — it's sized at construction to the max file id + 1).
+ */
+class Bitset {
+  /**
+   * @param {number} size — Number of bits (i.e. max file id + 1).
+   */
+  constructor(size) {
+    this.size = size;
+    this.words = new Uint32Array(Math.ceil(size / 32));
+  }
+
+  /** Set bit `i` to 1. */
+  set(i) { this.words[i >>> 5] |= (1 << (i & 31)); }
+
+  /** Clear bit `i` (set to 0). */
+  clear(i) { this.words[i >>> 5] &= ~(1 << (i & 31)); }
+
+  /** Return true if bit `i` is set. */
+  has(i) { return (this.words[i >>> 5] & (1 << (i & 31))) !== 0; }
+
+  /** Bitwise OR. Returns a new Bitset sized to the larger of the two. */
+  or(other) {
+    const r = new Bitset(Math.max(this.size, other.size));
+    for (let i = 0; i < this.words.length; i++) r.words[i] = this.words[i];
+    for (let i = 0; i < other.words.length; i++) r.words[i] |= other.words[i];
+    return r;
+  }
+
+  /** Bitwise AND. Returns a new Bitset sized to `this`. */
+  and(other) {
+    const r = new Bitset(this.size);
+    const len = Math.min(this.words.length, other.words.length);
+    for (let i = 0; i < len; i++) r.words[i] = this.words[i] & other.words[i];
+    return r;
+  }
+
+  /** Bitwise AND-NOT (this & ~other). Returns a new Bitset sized to `this`. */
+  andNot(other) {
+    const r = new Bitset(this.size);
+    for (let i = 0; i < this.words.length; i++) {
+      r.words[i] = this.words[i] & ~(other.words[i] || 0);
+    }
+    return r;
+  }
+
+  /**
+   * In-place bitwise OR. Mutates `this` to `this | other` and returns `this`.
+   *
+   * Used by hot BFS loops (`blastRadius`, `simulateChangeImpact`)
+   * to avoid the per-hop Uint32Array allocation that `.or()` performs.
+   * If `other` is wider than `this`, words beyond `this.size` are dropped —
+   * callers that care about size growth should resize before calling.
+   */
+  orInPlace(other) {
+    const n = Math.min(this.words.length, other.words.length);
+    for (let i = 0; i < n; i++) this.words[i] |= other.words[i];
+    return this;
+  }
+
+  /**
+   * In-place bitwise AND-NOT. Mutates `this` to `this & ~other` and
+   * returns `this`.
+   */
+  andNotInPlace(other) {
+    const n = Math.min(this.words.length, other.words.length);
+    for (let i = 0; i < n; i++) this.words[i] &= ~other.words[i];
+    // If `other` is shorter than `this`, the trailing words stay as-is —
+    // andNot with implicit-zero is a no-op.
+    return this;
+  }
+
+  /**
+   * Copy bits from `other` into `this` (mutates and returns `this`).
+   *
+   * The receiver's word array is reused — no new allocation.
+   * If sizes differ, the overlapping prefix is copied and the suffix of
+   * `this` (if any) is zeroed so stale bits don't leak.
+   */
+  copyFrom(other) {
+    if (this.words.length === other.words.length) {
+      this.words.set(other.words);
+    } else {
+      const n = Math.min(this.words.length, other.words.length);
+      for (let i = 0; i < n; i++) this.words[i] = other.words[i];
+      for (let i = n; i < this.words.length; i++) this.words[i] = 0;
+    }
+    return this;
+  }
+
+  /**
+   * Set every word to `value` (defaults to 0). Mutates and returns `this`.
+   *
+   * The hot use case is `setAll(0)` to clear the transient
+   * frontier bitset between BFS hops without reallocating its 32-bit lanes.
+   */
+  setAll(value = 0) {
+    this.words.fill(value >>> 0);
+    return this;
+  }
+
+  /**
+   * Population count (number of set bits).
+   * Hamming weight via parallel bit-fold — same kernel as POPCNT but portable
+   * across Node runtimes that don't expose hardware popcount.
+   */
+  popcount() {
+    let c = 0;
+    for (let i = 0; i < this.words.length; i++) {
+      let v = this.words[i];
+      v = v - ((v >>> 1) & 0x55555555);
+      v = (v & 0x33333333) + ((v >>> 2) & 0x33333333);
+      c += (((v + (v >>> 4)) & 0x0F0F0F0F) * 0x01010101) >>> 24;
+    }
+    return c;
+  }
+
+  /**
+   * Returns an array of every set bit's index (file id), in ascending order.
+   * Uses the `v & -v` LSB trick to skip zero lanes early.
+   */
+  iterate() {
+    const result = [];
+    for (let w = 0; w < this.words.length; w++) {
+      let v = this.words[w];
+      while (v) {
+        const bit = v & (-v);
+        result.push((w << 5) + (31 - Math.clz32(bit)));
+        v ^= bit;
+      }
+    }
+    return result;
+  }
+
+  /** Deep copy. */
+  clone() {
+    const r = new Bitset(this.size);
+    r.words.set(this.words);
+    return r;
+  }
+
+  /**
+   * Serialize to a Buffer. Layout: just the raw word bytes (LE on x86/ARM
+   * Node hosts; Carto runs Node so endianness matches the running machine —
+   * the on-disk file is per-machine just like SQLite's `carto.db`).
+   *
+   * Size in bytes: `Math.ceil(size / 32) * 4`. Caller persists the `size`
+   * separately so deserialize knows how many bits to expect.
+   */
+  serialize() {
+    return Buffer.from(this.words.buffer, this.words.byteOffset, this.words.byteLength);
+  }
+
+  /**
+   * Reconstruct a Bitset from a Buffer + the original size.
+   *
+   * @param {Buffer} buf — Bytes produced by `serialize()`.
+   * @param {number} size — Bit count (must match the serialized bitset).
+   * @returns {Bitset}
+   */
+  static deserialize(buf, size) {
+    const r = new Bitset(size);
+    // Copy through a Uint8Array view so we don't alias buf's underlying
+    // ArrayBuffer (which may be shared with Node's internal pool).
+    const expected = r.words.byteLength;
+    if (buf.length < expected) {
+      throw new Error(`Bitset.deserialize: buffer too small (got ${buf.length}, need ${expected})`);
+    }
+    const u8 = new Uint8Array(r.words.buffer, r.words.byteOffset, expected);
+    for (let i = 0; i < expected; i++) u8[i] = buf[i];
+    return r;
+  }
+}
+
+module.exports = { Bitset };