Sunrisepeak · Nov 15, 2026 · Nov 15, 2026
diff --git a/README.md b/README.md
@@ -13,7 +13,7 @@ The upstream C++ API is unchanged: `rcConfig`, `rcBuildPolyMesh`, `dtNavMesh`,
 `dtNavMeshQuery`, `dtCrowd`, `dtTileCache` and every other public name are the
 upstream entities, in the global namespace where upstream declares them. The
 four libraries (Recast, Detour, DetourCrowd, DetourTileCache) are vendored from
-the 1.6.0 release with three patches (`patches/`) and built with no contraction
+the 1.6.0 release with four patches (`patches/`) and built with no contraction
 of `a*b+c` into an FMA and no fast math (`-ffp-contract=off`, `-fno-fast-math`),
 the same in every profile (`NDEBUG`: upstream's asserts off). The trigonometry
 does not come from the platform's C library, whose `cosf` and `sinf` differ by
@@ -34,7 +34,7 @@ tests, compared on every CI row).
 ## Getting started
 
 ```sh
-mcpp new path --template mcpplibs.mwrap.recastnavigation@1.6.0.1
+mcpp new path --template mcpplibs.mwrap.recastnavigation@1.6.0.2
 cd path
 mcpp run
 ```
@@ -44,7 +44,7 @@ it and prints the way from one side of the wall to the other, around its end.
 
 ```toml
 [dependencies]
-mcpplibs.mwrap.recastnavigation = "1.6.0.1"
+mcpplibs.mwrap.recastnavigation = "1.6.0.2"
 
 # For the same path everywhere: Detour's inline math is compiled in your units too.
 [build]
@@ -86,11 +86,11 @@ The 64-bit polygon references (`DT_POLYREF64`) and the virtual query filter
 - Not here: DebugUtils (drawing a mesh through a debug renderer's callbacks), the demo (RecastDemo) and upstream's tests. The assert hooks (`rcAssertFailSetCustom`, `dtAssertFailSetCustom`) exist only in a build without `NDEBUG`; the package builds with it in every profile.
 - Not `compat.recastnavigation`: the index also has the same release as a traditional package; this wrap vendors its own copy because the same bits everywhere need the library itself compiled with `-ffp-contract=off`, no fast math and `NDEBUG` in every profile, which a dependency's build does not take from its consumer. A program links one of the two, not both.
 - A tile cache needs a compressor: `dtTileCacheCompressor` is an interface the program implements (upstream's demo uses FastLZ, which is not vendored).
-- Patched (`patches/`, the files' tarball and patched checksums in `upstream/documents.json`): 0001 makes DetourCrowd's obstacle avoidance and DetourTileCache's oriented boxes call Detour's own `dtMathCosf` and `dtMathSinf` instead of `cosf` and `sinf` (upstream-worthy); 0002 makes the hooks and Recast's two `cosf` calls use `port/` (wrapper-specific). 0003 fixes the adaptive sampling pattern for an even `adaptiveDivs`: a ring's last direction was written one slot past the one counted, so the crowd sampled an uninitialised direction, and past the array's end at 32 divisions and 4 rings (upstream issues #496, #798). An infinity or a NaN gives the one quiet NaN `0x7fc00000` on every platform.
