fix(build): emit ASCII-only bundles so doctrenderer's no-ICU V8 can parse them

.github/workflows/check-build.yml

@@ -145,6 +145,22 @@ jobs:
           npm ci --prefix build
           npm run --prefix build build
 
+      # doctrenderer/x2t execute these bundles on a V8 built with
+      # v8_enable_i18n_support=false, whose reduced Unicode tables reject
+      # supplementary-plane characters used as identifiers. Terser emitted the
+      # LaTeX symbol table's astral keys bare, and that V8 then failed the whole
+      # bundle with "SyntaxError: Invalid or unexpected token", aborting x2t on
+      # first start (issue #80). Every JS gate in this workflow runs on Node and
+      # headless Chromium -- both full-ICU -- so none of them can catch it.
+      #
+      # The check is a script rather than a grep because the residual failure
+      # mode is pure ASCII: with format.ascii_only but no format.quote_keys,
+      # Terser emits `\u{1d552}:"\\doublea"`, which no byte-range test flags.
+      - name: Assert built bundles are ASCII-only with no bare astral keys (#80)
+        run: |
+          cd sdkjs
+          node build/scripts/check-bundle-ascii.cjs
+
       - name: Install QUnit runner dependencies
         run: |
           sudo apt-get update

build/scripts/check-bundle-ascii.cjs

@@ -0,0 +1,103 @@
+/**
+ * SPDX-FileCopyrightText: 2026 Euro-Office contributors
+ * SPDX-License-Identifier: AGPL-3.0-only
+ */
+
+/**
+ * Guards the two invariants the built sdkjs bundles must satisfy for
+ * doctrenderer/x2t, which run them on a V8 built with
+ * v8_enable_i18n_support=false (core: Common/3dParty/v8/tools/8.9/*).
+ *
+ *   1. No raw non-ASCII byte.
+ *   2. No object key that is a bare supplementary-plane ("astral") identifier.
+ *
+ * (2) is the subtle one and the reason a plain "is it ASCII" grep is not
+ * enough: with format.ascii_only but without format.quote_keys, Terser emits
+ *
+ *     \u{1d552}:"\\doublea"
+ *
+ * which contains no byte > 0x7F yet still parses as a bare identifier whose
+ * code point is U+1D552. A no-ICU V8 rejects that exactly as it rejects the
+ * raw UTF-8 spelling, so invariant (1) alone would let issue #80 regress
+ * silently. See build/webpack.sdk.factory.mjs.
+ *
+ * Usage: node build/scripts/check-bundle-ascii.cjs [file...]
+ * With no arguments, scans the built sdk-all bundles under the deploy root.
+ */
+
+'use strict';
+
+const fs   = require('fs');
+const path = require('path');
+const { sync: globSync } = require('glob');
+const { resolveBuildRoot } = require('../lib/env.cjs');
+
+// An unquoted object key in `{`/`,` position that starts with either a raw
+// astral character or a unicode escape. A quoted key cannot match: the quote
+// sits between the delimiter and the escape.
+const BARE_ASTRAL_RAW = /[{,]\s*[\u{10000}-\u{10FFFF}]/u;
+const BARE_ASTRAL_ESC = /[{,]\s*\\u(?:\{[0-9a-fA-F]{5,6}\}|[dD][89abAB][0-9a-fA-F]{2})/;
+
+function sampleAround(text, regex, limit = 3) {
+    const global = new RegExp(regex.source, regex.flags.includes('g') ? regex.flags : regex.flags + 'g');
+    const out = [];
+    let m;
+    while ((m = global.exec(text)) !== null && out.length < limit) {
+        out.push(text.slice(m.index, Math.min(m.index + 40, text.length)));
+    }
+    return out;
+}
+
+function checkFile(file) {
+    const buf  = fs.readFileSync(file);
+    const problems = [];
+
+    let nonAscii = 0;
+    for (const byte of buf) if (byte > 0x7f) nonAscii++;
+    if (nonAscii > 0) {
+        problems.push(`${nonAscii} raw non-ASCII byte(s); Terser format.ascii_only must stay enabled`);
+    }
+
+    const text = buf.toString('utf8');
+    for (const [label, re] of [['raw', BARE_ASTRAL_RAW], ['escaped', BARE_ASTRAL_ESC]]) {
+        if (re.test(text)) {
+            const samples = sampleAround(text, re).map((s) => JSON.stringify(s)).join(', ');
+            problems.push(`bare astral object key (${label} form); Terser format.quote_keys must stay enabled -- e.g. ${samples}`);
+        }
+    }
+
+    return problems;
+}
+
+function main(argv) {
+    let files = argv.slice(2);
+    if (files.length === 0) {
+        const buildDir  = path.resolve(__dirname, '..');
+        const deployRoot = resolveBuildRoot(buildDir);
+        files = globSync('*/sdk-all{,-min}.js', { cwd: deployRoot, absolute: true }).sort();
+        if (files.length === 0) {
+            console.error(`check-bundle-ascii: no sdk-all bundles found under ${deployRoot} -- nothing was verified`);
+            return 1;
+        }
+    }
+
+    let failed = 0;
+    for (const file of files) {
+        const problems = checkFile(file);
+        if (problems.length === 0) {
+            console.log(`ok: ${file}`);
+            continue;
+        }
+        failed++;
+        for (const problem of problems) {
+            console.error(`::error file=${file}::${problem} (see issue #80)`);
+        }
+    }
+
+    console.log(`check-bundle-ascii: ${files.length} bundle(s) scanned, ${failed} failed`);
+    return failed === 0 ? 0 : 1;
+}
+
+if (require.main === module) process.exit(main(process.argv));
+
+module.exports = { checkFile, BARE_ASTRAL_RAW, BARE_ASTRAL_ESC };

