Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias
backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.

tests/integration/test_eval_files.py

@@ -0,0 +1,23 @@
+from pathlib import Path
+
+from kda.training.eval_mt import _instruction
+from kda.training.prompts import instruction_prompt
+
+_ROOT = Path(__file__).resolve().parents[2] / "data" / "eval"
+
+
+def _lines(name: str) -> list[str]:
+    return [ln.strip() for ln in (_ROOT / name).read_text(encoding="utf-8").splitlines() if ln.strip()]
+
+
+def test_frozen_eval_files_are_aligned():
+    zh_src, zh_ref = _lines("zh2en.src.txt"), _lines("zh2en.ref.txt")
+    en_src, en_ref = _lines("en2zh.src.txt"), _lines("en2zh.ref.txt")
+    assert len(zh_src) == len(zh_ref) >= 16
+    assert len(en_src) == len(en_ref) >= 16
+    assert all("\t" not in s for s in zh_src + en_src)
+
+
+def test_eval_instruction_is_the_sft_template():
+    assert _instruction("q", "en") == instruction_prompt("q", "en")
+    assert _instruction("q", "zh") == instruction_prompt("q", "zh")