+- Patched (`patches/`, the files' tarball and patched checksums in `upstream/documents.json`): 0001 makes DetourCrowd's obstacle avoidance and DetourTileCache's oriented boxes call Detour's own `dtMathCosf` and `dtMathSinf` instead of `cosf` and `sinf` (upstream-worthy); 0002 makes the hooks and Recast's two `cosf` calls use `port/` (wrapper-specific). 0003 fixes the adaptive sampling pattern for an even `adaptiveDivs`: a ring's last direction was written one slot past the one counted, so the crowd sampled an uninitialised direction, and past the array's end at 32 divisions and 4 rings (upstream issues #496, #798). 0004 makes the BV tree and the contours' hole merging sort with `std::stable_sort` instead of `qsort`, which is not stable: glibc, macOS, the Windows UCRT and emscripten order equal keys differently, so a mesh with ties (common on a grid) had different bytes by platform (upstream-worthy). An infinity or a NaN gives the one quiet NaN `0x7fc00000` on every platform.
 
 ## Verification
 
-- `mcpp test`: a navmesh built from a floor and a wall, the same bytes twice; a path around the wall, its straight path, a ray along the mesh stopped by the wall; a crowd of four walked through the doorway to the far side; eight agents crossing a circle through each other with four obstacle-avoidance settings (3 to 32 sample directions); the mesh's bytes and both crowds' states pinned; the adapted constants. `tests/trig_test.cpp`: sin, cos and atan2 against the double reference over a dense sweep (under 1 ulp; the results fingerprinted), C's special cases; `RECAST_TRIG_EXHAUSTIVE=1` sweeps every float.
+- `mcpp test`: a navmesh built from a floor and a wall, the same bytes twice; a path around the wall, its straight path, a ray along the mesh stopped by the wall; a crowd of four walked through the doorway to the far side; eight agents crossing a circle through each other with four obstacle-avoidance settings (3 to 32 sample directions); a hall tiled in four, its pillars and rooms on a grid so the BV items and contours tie, eight agents walked across it; the meshes' bytes and the crowds' states pinned; the adapted constants. `tests/trig_test.cpp`: sin, cos and atan2 against the double reference over a dense sweep (under 1 ulp; the results fingerprinted), C's special cases; `RECAST_TRIG_EXHAUSTIVE=1` sweeps every float.
 - `python3 tools/check_exports.py`: the module is exactly `tools/surface.json`; given a clang++ (required on CI's LLVM row) the record is derived again from the headers and compared.
 - `python3 tools/test_template.py`: `mcpp new` from an installed copy, a strict build, and a run of the generated project.
 - CI: Linux x86-64 and arm64 (GCC 16, LLVM 23), macOS and Windows (LLVM 23), each in the dev and release profiles, and wasm32 under node — the same pinned hashes on every row.

diff --git a/README.zh.md b/README.zh.md
@@ -11,7 +11,7 @@
 
 上游的 C++ API 保持原样：`rcConfig`、`rcBuildPolyMesh`、`dtNavMesh`、`dtNavMeshQuery`、`dtCrowd`、`dtTileCache`
 以及其他公开名字都是上游的实体，在上游声明它们的全局命名空间里。四个库（Recast、Detour、DetourCrowd、
-DetourTileCache）取自 1.6.0 发布包并打了三个补丁（`patches/`），构建时不把 `a*b+c` 合并为 FMA、不用快速数学
+DetourTileCache）取自 1.6.0 发布包并打了四个补丁（`patches/`），构建时不把 `a*b+c` 合并为 FMA、不用快速数学
 （`-ffp-contract=off`、`-fno-fast-math`），且每种配置都相同（`NDEBUG`：不打开上游的断言）。三角函数不取自平台的
 C 库（glibc、macOS、Windows UCRT 与 emscripten 的 `cosf`、`sinf` 相差一个 ulp）：Detour 的钩子 `dtMathCosf`、
 `dtMathSinf`、`dtMathAtan2f`（群体与瓦片缓存现在按上游本意调用它们）和 Recast 求可行走坡度的 `cosf` 都由
@@ -27,7 +27,7 @@ C 库（glibc、macOS、Windows UCRT 与 emscripten 的 `cosf`、`sinf` 相差
 ## 快速开始
 
