diff --git a/.github/workflows/R-CMD-check.yml b/.github/workflows/R-CMD-check.yml
index b9472d53..85b28879 100644
--- a/.github/workflows/R-CMD-check.yml
+++ b/.github/workflows/R-CMD-check.yml
@@ -21,9 +21,7 @@ on:
       - "**.R[dD]ata"
       - "**.Rpro*"
   pull_request:
-    branches:
-      - main
-      - master
+    branches: ["*"]
     paths-ignore:
       - "Meta**"
       - "memcheck**"
diff --git a/AGENTS.md b/AGENTS.md
index 2f0146ac..2b0ad10c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -34,3 +34,4 @@ binary 0/1 in an underlying `raw` object.
   before moving on to the next task.
 - Increment the `.900X` dev version suffix in `DESCRIPTION` with each
   `NEWS.md` update.
+- Check that existing tests cover all new code. (The GHA test suite uses codecov.)
diff --git a/NEWS.md b/NEWS.md
index 85481258..76e7bcb7 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,7 +1,8 @@
 # TreeTools 2.1.0.9005 (2026-03-13) #
 
-- `SplitFrequency(reference = NULL)` split normalization moved to C++,
-  eliminating an R-level per-split loop.
+- `SplitFrequency(reference = NULL)` performance improvements:
+  split normalization moved to C++; internal split de-duplication uses
+  hash map instead of ordered map.
 
 # TreeTools 2.1.0.9003 (2026-03-09) #
 
diff --git a/src/consensus.cpp b/src/consensus.cpp
index 93a3b02b..e6585df1 100644
--- a/src/consensus.cpp
+++ b/src/consensus.cpp
@@ -6,7 +6,8 @@ using namespace Rcpp;
 
 #include <algorithm> /* for fill */
 #include <array> /* for array */
-#include <map> /* for map */
+#include <string> /* for string (hash key) */
+#include <unordered_map> /* for unordered_map */
 
 using TreeTools::ct_stack_threshold;
 using TreeTools::ct_max_leaves_heap;
@@ -178,12 +179,15 @@ List calc_split_frequencies(
 
   StackEntry *const S_start = S.data();
   
-  // Use a map to store unique splits and their counts
-  // Key: split bit pattern; Value: index in output
-  std::map<std::vector<Rbyte>, int32> split_map;
+  // Hash map for O(1) amortized split deduplication
+  std::unordered_map<std::string, int32> split_map;
+  split_map.reserve(ntip_3 * 2);
   std::vector<std::vector<Rbyte>> split_patterns;
   std::vector<int32> counts;
 
+  // Reusable key buffer — avoids per-split heap allocation
+  std::string key(nbin, '\0');
+
   for (int32 i = 0; i < n_trees; ++i) {
     if (tables[i].NOSWX(ntip_3)) {
       continue;
@@ -247,21 +251,21 @@ List calc_split_frequencies(
       const int32 end   = tables[i].X_right(k + 1);
       if (start == 0 && end == 0) continue; // No valid cluster at this position
       
-      // Build the bit pattern for this split
-      std::vector<Rbyte> pattern(nbin, 0);
+      // Build the bit pattern into the reusable key buffer
+      std::fill(key.begin(), key.end(), '\0');
       for (int32 j = start; j <= end; ++j) {
         const int32 leaf_idx = tables[i].DECODE(j) - 1; // 0-based
         const int32 byte_idx = leaf_idx >> 3;
         const int32 bit_idx  = leaf_idx & 7;
-        pattern[byte_idx] |= (Rbyte(1) << bit_idx);
+        key[byte_idx] |= static_cast<char>(1 << bit_idx);
       }
       
-      auto it = split_map.find(pattern);
+      auto it = split_map.find(key);
       if (it == split_map.end()) {
         // New split: record it with count from this reference tree
         const int32 idx = split_patterns.size();
-        split_map[pattern] = idx;
-        split_patterns.push_back(std::move(pattern));
+        split_map.emplace(key, idx);
+        split_patterns.emplace_back(key.begin(), key.end());
         counts.push_back(split_count[k]);
       }
       // If already found, the first reference tree that found it has the
diff --git a/tests/testthat/test-Support.R b/tests/testthat/test-Support.R
index 7b249527..2fa0e342 100644
--- a/tests/testthat/test-Support.R
+++ b/tests/testthat/test-Support.R
@@ -59,6 +59,19 @@ test_that("Node support colours consistent", {
                c("oor", "1", "34", "67", "101"))
 })
 
+test_that("normalize_splits() covers all branches", {
+  # n_spare == 0 (nTip multiple of 8): complement without trailing-bit mask
+  trees16 <- c(PectinateTree(16), PectinateTree(16))
+  freq16 <- SplitFrequency(trees16)
+  expect_equal(length(freq16), NSplits(trees16[[1]]))
+  expect_true(all(attr(freq16, "count") == 2L))
+
+  # Trivial splits are filtered: a star tree produces no non-trivial splits
+  star <- c(StarTree(10), StarTree(10))
+  freq_star <- SplitFrequency(star)
+  expect_equal(length(freq_star), 0)
+})
+
 test_that("SplitFrequency() handles four-split trees", {
   trees <- AddTipEverywhere(BalancedTree(3))
   trees <- c(trees[1], trees)