fix(coding-agent): bundle mupdf into the compiled binary so PDF reads work

The vendored markit engine kept `mupdf` external, but a single-file
`bun --compile` binary has no node_modules to resolve it from, so the
standalone binary aborted at startup with `Cannot find package 'mupdf'`
— the otherwise-lazy import is resolved eagerly at boot. Bundle mupdf and
embed its WASM blob (scripts/embed-mupdf-wasm.ts, reset after the build);
npm and source installs still load mupdf from node_modules.

Import mupdf lazily inside the PDF converter so the bundled markit chunk's
init stays synchronous: mupdf's top-level await otherwise made the chunk
init async and bun's compiled bundler failed to await it through the
barrel, exposing the converters before their module-level const tables
initialized (undefined EXTENSIONS). Also keeps the ~10MB wasm off non-PDF
document conversions.

packages/coding-agent/scripts/build-binary.ts

@@ -44,6 +44,7 @@ async function main(): Promise<void> {
 		await runCommand(["bun", "--cwd=../stats", "scripts/generate-client-bundle.ts", "--generate"]);
 		await runCommand(["bun", "scripts/generate-docs-index.ts", "--generate"]);
 		await runCommand(["bun", "--cwd=../natives", "run", "embed:native"]);
+		await runCommand(["bun", "scripts/embed-mupdf-wasm.ts", "--generate"]);
 		try {
 			const buildEnv = shouldAdhocSignDarwinBinary() ? { ...Bun.env, BUN_NO_CODESIGN_MACHO_BINARY: "1" } : Bun.env;
 			await runCommand(
@@ -61,8 +62,6 @@ async function main(): Promise<void> {
 					"--define",
 					`process.env.PI_TINY_TRANSFORMERS_VERSION=${JSON.stringify(transformersVersion)}`,
 					"--external",
-					"mupdf",
-					"--external",
 					"fastembed",
 					"--external",
 					"onnxruntime-node",
@@ -97,6 +96,7 @@ async function main(): Promise<void> {
 				await runCommand(["codesign", "--force", "--sign", "-", outputPath]);
 			}
 		} finally {
+			await runCommand(["bun", "scripts/embed-mupdf-wasm.ts", "--reset"]);
 			await runCommand(["bun", "--cwd=../natives", "run", "embed:native", "--reset"]);
 		}
 	} finally {

packages/coding-agent/scripts/embed-mupdf-wasm.ts

@@ -0,0 +1,67 @@
+#!/usr/bin/env bun
+
+// Embeds mupdf's `mupdf-wasm.wasm` into the compiled single-file binary.
+//
+// mupdf loads its wasm by reading the `mupdf-wasm.wasm` sibling of its own
+// module via `new URL(..., import.meta.url)` + `readFileSync`. A `bun --compile`
+// binary has no node_modules, so that read fails (`ENOENT .../mupdf-wasm.wasm`),
+// and marking mupdf `--external` instead makes `bun --compile` eagerly fail to
+// resolve the package at startup (the static `import * as mupdf` lives in a lazy
+// chunk but is hoisted). So the binary build bundles mupdf and embeds the wasm
+// bytes here, handing them to the WASM module as `$libmupdf_wasm_Module.wasmBinary`
+// (see src/utils/markit.ts).
+//
+// `--generate` copies the wasm next to src/utils/mupdf-wasm-embed.ts and rewrites
+// that module to import it via `with { type: "file" }`; `--reset` restores the
+// checked-in placeholder and removes the copy. The npm `dist/cli.js` bundle never
+// runs this — it keeps mupdf external and loads the wasm from node_modules.
+
+import * as fs from "node:fs/promises";
+import { createRequire } from "node:module";
+import * as path from "node:path";
+
+const utilsDir = path.join(import.meta.dir, "..", "src", "utils");
+const helperPath = path.join(utilsDir, "mupdf-wasm-embed.ts");
+const wasmCopyPath = path.join(utilsDir, "mupdf-wasm.wasm");
+
+const placeholder = `// AUTOGENERATED -- managed by scripts/embed-mupdf-wasm.ts. Do not edit by hand.
+//
+// Compiled single-file binaries cannot let mupdf resolve its \`mupdf-wasm.wasm\`
+// sibling from the read-only bunfs, so the binary build (scripts/build-binary.ts
+// and scripts/ci-release-build-binaries.ts) regenerates this module to embed the
+// wasm bytes via \`with { type: "file" }\` and copies the wasm next to it. Source
+// checkouts, \`bun test\`, and the npm \`dist/cli.js\` bundle keep mupdf external and
+// load the wasm from node_modules, so this placeholder returns undefined and the
+// build resets back to it afterward.
+export function loadEmbeddedMupdfWasm(): Uint8Array | undefined {
+\treturn undefined;
+}
+`;
+
+const generated = `// AUTOGENERATED -- managed by scripts/embed-mupdf-wasm.ts. Do not edit or commit.
+import { readFileSync } from "node:fs";
+import wasmPath from "./mupdf-wasm.wasm" with { type: "file" };
+
+export function loadEmbeddedMupdfWasm(): Uint8Array | undefined {
+\treturn readFileSync(wasmPath);
+}
+`;
+
+if (process.argv.includes("--reset")) {
+	await Bun.write(helperPath, placeholder);
+	try {
+		await fs.unlink(wasmCopyPath);
+	} catch (err) {
+		if ((err as NodeJS.ErrnoException).code !== "ENOENT") throw err;
+	}
+	process.exit(0);
+}
+
+const wasmSource = path.join(path.dirname(createRequire(import.meta.url).resolve("mupdf")), "mupdf-wasm.wasm");
+const wasmFile = Bun.file(wasmSource);
+if (!(await wasmFile.exists())) {
+	throw new Error(`mupdf wasm not found at ${wasmSource}; run \`bun install\` first.`);
+}
+await Bun.write(wasmCopyPath, wasmFile);
+await Bun.write(helperPath, generated);
+console.log(`Embedded mupdf wasm (${wasmFile.size} bytes) into ${path.relative(process.cwd(), wasmCopyPath)}`);