Keep attnres on resume, fix final chunk_index, default 0.5b to Yi-6B

CLI default attnres=off was overwriting block checkpoints on resume
so later loads hit Unexpected key(s). Only apply flags the user
passed. Track next_chunk so a budget-exit save does not skip the
untrained yield. 0.5b now uses 01-ai/Yi-6B (64k); refuse resume
when the ckpt tokenizer does not match.

kda/models/k3_config.py

@@ -9,13 +9,15 @@ Hybrid Attention (K3): 每 4 层 1 次 Gated MLA, 末层强制 MLA.
 
 Presets:
     toy  — ~8M, 自训 8k SP, 本地过拟合
-    0.5b — ~482M, Qwen3 词表, 32–40GB bf16；默认 step 是冒烟，翻译前置用 --max-tokens
+    0.5b — ~415M, Yi-6B 词表 (64k), 32–40GB bf16；默认 step 是冒烟，翻译前置用 --max-tokens
 """
 from __future__ import annotations
 
 from dataclasses import dataclass
 
-# Qwen3 config.json; train_k3 overrides with len(tokenizer).
+# 01-ai/Yi-6B config.json; train_k3 overrides with len(tokenizer).
+YI6B_VOCAB_SIZE = 64000
+# Kept for old Qwen3 checkpoints / docs.
 QWEN3_VOCAB_SIZE = 151936
 
 
@@ -24,7 +26,7 @@ class K3Config:
     # 主干
     hidden_size: int = 256
     num_hidden_layers: int = 4
-    vocab_size: int = 8192               # toy: data/spm_4k; 0.5b: Qwen3
+    vocab_size: int = 8192               # toy: data/spm_4k; 0.5b: Yi-6B 64k
     initializer_range: float = 0.02
     norm_eps: float = 1e-6
     tie_word_embeddings: bool = False
@@ -76,11 +78,11 @@ class K3Config:
             return cls()
         if name in {"0.5b", "500m"}:
             # H * head_dim == hidden. Routed 16 Top-2; LatentMoE padded bmm.
-            # ~482M with tied Qwen3 embeddings. 6×(3 KDA + 1 MLA).
+            # ~415M with tied Yi-6B embeddings. 6×(3 KDA + 1 MLA).
             return cls(
                 hidden_size=768,
                 num_hidden_layers=24,
-                vocab_size=QWEN3_VOCAB_SIZE,
+                vocab_size=YI6B_VOCAB_SIZE,
                 tie_word_embeddings=True,
                 max_position_embeddings=2048,
                 num_heads=12,