fix: clear BOM state on fatal errors after stream end

Author: ChALkeR

fallback/encoding.js

@@ -110,17 +110,19 @@ export class TextDecoder {
             u = u.subarray(bom)
           }
         }
+      } else if (!stream && !this.ignoreBOM) {
+        this.#canBOM = true
       }
 
       if (!this.#decode) this.#decode = unicodeDecoder(this.encoding, !this.fatal)
       try {
         const res = (prefix ? this.#decode(prefix) : '') + this.#decode(u) + suffix
         if (res.length > 0 && stream) this.#canBOM = false
-
-        if (!stream) this.#canBOM = !this.ignoreBOM
         return res
       } catch (err) {
         this.#chunk = null // reset unfinished chunk on errors
+        // The correct way per spec seems to be not destroying the decoder state (aka BOM here) in stream mode
+        // See also multi-byte.js
         throw err
       }
 

tests/encoding/mistakes.test.js

@@ -715,6 +715,23 @@ describe('Common implementation mistakes', () => {
       }
     })
 
+    test('utf-8 BOM handling', (t) => {
+      // Firefox fails on this
+      {
+        const d = new TextDecoder('utf-8', { fatal: true })
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf), { stream: true }).length, 0) // BOM
+        t.assert.throws(() => d.decode(u(0xff)))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 0)
+      }
+
+      // Chrome and Node.js fail on this. Firefox, WebKit and Deno pass
+      {
+        const d = new TextDecoder('utf-8', { fatal: true })
+        t.assert.throws(() => d.decode(u(0xef, 0xbb, 0xbf, 0xff), { stream: true }))
+        t.assert.strictEqual(d.decode(u(0xef, 0xbb, 0xbf)).length, 1)
+      }
+    })
+
     // https://github.com/facebook/hermes/pull/1855#issuecomment-3632349129
     for (const encoding of ['utf-16le', 'utf-16be']) {
       test(encoding, (t) => {