fix(cli): preserve literal WebVTT cue text (#5236)

packages/cli/src/whisper/normalize.test.ts

@@ -281,6 +281,33 @@ Short format
 });
 
 describe("caption formatting", () => {
+  it.each([
+    ["R&D <config> next", "R&amp;D &lt;config&gt; next"],
+    ["Literal &lt; &amp; references", "Literal &amp;lt; &amp;amp; references"],
+    ["3 < 4 and 5 > 2", "3 &lt; 4 and 5 &gt; 2"],
+    ["R&D --> next", "R&amp;D --&gt; next"],
+  ])("preserves literal WebVTT text %j on export and reimport", (text, payload) => {
+    const words = [{ text, start: 1, end: 2, id: "w0" }];
+    const output = formatVtt(words);
+    expect(output).toBe(`WEBVTT\n\n00:00:01.000 --> 00:00:02.000\n${payload}\n`);
+    expect(loadTranscript(tmpFile("literal.vtt", output)).words).toEqual(words);
+  });
+
+  it.each([
+    ["R&amp;D &lt;config&gt; next", "R&D <config> next"],
+    ["&quot;quoted&quot; &apos;text&apos;", "\"quoted\" 'text'"],
+    ["&#38; &#x3c; &#60;", "& < <"],
+    ["Literal &amp;lt; &amp;amp; references", "Literal &lt; &amp; references"],
+    ["<b>R&amp;D</b> &lt;config&gt;", "R&D <config>"],
+    ["A&nbsp;B&lrm;&rlm;", "A\u00a0B\u200e\u200f"],
+    ["Keep &unknown;", "Keep &unknown;"],
+  ])("decodes the WebVTT cue payload %j once", (payload, text) => {
+    const input = `WEBVTT\n\n00:00:01.000 --> 00:00:02.000\n${payload}\n`;
+    expect(loadTranscript(tmpFile("entities.vtt", input)).words).toEqual([
+      { text, start: 1, end: 2, id: "w0" },
+    ]);
+  });
+
   it("round-trips SRT cues through normalized words", () => {
     const srt = `1
 00:00:01,000 --> 00:00:03,500

packages/cli/src/whisper/normalize.ts

@@ -1,5 +1,6 @@
 import { readFileSync, readdirSync, writeFileSync } from "node:fs";
 import { extname, join } from "node:path";
+import { parseHTML } from "linkedom";
 
 export interface Word {
   /** Stable identifier for referencing this word in overrides and compositions.
@@ -217,6 +218,14 @@ function parseSrt(content: string): Word[] {
   return words;
 }
 
+function decodeVttText(text: string): string {
+  if (!text.includes("&")) return text;
+  const { document } = parseHTML("<html><body></body></html>");
+  const element = document.createElement("span");
+  element.innerHTML = text.replace(/</g, "&lt;");
+  return element.textContent ?? "";
+}
+
 function parseVtt(content: string): Word[] {
   const body = content.replace(/^WEBVTT[^\n]*\n/, "");
   // VTT is structurally similar to SRT (without numeric indices)
@@ -239,7 +248,7 @@ function parseVtt(content: string): Word[] {
     if (!text) continue;
 
     words.push({
-      text,
+      text: decodeVttText(text),
       start: parseVttTimestamp(startStr),
       end: parseVttTimestamp(endStr),
     });
@@ -468,7 +477,8 @@ export function formatVtt(words: Word[], opts?: WordsToCuesOptions): string {
     "WEBVTT\n\n" +
     cues
       .map(
-        (cue) => `${formatVttTimestamp(cue.start)} --> ${formatVttTimestamp(cue.end)}\n${cue.text}`,
+        (cue) =>
+          `${formatVttTimestamp(cue.start)} --> ${formatVttTimestamp(cue.end)}\n${cue.text.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;")}`,
       )
       .join("\n\n") +
     "\n"