feat(mcp): OOXML reference engine

.gitignore

@@ -5,4 +5,10 @@ dev/
 .wrangler/
 .env
 .mcp.json
-.vscode/
+.vscode/
+
+# Local-only planning doc (public repo)
+PLAN.md
+
+# XSD/spec artifacts: pulled by scripts/fetch-xsd.ts; never committed.
+data/xsd-cache/

CLAUDE.md

@@ -29,15 +29,22 @@ apps/
     src/data/docs.ts   ← All doc pages live here (single source of truth)
     src/components/    UI components (Sidebar, SuperDocPreview, etc.)
     src/pages/         Route pages (Home, Docs, SpecExplorer, Mcp)
-  mcp-server/          Cloudflare Worker — MCP server for AI spec search
+  mcp-server/          Cloudflare Worker - MCP server (semantic + structural tools)
 packages/
   shared/              Database client, embedding client, types
 scripts/
-  ingest/              PDF → chunks → embeddings → database pipeline
+  ingest-pdf/          ECMA PDF -> spec_content (semantic search corpus)
+  ingest-xsd/          ECMA XSDs -> schema graph (structural query corpus)
+  sources-sync.ts      data/sources.json -> reference_sources
+  db-migrate.ts        Apply db/migrations/*.sql in order
 db/
-  schema.sql           PostgreSQL + pgvector schema
+  schema.sql           PostgreSQL + pgvector + XSD schema graph
+  migrations/          Numbered, idempotent SQL migrations
+data/
+  sources.json         Source manifest (artifact URLs, sha256, license notes)
+  xsd-cache/           Local-only XSD download cache (gitignored)
 dev/
-  data/                Extracted/chunked/embedded spec content
+  data/                Extracted/chunked/embedded PDF content
 ```
 
 ## Commands
@@ -97,23 +104,52 @@ The XML you provide is wrapped in a minimal `w:document > w:body` structure auto
 
 ## MCP Server
 
-Cloudflare Worker exposing three MCP tools for semantic spec search:
+Cloudflare Worker exposing two flavors of MCP tools backed by the same database.
 
-- `search_ecma_spec` — semantic vector search across 18,000+ spec chunks
-- `get_section` — fetch a specific section by ID (e.g., "17.3.1.24")
-- `list_parts` — browse the spec structure
+Semantic search over the spec PDF (powered by `spec_content`):
+
+- `search_ecma_spec` - semantic vector search across 18,000+ spec chunks
+- `get_section` - fetch a specific section by ID (e.g., "17.3.1.24")
+- `list_parts` - browse the spec structure
+
+Structural queries over the XSD schema graph (powered by `xsd_*` tables):
+
+- `ooxml_lookup_element` / `ooxml_lookup_type` - canonical symbol info
+- `ooxml_children` - legal children of an element/type/group, in document order
+- `ooxml_attributes` - attributes including those inherited and unfolded from attributeGroup refs
+- `ooxml_enum` - simpleType enumeration values
+- `ooxml_namespace_info` - vocabularies and per-profile symbol counts for a namespace URI
 
 Uses PostgreSQL with pgvector (Neon serverless in production, Docker locally).
 
-## Data Pipeline
+## Data Pipelines
+
+Two ingest paths feed the same database. Both are reproducible from `data/sources.json`.
 
-Ingests ECMA-376 PDFs into the vector database:
+**PDF (semantic corpus, into `spec_content`)**:
 
 ```
 PDF → extract (Python) → chunk (6KB) → embed (Voyage) → upload (PostgreSQL)
 ```
 
-Run the full pipeline: `bun scripts/ingest/pipeline.ts`
+```bash
+bun run pdf:ingest 1 ./pdfs/ECMA-376-Part1.pdf   # full pipeline for one part
+```
+
+See `scripts/ingest-pdf/README.md`.
+
+**XSD (structural corpus, into `xsd_*` tables)**:
+
+```
+ECMA Part 4 zip → fetch+verify (sha256) → parse → ingest (single transaction)
+```
+
+```bash
+bun run xsd:fetch                                    # URL + sha256 from data/sources.json
+bun run xsd:ingest
+```
+
+See `scripts/ingest-xsd/README.md`.
 
 ## Database
 

README.md

@@ -10,9 +10,10 @@ The OOXML spec, explained by people who actually implemented it.
 
 An interactive reference for ECMA-376 (Office Open XML) built by the [SuperDoc — DOCX editing and tooling](https://superdoc.dev) team. Every page combines XML structure, live rendered previews, and implementation notes that tell you what the spec doesn't.
 
-- **Live previews** — Edit XML and see it render in real-time. Every example is a working document.
-- **Implementation notes** — Where Word diverges from the spec, what will break your code, and what to do about it.
-- **Semantic spec search** — 18,000+ spec chunks searchable by meaning via MCP server.
+- **Live previews** - Edit XML and see it render in real-time. Every example is a working document.
+- **Implementation notes** - Where Word diverges from the spec, what will break your code, and what to do about it.
+- **Semantic spec search** - 18,000+ spec chunks searchable by meaning via MCP server.
+- **Structural schema lookup** - Element children, attributes, types, enums, namespaces. Same MCP server, deterministic answers from the parsed XSDs.
 
 ## Why?
 
@@ -22,13 +23,16 @@ We faced this at SuperDoc — building a document engine on native OOXML with no
 
 ## MCP Server
 
-Search the ECMA-376 spec with AI. Ask questions in natural language, get answers grounded in the actual specification.
+Ask questions in natural language and get answers grounded in the spec, or query the schema graph for precise structural answers.
 
 ```bash
 claude mcp add --transport http ecma-spec https://api.ooxml.dev/mcp
 ```
 
-Works with Claude Code, Cursor, and any MCP-compatible client. Three tools: `search_ecma_spec` (semantic search), `get_section` (by ID), and `list_parts` (browse structure).
+Works with Claude Code, Cursor, and any MCP-compatible client. Two flavors of tools share one server:
+
+- **Semantic** (over the spec PDF): `search_ecma_spec`, `get_section`, `list_parts`
+- **Structural** (over the parsed XSDs): `ooxml_lookup_element`, `ooxml_lookup_type`, `ooxml_children`, `ooxml_attributes`, `ooxml_enum`, `ooxml_namespace_info`
 
 ## Development
 

apps/mcp-server/src/mcp.ts

@@ -7,6 +7,7 @@
 import { createDb } from "./db";
 import { embedQuery } from "./embeddings";
 import type { Env } from "./index";
+import { callOoxmlTool, isOoxmlTool, OOXML_TOOL_DEFS } from "./ooxml-tools";
 
 // JSON-RPC types
 interface JsonRpcRequest {
@@ -136,9 +137,7 @@ function handleToolsList(id: number | string | null): JsonRpcResponse {
 	return {
 		jsonrpc: "2.0",
 		id,
-		result: {
-			tools: TOOLS,
-		},
+		result: { tools: [...TOOLS, ...OOXML_TOOL_DEFS] },
 	};
 }
 
@@ -162,6 +161,17 @@ async function handleToolsCall(
 	try {
 		let resultText: string;
 
+		// Structural OOXML tools share the dispatch with the existing semantic
+		// tools below.
+		if (isOoxmlTool(name)) {
+			resultText = await callOoxmlTool(name, args ?? {}, env);
+			return {
+				jsonrpc: "2.0",
+				id,
+				result: { content: [{ type: "text", text: resultText }] },
+			};
+		}
+
 		switch (name) {
 			case "search_ecma_spec": {
 				const query = args?.query as string;