build/scripts/deploy-assets.cjs

@@ -89,7 +89,20 @@ async function deployJsFile(srcPath, destPath) {
     const result  = await minify(source, {
         compress: false,
         mangle:   false,
-        format:   { comments: false },
+        format:   {
+            comments: false,
+            // Same invariant as the sdk-all bundles in webpack.sdk.factory.mjs:
+            // several of the files deployed here (Native/*.js, and the
+            // libfont/engine/fonts_*.js that doctrenderer concatenates into the
+            // script it compiles) are executed by a V8 built with
+            // v8_enable_i18n_support=false, which rejects supplementary-plane
+            // characters used as identifiers. These inputs happen to be
+            // ASCII-clean today, so this is prophylactic rather than a fix --
+            // but nothing else stops the next non-ASCII string landing in one of
+            // them and reproducing issue #80 outside the bundle scan's reach.
+            ascii_only: true,
+            quote_keys: true,
+        },
     });
     const content = licenseText + '\n' + (result.code != null ? result.code : source);
     fs.mkdirSync(path.dirname(destPath), { recursive: true });

build/test/webpack-sdk-terser-options.test.cjs

@@ -21,6 +21,10 @@ const assert = require('node:assert/strict');
 const path   = require('node:path');
 const url    = require('node:url');
 
