Portable golden reference (JSON instead of LFS npz); doc links updated

.gitignore

@@ -21,7 +21,7 @@ env/
 # Model weights and checkpoints (this release ships no weights)
 *.npz
 *.npy
-!tests/golden/*.npz
+!tests/golden/*.json
 *.pt
 *.pth
 *.bin

docs/ZeroPoint.md

@@ -181,7 +181,7 @@ A = ( a ⊙ λ^t ) @ ( w ⊙ λ^(−u) )ᵀ    再把 u > t 的部分清零（
 **递推形式（推理用）**。逐字更新 `S`，每层只存 M·d 个浮点数，跟文本多长无关。
 
 两种形式结果一致：自检在**所有 64 种开关组合**下、300 个字（跨多个块）上逐位比对，float64 误差 < 1e-8。
-另有一份固定的参考数据 `tests/golden/zeropoint_ref.npz`，以后任何实现（包括内核里的 C 版本）都拿它对答案。
+另有一份固定的参考数据 `tests/golden/zeropoint_ref.json`，以后任何实现（包括内核里的 C 版本）都拿它对答案。
 **训练好的模型可以直接用递推形式放进内核**：不用 KV 缓存，每层固定一块 M × d 的状态，不需要动态分配。
 
 关于"最长路径 O(1)"要说实话：u 处写入的信息确实一步就能被 t 处读到，但已经乘上了 λ^(t−u)。

docs/ZeroPoint_EN.md

@@ -157,7 +157,7 @@ Within a block it is matrix multiplication, and only an M × d state is passed b
 
 **Recurrent form (for inference)**. Update `S` character by character; each layer stores only M·d floating-point numbers, regardless of how long the text is.
 
-The two forms agree: the self-check compares them bit by bit on **all 64 switch combinations** over 300 characters (crossing several blocks), with float64 error < 1e-8. There is also a fixed reference file `tests/golden/zeropoint_ref.npz`; any future implementation (including a C version in the kernel) can check its answers against it. **A trained model can be put directly into the kernel in recurrent form**: no KV cache, a fixed M × d state per layer, no dynamic allocation.
+The two forms agree: the self-check compares them bit by bit on **all 64 switch combinations** over 300 characters (crossing several blocks), with float64 error < 1e-8. There is also a fixed reference file `tests/golden/zeropoint_ref.json`; any future implementation (including a C version in the kernel) can check its answers against it. **A trained model can be put directly into the kernel in recurrent form**: no KV cache, a fixed M × d state per layer, no dynamic allocation.
 
 About "maximum path O(1)", we must be honest: information written at u can indeed be read at t in one step, but it has already been multiplied by λ^(t−u). Attention can "retrieve verbatim" a character from arbitrarily far away; ZeroPoint can only retrieve information **that has not yet been forgotten and has not been washed out by other writes**.
 

scripts/upload.md

@@ -69,7 +69,7 @@ hf upload <your-username>/ZeroPoint . --type model \
 ```
 
 This is a single-commit upload of the whole tree. It includes
-`tests/golden/zeropoint_ref.npz` (the fixed reference data required by the self-checks)
+`tests/golden/zeropoint_ref.json` (the fixed reference data required by the self-checks)
 and must not include model weights, training runs, corpora, or tokens — see
 `.gitignore`; the release tree itself contains none of those.
 

tests/make_golden.py

@@ -1,10 +1,11 @@
-"""生成 ZeroPoint 的参考数据 tests/golden/zeropoint_ref.npz：固定的参数 + 输入 + 标准输出。
+"""生成 ZeroPoint 的参考数据 tests/golden/zeropoint_ref.json：固定的参数 + 输入 + 标准输出。
 
 所有实现（zerocg 并行形式、numpy 递推形式、以后内核里的 C 版本）都拿它对答案，
 谁和它对不上谁就有 bug。**只有在有意修改架构时才重新生成**，并在提交说明里写清楚为什么。
 
     python3 tests/make_golden.py
 """
+import json
 import os
 import sys
 
@@ -26,7 +27,11 @@ if __name__ == "__main__":
         p.data = p.data + np.random.randn(*p.shape) * 0.3
     idx = np.random.randint(0, CONFIG["vocab_size"], size=(2, 70))   # 跨 3 个块
     logits = model(idx).data
-    out = os.path.join(HERE, "golden", "zeropoint_ref.npz")
-    np.savez(out, idx=idx, logits=logits, **{"param:" + k: v for k, v in model.state_dict().items()},
-             **{"config:" + k: v for k, v in CONFIG.items()})
+    out = os.path.join(HERE, "golden", "zeropoint_ref.json")
+    data = {"idx": idx.tolist(), "logits": logits.tolist()}
+    data.update({"param:" + k: v.tolist() for k, v in model.state_dict().items()})
+    data.update({"config:" + k: int(v) for k, v in CONFIG.items()})
+    with open(out, "w", encoding="utf-8") as f:
+        json.dump(data, f, separators=(",", ":"), ensure_ascii=False)
+        f.write("\n")
     print(f"写入 {out}：输入 {idx.shape}，输出 {logits.shape}，{len(model.state_dict())} 个参数张量")

tests/test_zeropoint.py

@@ -1,5 +1,6 @@
 """ZeroPoint 架构自检。运行: python3 tests/test_zeropoint.py"""
 import itertools
+import json
 import os
 import sys
 
@@ -35,13 +36,16 @@ def test_parallel_equals_recurrent():
 
 
 def test_matches_golden_reference():
-    """和 tests/golden/zeropoint_ref.npz 对答案：防止以后改代码时并行或递推形式悄悄变了。"""
-    ref = np.load(os.path.join(HERE, "golden", "zeropoint_ref.npz"))
-    cfg = {k[7:]: int(ref[k]) for k in ref.files if k.startswith("config:")}
+    """和 tests/golden/zeropoint_ref.json 对答案：防止以后改代码时并行或递推形式悄悄变了。"""
+    with open(os.path.join(HERE, "golden", "zeropoint_ref.json"), encoding="utf-8") as f:
+        ref = json.load(f)
+    cfg = {k[7:]: int(v) for k, v in ref.items() if k.startswith("config:")}
     model = ZeroX(**cfg).eval()
-    model.load_state_dict({k[6:]: ref[k] for k in ref.files if k.startswith("param:")})
-    assert np.abs(model(ref["idx"]).data - ref["logits"]).max() < 1e-10
-    assert np.abs(model.forward_recurrent(ref["idx"]) - ref["logits"]).max() < 1e-8
+    model.load_state_dict({k[6:]: np.array(v, dtype=np.float64) for k, v in ref.items()
+                           if k.startswith("param:")})
+    idx, logits = np.array(ref["idx"]), np.array(ref["logits"])
+    assert np.abs(model(idx).data - logits).max() < 1e-10
+    assert np.abs(model.forward_recurrent(idx) - logits).max() < 1e-8
 
 
 def test_causal_across_chunks():