diff --git a/.claude/skills/xce-build/SKILL.md b/.claude/skills/xce-build/SKILL.md
index 8a0f5a40..7dc91eeb 100644
--- a/.claude/skills/xce-build/SKILL.md
+++ b/.claude/skills/xce-build/SKILL.md
@@ -21,7 +21,7 @@ Args: `[component] [profile|build_dir] [build_dir] [--incremental]` — passed a
    awk '/MemAvailable/ {print int($2/1024)}' /proc/meminfo   # available memory, MB (reclaimable-aware, unlike `free`'s raw "free" column)
    cut -d' ' -f1 /proc/loadavg                                # current 1-min load average = existing run-queue length from other activity
    ```
-2. Pick a per-job memory budget by profile — ASAN+debug is the heaviest (debug info + sanitizer instrumentation per translation unit): **~1536 MB/job for `asan`** (the default profile), **~1024 MB/job for `debug`/`release`**.
+2. Pick a per-job memory budget by profile — ASAN+debug is the heaviest (debug info + sanitizer instrumentation per translation unit): **~512 MB/job for `asan`** (the default profile; recalibrated in build/de-nix after the capnp compile-once and scheduler test-archive dedup dropped the measured per-job RSS ceiling to ~125-150MB — the old 1536 was calibrated around the 293MB-peak serial capnp compile that no longer exists), **~1024 MB/job for `debug`/`release`**.
 3. Compute, reserving ~15% of available memory as headroom for the OS and other sessions, and dampening the CPU term for how many *live OS processes* one "job" actually costs:
    - `mem_jobs = floor(available_mem_mb * 0.85 / per_job_mb)`
    - `PROCESS_FACTOR = 0.5` for `asan`/`debug` profiles, `0.7` for `release` — each in-flight ninja job is not one process: the compiler driver (`cc`/`c++`) forks a child compiler (`cc1`/`cc1plus`) and then an assembler (`as`), and `asan`/`debug` stretch out how long those extra children stay resident (sanitizer instrumentation, full debug-info emission) versus `release`'s leaner, shorter-lived ones. `nproc - current_load_avg` alone treats 1 job ≈ 1 process, so it under-throttles by roughly this factor — and `current_load_avg` is a smoothed 1-minute average, so it lags a fresh burst of forked children rather than pre-empting it.
diff --git a/src/tools/xce-build.sh b/src/tools/xce-build.sh
index fc77e3c3..63eb346b 100755
--- a/src/tools/xce-build.sh
+++ b/src/tools/xce-build.sh
@@ -157,6 +157,16 @@ else
     # 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).
+    #
+    # asan's 512MB budget (build/de-nix): the two memory hogs the original
+    # 1536MB figure was calibrated around are both gone -- the 293MB-peak
+    # serial capnp compile (批1: now compiled once, still far below the old
+    # budget) and the 17 scheduler test executables each statically linking
+    # all of scheduler_srcs (批2: shared sched_srcs_lib, link is the only
+    # heavy step and lld/ld peak ~125-150MB with -g1 debug info). Measured
+    # per-job RSS ceiling on t1 post-dedup is ~125-150MB; 512MB leaves ~3x
+    # headroom for a future pathological TU while letting an 8GB box run ~8
+    # concurrent compiles instead of 2-3.
     default_nproc="$(nproc --all 2>/dev/null || echo 1)"
     default_mem_mb="$(awk '/MemAvailable/ {print int($2/1024)}' /proc/meminfo 2>/dev/null)"
     default_load="$(cut -d' ' -f1 /proc/loadavg 2>/dev/null)"
@@ -164,7 +174,7 @@ else
     : "${default_load:=0}"
 
     case "$PROFILE" in
-        asan)    default_per_job_mb=1536; default_process_factor=0.5 ;;
+        asan)    default_per_job_mb=512;  default_process_factor=0.5 ;;
         debug)   default_per_job_mb=1024; default_process_factor=0.5 ;;
         release) default_per_job_mb=1024; default_process_factor=0.7 ;;
     esac