+// Share one definition of "bare astral key" with the CI bundle scan so the two
+// cannot drift apart.
+const { BARE_ASTRAL_RAW, BARE_ASTRAL_ESC } = require('../scripts/check-bundle-ascii.cjs');
+
 async function loadSdkConfig() {
     const mod = await import(url.pathToFileURL(path.join(__dirname, '..', 'webpack.sdk.factory.mjs')));
     return mod.sdkConfig;
@@ -104,3 +108,68 @@ test('sdkConfig: DROP_CONSOLE is opt-in, not the default, for non-desktop/mobile
         }
     }
 });
+
+// --- Regression coverage for issue #80 ---------------------------------------
+// doctrenderer/x2t execute these bundles on a V8 built with
+// v8_enable_i18n_support=false, whose reduced Unicode tables do not accept
+// supplementary-plane ("astral") characters as identifiers. The webpack
+// pipeline emitted the LaTeX symbol table's astral keys as bare identifiers
+// (the reverse map in word/Math/NamesOfLiterals.js), and that V8 rejected the
+// whole bundle with "SyntaxError: Invalid or unexpected token" -- aborting x2t
+// on first start and making Euro-Office 9.3.4 unusable. ASCII-only output is
+// the invariant the previous Closure Compiler build provided implicitly.
+
+for (const platform of ['', 'desktop', 'mobile']) {
+    test(`sdkConfig: Terser is configured for ASCII-only output on platform '${platform || 'web'}' (#80)`, async () => {
+        const sdkConfig = await loadSdkConfig();
+        const terserOptions = withPlatform(platform, () => terserOptionsOf(sdkConfig('word')));
+
+        assert.equal(terserOptions.format.ascii_only, true);
+        assert.equal(terserOptions.format.quote_keys, true);
+    });
+}
+
+// terser-webpack-plugin derives an `ecma` from webpack's target and injects it
+// into terserOptions; Terser's own default (ecma unset) is conservative and
+// quotes astral keys regardless. Calling terser.minify() without ecma therefore
+// exercises a code path the real build never takes -- and would pass even with
+// quote_keys deleted. Pin it so these assertions test the shipped behaviour.
+const PIPELINE_ECMA = 2020;
+
+// Shape lifted from word/Math/NamesOfLiterals.js's reverse symbol table:
+// U+2219 (BMP) plus U+1D552 / U+1D538 (astral) used as object keys.
+const SYMBOL_TABLE_SRC = 'var Reverse={"\u2219":"\\bullet","\u{1d552}":"\\doublea","\u{1d538}":"\\doubleA"};';
+
+test('sdkConfig: Terser escapes astral object keys rather than emitting them bare (#80)', async () => {
+    const terser        = require('terser');
+    const sdkConfig     = await loadSdkConfig();
+    const terserOptions = withPlatform('', () => terserOptionsOf(sdkConfig('word')));
+
+    const { code } = await terser.minify(SYMBOL_TABLE_SRC, { ...terserOptions, ecma: PIPELINE_ECMA });
+
+    assert.ok(!/[^\x00-\x7F]/.test(code),
+        `Terser emitted non-ASCII output, which doctrenderer's no-ICU V8 rejects: ${code}`);
+    assert.equal(BARE_ASTRAL_RAW.test(code), false,
+        `Terser emitted a raw bare astral object key: ${code}`);
+    assert.equal(BARE_ASTRAL_ESC.test(code), false,
+        `Terser emitted an escaped bare astral object key, which is still a bare astral identifier to a no-ICU V8: ${code}`);
+});
+
+// Guards the guard: without quote_keys the same input must produce exactly the
+// failure this PR fixes. If this ever stops failing, the assertions above have
+// gone blind and the ones in the test before it prove nothing.
+test('sdkConfig: dropping quote_keys reintroduces the bare astral key (#80)', async () => {
+    const terser        = require('terser');
+    const sdkConfig     = await loadSdkConfig();
+    const terserOptions = withPlatform('', () => terserOptionsOf(sdkConfig('word')));
+
+    const withoutQuoteKeys = {
+        ...terserOptions,
+        ecma:   PIPELINE_ECMA,
+        format: { ...terserOptions.format, quote_keys: false },
+    };
+    const { code } = await terser.minify(SYMBOL_TABLE_SRC, withoutQuoteKeys);
+
+    assert.equal(BARE_ASTRAL_ESC.test(code), true,
+        `Expected a bare escaped astral key without quote_keys, got: ${code}`);
+});

build/webpack.sdk.factory.mjs

@@ -296,6 +296,38 @@ export function sdkConfig(moduleName) {
                                 // text also matches the identical per-file header repeated in
                                 // all ~400+ concatenated source files.
                                 comments: /@@license-banner@@/,
+
+                                // BOTH of the options below are required; neither is
+                                // sufficient alone. doctrenderer/x2t run these bundles on a
+                                // V8 built with v8_enable_i18n_support=false, whose reduced
+                                // Unicode tables do not classify supplementary-plane
+                                // ("astral") characters as ID_Start. The LaTeX symbol table
+                                // in word/Math/NamesOfLiterals.js keys an object on such
+                                // characters, and this pipeline emits them unquoted.
+                                //
+                                // ascii_only escapes non-ASCII *characters*, but an escaped
+                                // bare key is still a bare astral identifier: with
+                                // ascii_only alone this build emits 131 keys of the form
+                                //     \u{1d552}:"\\doublea"
+                                // which is pure ASCII text yet still resolves to U+1D552 for
+                                // ID_Start classification, so that V8 rejects it exactly as
+                                // it rejects the raw UTF-8 form. Measured: with ascii_only
+                                // alone, `x2t -create-js-cache` still aborts and writes a
+                                // 0-byte sdk-all.cache.
+                                //
+                                // quote_keys is therefore the load-bearing option -- it
+                                // turns the key into a string literal, which ascii_only then
+                                // escapes into `"\u{1d552}"`. ascii_only remains necessary
+                                // in its own right so no raw non-ASCII byte reaches the
+                                // bundle at all.
+                                //
+                                // quote_keys must live inside `format`; Terser rejects a
+                                // top-level one with "`quote_keys` is not a supported
+                                // option". Covered by
+                                // build/test/webpack-sdk-terser-options.test.cjs and by the
+                                // bundle scan in check-build.yml. See #80.
+                                ascii_only: true,
+                                quote_keys: true,
                             },
                             compress: (platform === 'desktop' || platform === 'mobile')
                                 // Old build-desktop.bat/build-mobile.command ran Closure's