Add verified diamond and heart pattern examples

README.md

@@ -8,7 +8,7 @@ make bench
   --position-batch 128 --devices all --repeat 3
 ```
 
-自定义图案用 `--pattern` 或 `--pattern-file`，支持史莱姆、非史莱姆和忽略格；见 [6×6 苦力怕示例](bench/README.md#自定义图案6x6-苦力怕)。
+自定义图案用 `--pattern` 或 `--pattern-file`，支持史莱姆、非史莱姆和忽略格；见 [已验证的 5×5 爱心](bench/README.md#已验证实例5x5-爱心)、[4×4 菱形](bench/README.md#已验证实例4x4-菱形)及 [6×6 苦力怕示例](bench/README.md#自定义图案6x6-苦力怕)。
 
 H100 编译：`make -C bench -B CUDA_ARCH=90`。需要 CUDA Toolkit 和 C++17 编译器。普通分支不包含拒绝重抽补扫，不能将它的结果声明为严格完整目录；参数和计时口径见 [CLI 使用说明](bench/README.md)。
 

bench/README.md

@@ -16,6 +16,56 @@ mkdir -p artifacts
 
 默认编译 `sm_86` 并附带 PTX，适合本机 RTX 3060 Laptop。换架构时用 `-B` 强制重建；仅改变 Make 变量不会使已有文件自动过期。
 
+## 已验证实例：5×5 爱心
+
+种子 **`22428888508636`**，左上角区块 **`(-5,-147)`**（方块原点 `(-80,-2352)`）。每格一个区块，向右为 `+X`、向下为 `+Z`；16 格 `#` 必须是史莱姆，9 格 `.` 必须不是：
+
+```text
+.#.#.
+#####
+#####
+.###.
+..#..
+```
+
+```sh
+./bench/bin/slime-bench --radius 256 \
+  --pattern-file bench/patterns/heart-5x5.txt \
+  --position-batch 256 --max-results 1000
+
+# 单点复查，收集全部普通分支匹配：应得到 1,181 个种子
+./bench/bin/slime-bench --x -5 --z -147 --radius 0 \
+  --pattern-file bench/patterns/heart-5x5.txt --max-results 2000
+```
+
+RTX 3060 Laptop 实测：131,585 个锚点，普通分支共找到 1,456 个种子/位置匹配，扫描约 2.439 秒（含设备初始化的端到端约 2.620 秒）。保存的 1,000 个搜索样本，以及单点复查的全部 1,181 个种子，都通过独立 Python 实现的 Java 整数回绕及 `Random.nextInt(10)` 精确逐格验证。已知种子及计时见 [results/heart-5x5-example.json](results/heart-5x5-example.json)。GPU 回归检查单点计数、已知种子及全部样本，搜索计数不包含拒绝重抽补扫。
+
+## 已验证实例：4×4 菱形
+
+种子 **`126066202091224`**，左上角区块 **`(-1,-63)`**（方块原点 `(-16,-1008)`）。每格是一个区块，向右为 `+X`、向下为 `+Z`；`#` 为史莱姆，`.` 必须不是：
+
+```text
+.##.
+####
+####
+.##.
+```
+
+```sh
+# 搜索原点周围；这里只限制保存的样本数，仍会完整扫描
+./bench/bin/slime-bench --radius 64 \
+  --pattern-file bench/patterns/diamond-4x4.txt \
+  --position-batch 128 --max-results 1000
+
+# 单点复查：普通分支应得到 11,142 个种子
+./bench/bin/slime-bench --x -1 --z -63 --radius 0 \
+  --pattern-file bench/patterns/diamond-4x4.txt --max-results 64
+```
+
+RTX 3060 Laptop 实测：8,321 个锚点，普通分支找到 2,457,370 个种子/位置匹配，扫描约 0.924 秒（含设备初始化的端到端约 1.098 秒）。保存的 1,000 个样本全部通过独立 Python 实现的 Java 整数回绕及 `Random.nextInt(10)` 精确验证，包括四个非史莱姆角。样本顺序由 GPU 调度决定，不保证固定种子在截取的前 64 个样本中。已知种子及验证记录见 [results/diamond-4x4-example.json](results/diamond-4x4-example.json)，GPU 回归会核对单点计数并独立验证返回样本。
+
+此实例证明自定义正负约束搜索可行；搜索计数仍不包含拒绝重抽补扫。
+
 ## 自定义图案：6×6 苦力怕
 
 `--pattern` 接受 `/` 分隔的行，或用 `--pattern-file` 读取每行一排的文本文件。`#` 必须是史莱姆区块，`.` 必须不是，`?` 不限制。支持非正方形，宽、高分别为 1～64，至少含一个 `#`；不能与 `--size` 同时指定。不自动旋转或镜像。

bench/patterns/diamond-4x4.txt

@@ -0,0 +1,4 @@
+.##.
