fix: seq2seq decode returns primary choice (strip alternates)

src/ml/imf/model.ts

@@ -11,7 +11,7 @@ import type { Tensor } from "../types.js"
 import { verifyAndRead, parseManifest, type IMFManifest } from "./loader.js"
 import { resolve } from "./registry.js"
 import { normalizeArabicInput, repetitionGuardCut } from "./guards.js"
-import { EOS_ID, PAD_ID, decode, encode } from "./tokens.js"
+import { EOS_ID, PAD_ID, decode, encode, firstAlternates } from "./tokens.js"
 import { translateWindowed } from "./windows.js"
 
 interface InputMeta {
@@ -102,7 +102,7 @@ export class IMFModel {
     const tokens = this.kv
       ? await this.greedyKv(hidden, maxLen, opts)
       : await this.greedyPlain(hidden, maxLen, opts)
-    return decode(tokens)
+    return firstAlternates(decode(tokens))
   }
 
   /** Normalized text -> encoder hidden states; null when the input

src/ml/imf/tokens.ts

@@ -30,3 +30,14 @@ export function decode(tokenIds: readonly number[]): string {
   }
   return decoder.decode(new Uint8Array(out))
 }
+
+/** Seq2seq output contract: models trained on multi-reference corpora
+ * emit '/'-separated alternates per token; the runtime returns the
+ * primary choice only. */
+export function firstAlternates(text: string): string {
+  return text
+    .split(/\s+/)
+    .filter((t) => t.length > 0)
+    .map((t) => t.split("/")[0])
+    .join(" ")
+}

test/ml/first-alternates.test.ts

@@ -0,0 +1,18 @@
+import { describe, expect, it } from "vitest"
+
+import { firstAlternates } from "../../src/ml/imf/tokens.js"
+
+describe("firstAlternates", () => {
+  it("keeps the primary choice per token", () => {
+    expect(firstAlternates("وِكَالَةُ/وِكَالَةُ اَلْأَرْصَادِ")).toBe("وِكَالَةُ اَلْأَرْصَادِ")
+  })
+
+  it("leaves plain text untouched", () => {
+    expect(firstAlternates("مرحبا بالعالم")).toBe("مرحبا بالعالم")
+  })
+
+  it("handles edges", () => {
+    expect(firstAlternates("")).toBe("")
+    expect(firstAlternates("a/b c/d")).toBe("a c")
+  })
+})