exp77: order-13 flat Dirichlet + phrase[36,28,20,16]

Extend n-gram to order-13 (PR openai#921 validates higher orders: 0.0939).
Trim phrase to [36,28,20,16] to fit eval budget.
Flat Dirichlet c=1.0 (highest match only — avoids hierarchical overhead).

Author: sofiabod

results.tsv

@@ -47,3 +47,4 @@ b224b23	1.1323	15.88	keep	TTT AdamW 5ep lr=0.0005 DDP-synced	BEATS SOTA! sw=1.13
 1a8ee89	0.2534	15.26	discard	hierarchical Dirichlet c=0.5 order-9	slightly worse than c=1.0, eval=532s
 cd10ecb	0.2463	15.39	keep	flat Dirichlet c=1.0 + phrase[28,20,16]	NEW BEST! phrase[28] adds -0.007, eval=529s
 fc5f627	0.2417	15.39	keep	flat Dirichlet c=1.0 + phrase[36,28,20,16]	NEW BEST! phrase[36] adds -0.005, eval=548s
+1b32847	0.2380	15.65	keep	flat Dirichlet c=1.0 + phrase[48,36,28,20,16]	NEW BEST! -0.004, eval=586s (14s spare)

train_gpt.py

@@ -896,7 +896,7 @@ def eval_val_sliding(
 class LongPhraseCache:
     """variable-length suffix matcher for verbatim repetition (PR #880).
     probes at lengths [48,36,28,20,16] using rolling hashes."""
-    PROBE_LENGTHS = [48, 36, 28, 20, 16]  # full probe set (matching PR #880)
+    PROBE_LENGTHS = [36, 28, 20, 16]  # trimmed to fit with order-13
     PRIMES = [np.uint64(p) for p in [
         36313, 27191, 51647, 81929, 131071, 174763, 233017, 299993, 350377,
         412391, 479909, 541267, 613651, 700897, 786433, 850001, 921587,
@@ -2116,7 +2116,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float:
     ngram_enabled = bool(int(os.environ.get("NGRAM_ENABLED", "1")))
     sw_seq_len = effective_eval_seq_len
     if ngram_enabled:
-        ngram_order = int(os.environ.get("NGRAM_ORDER", "9"))
+        ngram_order = int(os.environ.get("NGRAM_ORDER", "13"))
         ngram_min_order = int(os.environ.get("NGRAM_MIN_ORDER", "2"))
         ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4194304"))
         ngram_min_count = int(os.environ.get("NGRAM_MIN_COUNT", "2"))