+####
+####
+.##.

bench/patterns/heart-5x5.txt

@@ -0,0 +1,5 @@
+.#.#.
+#####
+#####
+.###.
+..#..

bench/results/diamond-4x4-example.json

@@ -0,0 +1,49 @@
+{
+  "pattern": [
+    ".##.",
+    "####",
+    "####",
+    ".##."
+  ],
+  "seed48": 126066202091224,
+  "chunk_x": -1,
+  "chunk_z": -63,
+  "block_origin": [
+    -16,
+    -1008
+  ],
+  "verified_rows": [
+    ".##.",
+    "####",
+    "####",
+    ".##."
+  ],
+  "verification": {
+    "method": "independent Python implementation of Java int overflow and Random.nextInt(10) rejection loop",
+    "verified_sample_count": 1000,
+    "slime_cells": 12,
+    "non_slime_cells": 4
+  },
+  "search": {
+    "center": [
+      0,
+      0
+    ],
+    "radius_chunks": 64,
+    "anchors": 8321,
+    "matches": 2457370,
+    "scan_wall_ms": 924.202261,
+    "elapsed_ms": 1097.678503,
+    "gpu": "NVIDIA GeForce RTX 3060 Laptop GPU",
+    "rejection_tail_included": false
+  },
+  "single_anchor": {
+    "matches": 11142,
+    "root_jobs": 235
+  },
+  "source_sha256": {
+    "bench/slime_bench.cu": "3c70cbe2e5babc3e737199fb5e62b78ea9c5daadc0020af77292c2d59516259e",
+    "include/slime/solver_core.hpp": "feb1200518716b4d796fb60f60a83b6b798155bbd4e795531793181a5c7f9eba",
+    "include/slime/upper_interval.hpp": "af3fddf7705295d91ffd6e2190f542004a3ecff847fce9d4c9b13125bdbdc2d5"
+  }
+}

bench/results/heart-5x5-example.json

@@ -0,0 +1,52 @@
+{
+  "pattern": [
+    ".#.#.",
+    "#####",
+    "#####",
+    ".###.",
+    "..#.."
+  ],
+  "seed48": 22428888508636,
+  "chunk_x": -5,
+  "chunk_z": -147,
+  "block_origin": [
+    -80,
+    -2352
+  ],
+  "verified_rows": [
+    ".#.#.",
+    "#####",
+    "#####",
+    ".###.",
+    "..#.."
