fix(ci): resolve macOS bit-exactness failure and Linux AVX-512 compile error

- **Linux (GCC)**: Include `avx512vpopcntdq` in `DEGLIB_TARGET_AVX512` definition so GCC allows inlining `_mm512_popcnt_epi64` into AVX-512 distance methods.
- **macOS (Clang/libc++)**: Add index-based tie-breaking (`a.first != b.first ? a.first < b.first : a.second < b.second`) in `quantize_single` and `quantize_batch` calls to `std::nth_element` in `evp_quantize.h`. This guarantees identical top-K element partitioning regardless of standard library implementation differences.

Author: Neiko2002

cpp/deglib/include/config.h

@@ -9,7 +9,7 @@
 
 // Compile methods with this attribute for AVX-512 functions on GCC/Clang
 #if defined(DEGLIB_X86) && (defined(__GNUC__) || defined(__clang__))
-#define DEGLIB_TARGET_AVX512 __attribute__((target("avx512f,avx512dq,avx512bw,fma")))
+#define DEGLIB_TARGET_AVX512 __attribute__((target("avx512f,avx512dq,avx512bw,avx512vpopcntdq,fma")))
 #else
 #define DEGLIB_TARGET_AVX512
 #endif

cpp/deglib/include/quantization/evp_quantize.h

@@ -43,7 +43,9 @@ inline std::vector<std::byte> quantize_single(const float* embedding, uint32_t d
         abs_vals[i] = {std::abs(embedding[i]), i};
     }
     std::nth_element(abs_vals.begin(), abs_vals.begin() + (dim - non_zeros), abs_vals.end(),
-                 [](const auto& a, const auto& b) { return a.first < b.first; });
+                 [](const auto& a, const auto& b) {
+                     return a.first != b.first ? a.first < b.first : a.second < b.second;
+                 });
 
     std::vector<uint8_t> is_top(dim, 0);
     for (uint32_t j = dim - non_zeros; j < dim; ++j) {
@@ -128,7 +130,9 @@ inline std::vector<std::byte> quantize_batch(const float* data, size_t count, ui
                     abs_vals[k] = {std::abs(emb[k]), k};
                 }
                 std::nth_element(abs_vals.begin(), abs_vals.begin() + (dim - non_zeros), abs_vals.end(),
-                             [](const auto& a, const auto& b) { return a.first < b.first; });
+                             [](const auto& a, const auto& b) {
+                                 return a.first != b.first ? a.first < b.first : a.second < b.second;
+                             });
 
                 is_top.assign(dim, 0);
                 for (uint32_t j = dim - non_zeros; j < dim; ++j) {
@@ -243,7 +247,9 @@ inline std::vector<std::byte> quantize_batch(const uint16_t* data, size_t count,
                     abs_vals[k] = {static_cast<uint16_t>(emb[k] & 0x7FFFu), k};
                 }
                 std::nth_element(abs_vals.begin(), abs_vals.begin() + (dim - non_zeros), abs_vals.end(),
-                             [](const auto& a, const auto& b) { return a.first < b.first; });
+                             [](const auto& a, const auto& b) {
+                                 return a.first != b.first ? a.first < b.first : a.second < b.second;
+                             });
 
                 is_top.assign(dim, 0);
                 for (uint32_t j = dim - non_zeros; j < dim; ++j) {
@@ -334,7 +340,9 @@ inline std::vector<std::byte> quantize_batch(const std::vector<std::vector<std::
                     abs_vals[k] = {static_cast<uint16_t>(emb[k] & 0x7FFFu), k};
                 }
                 std::nth_element(abs_vals.begin(), abs_vals.begin() + (dim - non_zeros), abs_vals.end(),
-                             [](const auto& a, const auto& b) { return a.first < b.first; });
+                             [](const auto& a, const auto& b) {
+                                 return a.first != b.first ? a.first < b.first : a.second < b.second;
+                             });
 
                 is_top.assign(dim, 0);
                 for (uint32_t j = dim - non_zeros; j < dim; ++j) {

cpp/test/src/integration/test_builder_integration.cpp

@@ -354,9 +354,9 @@ TEST(DeglibBuilderIntegration, EvpDatasetBitExactness)
 
     // Hardcoded expected hashes computed on Windows (10k dataset)
     // Checksum verification ensures 100% bit-exact dataset generation and groundtruth across OS/compilers.
-    EXPECT_EQ(base_hash, 0xe8de72f535d6e07dULL) << "base_evp checksum mismatch across platforms!";
-    EXPECT_EQ(query_hash, 0xccb8e62c3224256fULL) << "query_evp checksum mismatch across platforms!";
-    EXPECT_EQ(gt_evp_hash, 0xc01f476e432ca611ULL) << "gt_evp checksum mismatch across platforms!";
+    EXPECT_EQ(base_hash, 0x524834ef619bf297ULL) << "base_evp checksum mismatch across platforms!";
+    EXPECT_EQ(query_hash, 0x942208b0d125a817ULL) << "query_evp checksum mismatch across platforms!";
+    EXPECT_EQ(gt_evp_hash, 0x6c0ad164a1cbbd8dULL) << "gt_evp checksum mismatch across platforms!";
 }
 
 TEST(DeglibBuilderIntegration, EVPInnerProduct_Recall)