 ```sh
-mcpp new path --template mcpplibs.mwrap.recastnavigation@1.6.0.1
+mcpp new path --template mcpplibs.mwrap.recastnavigation@1.6.0.2
 cd path
 mcpp run
 ```
@@ -36,7 +36,7 @@ mcpp run
 
 ```toml
 [dependencies]
-mcpplibs.mwrap.recastnavigation = "1.6.0.1"
+mcpplibs.mwrap.recastnavigation = "1.6.0.2"
 
 # 各平台得到同样的路线：Detour 的内联数学也在你的编译单元中编译。
 [build]
@@ -73,11 +73,11 @@ int main() {
 - 不在此：DebugUtils（通过调试渲染器的回调绘制网格）、演示程序（RecastDemo）与上游测试。断言钩子（`rcAssertFailSetCustom`、`dtAssertFailSetCustom`）只在未定义 `NDEBUG` 的构建中存在；本包在每种配置都定义它。
 - 不是 `compat.recastnavigation`：索引里也有同一发布的传统包；本封装自带一份副本，因为“处处同位”要求库本身以 `-ffp-contract=off`、不用快速数学、各配置都定义 `NDEBUG` 编译，而依赖的构建不接受消费者的这些标志。一个程序只链接两者之一。
 - 分块缓存需要一个压缩器：`dtTileCacheCompressor` 是由程序实现的接口（上游演示用的 FastLZ 未收录）。
-- 补丁（`patches/`，文件在发布包中与打补丁后的校验和见 `upstream/documents.json`）：0001 让 DetourCrowd 的避障与 DetourTileCache 的有向盒改调 Detour 自己的 `dtMathCosf`、`dtMathSinf`，不再直接调 `cosf`、`sinf`（值得回馈上游）；0002 让这些钩子和 Recast 的两处 `cosf` 使用 `port/`（本封装专有）。0003 修正 `adaptiveDivs` 为偶数时的自适应采样模式：每圈最后一个方向被写到计数槽位的下一个，群体因而采样一个未初始化的方向，在 32 个方向、4 圈时还写出数组末尾（上游 issue #496、#798）。无穷与 NaN 在每个平台都得到同一个静默 NaN `0x7fc00000`。
+- 补丁（`patches/`，文件在发布包中与打补丁后的校验和见 `upstream/documents.json`）：0001 让 DetourCrowd 的避障与 DetourTileCache 的有向盒改调 Detour 自己的 `dtMathCosf`、`dtMathSinf`，不再直接调 `cosf`、`sinf`（值得回馈上游）；0002 让这些钩子和 Recast 的两处 `cosf` 使用 `port/`（本封装专有）。0003 修正 `adaptiveDivs` 为偶数时的自适应采样模式：每圈最后一个方向被写到计数槽位的下一个，群体因而采样一个未初始化的方向，在 32 个方向、4 圈时还写出数组末尾（上游 issue #496、#798）。0004 让 BV 树与轮廓的孔洞合并改用 `std::stable_sort` 排序，不再用不稳定的 `qsort`：glibc、macOS、Windows UCRT 与 emscripten 对相等键的排列不同，带并列键的网格（网格状布局很常见）因而在各平台得到不同的字节（值得回馈上游）。无穷与 NaN 在每个平台都得到同一个静默 NaN `0x7fc00000`。
 
 ## 验证
 
-- `mcpp test`：由地面与墙构建导航网格，两次得到相同字节；绕墙的路径、其直线路径、沿网格被墙挡住的射线；四个智能体组成的群体穿过门洞走到另一侧；八个智能体以四种避障设置（每圈 3 到 32 个采样方向）相互穿过一个圆；钉住网格字节与两个群体状态的哈希；改写的常量。`tests/trig_test.cpp`：sin、cos、atan2 在密集扫描上与 double 参考比较（小于 1 ulp，结果取指纹）以及 C 的特殊情形；`RECAST_TRIG_EXHAUSTIVE=1` 遍历全部 float。
+- `mcpp test`：由地面与墙构建导航网格，两次得到相同字节；绕墙的路径、其直线路径、沿网格被墙挡住的射线；四个智能体组成的群体穿过门洞走到另一侧；八个智能体以四种避障设置（每圈 3 到 32 个采样方向）相互穿过一个圆；分成四块的大厅，柱子与房间排在网格上，使 BV 项与轮廓出现并列键，八个智能体穿过它；钉住网格字节与群体状态的哈希；改写的常量。`tests/trig_test.cpp`：sin、cos、atan2 在密集扫描上与 double 参考比较（小于 1 ulp，结果取指纹）以及 C 的特殊情形；`RECAST_TRIG_EXHAUSTIVE=1` 遍历全部 float。
 - `python3 tools/check_exports.py`：模块与 `tools/surface.json` 完全一致；给定 clang++（CI 的 LLVM 行必需）时再次从头文件推导记录并比对。
 - `python3 tools/test_template.py`：从安装副本 `mcpp new`、严格构建并运行生成的项目。
 - CI：Linux x86-64 与 arm64（GCC 16、LLVM 23），macOS 与 Windows（LLVM 23），各在 dev 与 release 配置下，以及 node 下的 wasm32——每一行都是同样的钉住哈希。

diff --git a/mcpp.toml b/mcpp.toml
@@ -2,7 +2,7 @@
 [package]
 namespace   = "mcpplibs.mwrap"
 name        = "recastnavigation"
-version     = "1.6.0.1"
+version     = "1.6.0.2"
 description = "Recast & Detour 1.6.0 C++ modules for mcpp (import mcpplibs.mwrap.recastnavigation;): navigation meshes built from triangles, path queries, crowds of agents, tiled meshes with obstacles"
 license     = "Apache-2.0"
 authors     = ["mcpp-community"]
@@ -29,6 +29,10 @@ kind = "lib"
 # float). sqrtf, floorf, ceilf and fabsf are exact under IEEE everywhere.
 # patches/0003: with an even adaptiveDivs the crowd sampled an uninitialised
 # direction (upstream issue #496), which no flag makes the same everywhere.
+# patches/0004: the BV tree and the contours' hole merging sort with
+# std::stable_sort instead of the C library's qsort, which is not stable and
+# orders equal keys differently between glibc, macOS, the Windows UCRT and
+# emscripten's musl, so a mesh with ties had different bytes by platform.
 # NDEBUG: the same library in every profile — upstream declares its assert
 # hooks (rcAssertFailSetCustom, dtAssertFailSetCustom) and checks only when
 # NDEBUG is not defined, which would make the module's names differ between a