+  ],
+  "verification": {
+    "method": "independent Python implementation of Java int overflow and Random.nextInt(10) rejection loop",
+    "verified_search_sample_count": 1000,
+    "verified_single_anchor_count": 1181,
+    "slime_cells": 16,
+    "non_slime_cells": 9
+  },
+  "search": {
+    "center": [
+      0,
+      0
+    ],
+    "radius_chunks": 256,
+    "anchors": 131585,
+    "matches": 1456,
+    "scan_wall_ms": 2438.534811,
+    "elapsed_ms": 2619.994137,
+    "gpu": "NVIDIA GeForce RTX 3060 Laptop GPU",
+    "rejection_tail_included": false
+  },
+  "single_anchor": {
+    "matches": 1181,
+    "root_jobs": 405
+  },
+  "source_sha256": {
+    "bench/slime_bench.cu": "3c70cbe2e5babc3e737199fb5e62b78ea9c5daadc0020af77292c2d59516259e",
+    "include/slime/solver_core.hpp": "feb1200518716b4d796fb60f60a83b6b798155bbd4e795531793181a5c7f9eba",
+    "include/slime/upper_interval.hpp": "af3fddf7705295d91ffd6e2190f542004a3ecff847fce9d4c9b13125bdbdc2d5"
+  }
+}

bench/test_cli.py

@@ -40,8 +40,36 @@ def main():
     with tempfile.TemporaryDirectory() as tmp:
         path=Path(tmp)/'pattern.txt';path.write_bytes(b'?##.\r\n#..?\r\n..#?\r\n')
         assert run('--radius',0,'--pattern-file',path,'--dry-run')['pattern']==pattern.split('/')
+    diamond_fixture=json.loads((HERE/'results/diamond-4x4-example.json').read_text())
+    assert (HERE/'patterns/diamond-4x4.txt').read_text().splitlines()==diamond_fixture['pattern']
+    assert all(slime(diamond_fixture['seed48'],diamond_fixture['chunk_x']+dx,diamond_fixture['chunk_z']+dz)==(ch=='#')
+               for dz,row in enumerate(diamond_fixture['pattern']) for dx,ch in enumerate(row))
+    heart_fixture=json.loads((HERE/'results/heart-5x5-example.json').read_text())
+    assert (HERE/'patterns/heart-5x5.txt').read_text().splitlines()==heart_fixture['pattern']
+    assert all(slime(heart_fixture['seed48'],heart_fixture['chunk_x']+dx,heart_fixture['chunk_z']+dz)==(ch=='#')
+               for dz,row in enumerate(heart_fixture['pattern']) for dx,ch in enumerate(row))
     summary={'cli_scope_and_errors':True,'pattern_inputs':True}
     if args.gpu:
+        heart_poc=run('--x',heart_fixture['chunk_x'],'--z',heart_fixture['chunk_z'],'--radius',0,
+                      '--pattern-file',HERE/'patterns/heart-5x5.txt','--max-results',2000)
+        heart_hits=heart_poc['rounds'][0]
+        assert heart_hits['matches']==heart_fixture['single_anchor']['matches']
+        assert heart_hits['root_jobs']==heart_fixture['single_anchor']['root_jobs']
+        assert not heart_hits['sample_truncated'] and len(heart_hits['sample'])==heart_hits['matches']
+        assert len(set(map(tuple,heart_hits['sample'])))==heart_hits['matches']
+        assert [heart_fixture['seed48'],heart_fixture['chunk_x'],heart_fixture['chunk_z']] in heart_hits['sample']
+        for seed,x,z in heart_hits['sample']:
+            assert all(slime(seed,x+dx,z+dz)==(ch=='#') for dz,row in enumerate(heart_fixture['pattern']) for dx,ch in enumerate(row))
+        summary['heart_pattern_verified_matches']=heart_hits['matches']
+        diamond_poc=run('--x',diamond_fixture['chunk_x'],'--z',diamond_fixture['chunk_z'],'--radius',0,
+                        '--pattern-file',HERE/'patterns/diamond-4x4.txt','--max-results',64)
+        diamond_hits=diamond_poc['rounds'][0]
+        assert diamond_hits['matches']==diamond_fixture['single_anchor']['matches']
+        assert diamond_hits['root_jobs']==diamond_fixture['single_anchor']['root_jobs']
+        assert len(diamond_hits['sample'])==64
+        for seed,x,z in diamond_hits['sample']:
+            assert all(slime(seed,x+dx,z+dz)==(ch=='#') for dz,row in enumerate(diamond_fixture['pattern']) for dx,ch in enumerate(row))
+        summary['diamond_pattern_verified_matches']=diamond_hits['matches']
         # A mixed-mask regression fixture; this is not the target creeper pattern.
         fixture=json.loads((HERE/'testdata/mixed-mask-regression.json').read_text())
         face=run('--x',fixture['chunk_x'],'--z',fixture['chunk_z'],'--radius',0,