Fix chrF fallback so English wiki garbage fails translation_success

Sacrebleu errors used to fall back to set-overlap unigrams, so any
English hyp scored ~70–80 against English refs and SFT reported
success 1.0. Use count-based char n-grams, keep BLEU failures from
clobbering chrF, and print a few hyps during eval.

kda/training/eval_mt.py

@@ -67,14 +67,18 @@ def evaluate_pairs(
         want = "zh" if target_lang.startswith("zh") else "en"
         lang_ok += int(_detect_lang(hyp) == want)
         chrf_sum += _chrf(hyp, ref)
-    corpus = {}
+    corpus = {"chrf": chrf_sum / max(n, 1), "bleu": None}
     try:
-        from sacrebleu.metrics import BLEU, CHRF
+        from sacrebleu.metrics import CHRF
 
         corpus["chrf"] = float(CHRF(word_order=2).corpus_score(hyps, [refs[:n]]).score)
+    except Exception:
+        pass
+    try:
+        from sacrebleu.metrics import BLEU
+
         corpus["bleu"] = float(BLEU().corpus_score(hyps, [refs[:n]]).score)
     except Exception:
-        corpus["chrf"] = chrf_sum / max(n, 1)
         corpus["bleu"] = None
     return {
         "n": n,
@@ -131,6 +135,8 @@ def main() -> None:
         )
         printable = {k: v for k, v in out.items() if k != "hyps"}
         print(json.dumps(printable, ensure_ascii=False, indent=2))
+        for i, hyp in enumerate(out["hyps"][: min(5, out["n"])]):
+            print(f"  [{i}] {hyp}")
     elif not args.prefix:
         raise SystemExit("pass --prefix and/or --src + --ref")