[ML] Fix macOS debug build failures and aarch64 timing regression

The debug build pipeline has been failing on macOS since Feb 10 because
the artifact_paths glob `*/**/unittest/ml_test_*` matched the test
executables themselves — debug binaries are enormous and exceeded
Buildkite's 10 GiB artifact limit. Remove the overly broad glob.

Disable unity builds on macOS (only 4 Orka cores — unity is
counterproductive) and revert test parallelism to -j 2 for <=4-core
machines to avoid starving timing-sensitive tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: edsavage

.buildkite/pipelines/build_macos.json.py

@@ -43,7 +43,7 @@
       "PATH": "/opt/homebrew/bin:$PATH",
       "ML_DEBUG": "0",
       "CPP_CROSS_COMPILE": "",
-      "CMAKE_FLAGS": "-DCMAKE_TOOLCHAIN_FILE=cmake/darwin-aarch64.cmake -DCMAKE_UNITY_BUILD=ON -DML_PCH=ON",
+      "CMAKE_FLAGS": "-DCMAKE_TOOLCHAIN_FILE=cmake/darwin-aarch64.cmake -DML_PCH=ON",
       "RUN_TESTS": "true",
       "BOOST_TEST_OUTPUT_FORMAT_FLAGS": "--logger=JUNIT,error,boost_test_results.junit",
     }
@@ -67,7 +67,7 @@ def main(args):
             "depends_on": "check_style",
             "key": f"build_test_macos-{arch}-{build_type}",
             "env": envs[arch],
-            "artifact_paths": "*/**/unittest/boost_test_results.junit;*/**/unittest/ml_test_*",
+            "artifact_paths": "*/**/unittest/boost_test_results.junit",
             "plugins": {
               "test-collector#v1.2.0": {                                                              
                 "files": "*/*/unittest/boost_test_results.junit",

build.gradle

@@ -126,11 +126,10 @@ project.ext.numCpus = Runtime.runtime.availableProcessors()
 // many suites simultaneously causes resource contention. These values were
 // determined empirically (see PR #2900).
 // On low-core machines (<=4, e.g. macOS CI Orka VMs with 4 cores),
-// cap at numCpus-1 to leave headroom for timing-sensitive tests
-// (e.g. CKMostCorrelatedTest/testScale). For higher core counts,
-// ceil(numCpus/2) balances parallelism vs contention — ceil(numCpus/3)
-// was too conservative on 8-core machines (52 min vs 39.5 min).
-project.ext.testParallel = isWindows ? 2 : (numCpus <= 4 ? Math.max(2, numCpus - 1) : Math.max(2, (int) Math.ceil(numCpus / 2.0)))
+// use 2 to avoid starving timing-sensitive tests like
+// CKMostCorrelatedTest/testScale.  For higher core counts,
+// ceil(numCpus/2) balances parallelism vs contention.
+project.ext.testParallel = isWindows ? 2 : (numCpus <= 4 ? 2 : Math.max(2, (int) Math.ceil(numCpus / 2.0)))
 project.ext.makeEnvironment = [ 'CPP_CROSS_COMPILE': cppCrossCompile,
                                 'VERSION_QUALIFIER': versionQualifier,
                                 'SNAPSHOT': (isSnapshot ? 'yes' : 'no'),

dev-tools/docker/docker_entrypoint.sh

@@ -110,7 +110,7 @@ if [ "x$1" = "x--test" ] ; then
     # For higher core counts, ceil(ncpus/2) balances parallelism vs
     # contention — ceil(ncpus/3) was too conservative on 8-core machines.
     if [ "$NCPUS" -le 4 ]; then
-        TEST_PARALLEL=$(( NCPUS > 2 ? NCPUS - 1 : 2 ))
+        TEST_PARALLEL=2
     else
         TEST_PARALLEL=$(( (NCPUS + 1) / 2 ))
     fi