diff --git a/src/tools/xce-build.sh b/src/tools/xce-build.sh
index 63eb346b..315a4ce3 100755
--- a/src/tools/xce-build.sh
+++ b/src/tools/xce-build.sh
@@ -149,14 +149,14 @@ fi
 # those the same value is used as a job-count ceiling instead.
 if [ -n "$JOBS" ]; then
     echo "=== Run-queue-length limit (explicit --jobs): $JOBS (meson compile -l load-average ceiling, meson test -j, go test -p) ==="
+    COMPILE_JOBS="$JOBS"
 else
     # Live default -- same formula the /xce-build skill computes by hand: a
     # per-profile memory budget per job (ASAN+debug's sanitizer
     # instrumentation and full debug info are the heaviest), reserving ~15%
-    # of available memory as headroom, and a per-profile PROCESS_FACTOR
-    # discounting nproc-minus-load-average for the compiler driver's forked
-    # cc1/cc1plus + as children (asan/debug keep those resident far longer
-    # than release's leaner ones).
+    # of available memory as headroom. The CPU term is raw nproc-minus-load;
+    # the forked-children discount is applied separately to the flat test
+    # caps below (see the COMPILE_JOBS comment), not to the -l ceiling.
     #
     # asan's 512MB budget (build/de-nix): the two memory hogs the original
     # 1536MB figure was calibrated around are both gone -- the 293MB-peak
@@ -179,11 +179,20 @@ else
         release) default_per_job_mb=1024; default_process_factor=0.7 ;;
     esac
 
+    # The memory term is where the per-job budget applies. The CPU term uses
+    # the RAW nproc-minus-load (no PROCESS_FACTOR discount) because the
+    # compile step consumes COMPILE_JOBS as ninja's -l LOAD-AVERAGE ceiling
+    # (see below), which re-checks the live run-queue before every launch and
+    # backs off by itself -- pre-emptively halving the ceiling for forked
+    # cc1/as children on top of that would roughly halve t1's compile
+    # parallelism for no added safety. The discount still matters for meson
+    # test -j / go test -p (flat caps with no live check), so those two get
+    # the discounted number.
     JOBS="$(awk -v nproc="$default_nproc" -v mem="$default_mem_mb" \
                 -v loadavg="$default_load" -v per_job="$default_per_job_mb" \
-                -v factor="$default_process_factor" 'BEGIN {
+                'BEGIN {
         mem_jobs = int(mem * 0.85 / per_job)
-        cpu_jobs = int((nproc - loadavg) * factor)
+        cpu_jobs = int(nproc - loadavg)
         if (cpu_jobs < 1) cpu_jobs = 1
         jobs = nproc
         if (mem_jobs < jobs) jobs = mem_jobs
@@ -191,9 +200,18 @@ else
         if (jobs < 1) jobs = 1
         print jobs
     }')"
-    echo "=== --jobs not passed -- computed live default: nproc=$default_nproc available=${default_mem_mb}MB load=$default_load profile=$PROFILE factor=$default_process_factor -> --jobs $JOBS ==="
+    COMPILE_JOBS="$JOBS"
+    echo "=== --jobs not passed -- computed live default: nproc=$default_nproc available=${default_mem_mb}MB load=$default_load profile=$PROFILE per_job=${default_per_job_mb}MB -> compile -l $COMPILE_JOBS, test/go -j/-p $JOBS ==="
 fi
-MESON_COMPILE_JOB_OPTS=(-l "$JOBS")
+
+# meson compile gets ninja's -l LOAD-AVERAGE ceiling at COMPILE_JOBS, NOT a
+# flat -j: -l re-checks the live run-queue before each launch and backs off
+# on its own, so the compile path needs no pre-emptive discount for the
+# compiler driver's forked cc1/as children -- ninja's own live check IS that
+# safety. meson test / go test have no load-average equivalent, so they get
+# the PROCESS_FACTOR-discounted number as flat process-count caps.
+: "${COMPILE_JOBS:=$(awk -v j="$JOBS" -v f="${default_process_factor:-0.5}" 'BEGIN { d = int(j * f); if (d < 1) d = 1; print d }')}"
+MESON_COMPILE_JOB_OPTS=(-l "$COMPILE_JOBS")
 MESON_TEST_JOB_OPTS=(-j "$JOBS")
 GO_TEST_JOB_OPTS=(-p "$JOBS")