diff --git a/patches/0004-stable-sorts.patch b/patches/0004-stable-sorts.patch
@@ -0,0 +1,79 @@
+Stable sorts: the same BV tree and contours on every platform
+
+dtCreateNavMeshData's BV tree (subdivide: compareItemX, compareItemY,
+compareItemZ) and rcBuildContours' hole merging (mergeRegionHoles:
+compareHoles, compareDiagDist) sort with the C library's qsort. qsort is not
+stable, and glibc, macOS libSystem, the Windows UCRT and emscripten's musl
+put items with equal keys in different orders; on a mesh laid out on a grid
+the keys tie often, so the same input gave different BV trees and contour
+merges, different mesh bytes, and a crowd on the mesh walked differently by
+platform. The five calls use std::stable_sort with the same comparators: a
+stable sort's result is determined by the comparator and the input order, so
+it is the same everywhere. Where the keys do not tie nothing changes.
+Upstream-worthy, not yet upstream.
+
+Apply from upstream/recastnavigation: git apply -p1 ../../patches/0004-stable-sorts.patch
+
+diff --git a/Detour/Source/DetourNavMeshBuilder.cpp b/Detour/Source/DetourNavMeshBuilder.cpp
+index e93a976..b52f041 100644
+--- a/Detour/Source/DetourNavMeshBuilder.cpp
++++ b/Detour/Source/DetourNavMeshBuilder.cpp
+@@ -17,6 +17,7 @@
+ //
+
+ #include <stdio.h>
++#include <algorithm>
+ #include <stdlib.h>
+ #include <string.h>
+ #include <float.h>
+@@ -142,17 +143,17 @@ static void subdivide(BVItem* items, int nitems, int imin, int imax, int& curNod
+ 		if (axis == 0)
+ 		{
+ 			// Sort along x-axis
+-			qsort(items+imin, inum, sizeof(BVItem), compareItemX);
++			std::stable_sort(items+imin, items+imin+inum, [](const BVItem& a, const BVItem& b) { return compareItemX(&a, &b) < 0; });
+ 		}
+ 		else if (axis == 1)
+ 		{
+ 			// Sort along y-axis
+-			qsort(items+imin, inum, sizeof(BVItem), compareItemY);
++			std::stable_sort(items+imin, items+imin+inum, [](const BVItem& a, const BVItem& b) { return compareItemY(&a, &b) < 0; });
+ 		}
+ 		else
+ 		{
+ 			// Sort along z-axis
+-			qsort(items+imin, inum, sizeof(BVItem), compareItemZ);
++			std::stable_sort(items+imin, items+imin+inum, [](const BVItem& a, const BVItem& b) { return compareItemZ(&a, &b) < 0; });
+ 		}
+
+ 		int isplit = imin+inum/2;
+diff --git a/Recast/Source/RecastContour.cpp b/Recast/Source/RecastContour.cpp
+index 5508a98..7e0f112 100644
+--- a/Recast/Source/RecastContour.cpp
++++ b/Recast/Source/RecastContour.cpp
+@@ -17,6 +17,7 @@
+ //
+
+ #include <math.h>
++#include <algorithm>
+ #include <string.h>
+ #include <stdio.h>
+ #include <stdlib.h>
+@@ -724,7 +725,7 @@ static void mergeRegionHoles(rcContext* ctx, rcContourRegion& region)
+ 	for (int i = 0; i < region.nholes; i++)
+ 		findLeftMostVertex(region.holes[i].contour, &region.holes[i].minx, &region.holes[i].minz, &region.holes[i].leftmost);
+
+-	qsort(region.holes, region.nholes, sizeof(rcContourHole), compareHoles);
++	std::stable_sort(region.holes, region.holes + region.nholes, [](const rcContourHole& a, const rcContourHole& b) { return compareHoles(&a, &b) < 0; });
+
+ 	int maxVerts = region.outline->nverts;
+ 	for (int i = 0; i < region.nholes; i++)
+@@ -770,7 +771,7 @@ static void mergeRegionHoles(rcContext* ctx, rcContourRegion& region)
+ 				}
+ 			}
+ 			// Sort potential diagonals by distance, we want to make the connection as short as possible.
+-			qsort(diags, ndiags, sizeof(rcPotentialDiagonal), compareDiagDist);
++			std::stable_sort(&diags[0], &diags[0] + ndiags, [](const rcPotentialDiagonal& a, const rcPotentialDiagonal& b) { return compareDiagDist(&a, &b) < 0; });
+
+ 			// Find a diagonal that is not intersecting the outline not the remaining holes.
+ 			index = -1;