feat(perf): RAPL package-power caps — watts are the thermal lever, not percent

Discovery 2026-07-19: the Dell BIOS ships PL1=PL2=135W on this 45W-TDP
i7-10875H. With no sustained-power brake, the only thing that ever slows
the CPU is the 100C thermal limiter — which is how a set peaked at 91C and
had to be aborted. Percent clock caps (--cool's max_perf_pct) don't bound
multi-core power; watt caps do.

All three modes now set package power limits via RAPL, saved/restored by
--stop:
- --optimize / --extreme: PL1 45W (chip spec) / PL2 90W — full nominal
  performance, bursts stay snappy, sustained runaway impossible.
- --cool: PL1 28W / PL2 50W by default (~2.5GHz all-core), tunable with
  COOL_PL1_W / COOL_PL2_W.
- --diagnose now prints the current PL1/PL2 against spec.

perf.sh

@@ -11,6 +11,15 @@ SAVED_MAX_PCT="/tmp/saved_cpu_max_pct.txt"
 # the editor's bursty UI work to clear quickly. Tune with COOL_MAX_PCT=NN
 # (lower = cooler but laggier UI; higher = snappier but warmer).
 COOL_MAX_PCT="${COOL_MAX_PCT:-85}"
+# Package power caps in watts (PL1 = sustained, PL2 = burst). The Dell BIOS
+# ships PL1=PL2=135W on this 45W-TDP i7-10875H, so nothing but the 100°C
+# thermal limiter ever brakes the CPU (measured 91°C mid-set 2026-07-18).
+# Watts bound heat directly; percent clock caps don't on multi-core load.
+COOL_PL1_W="${COOL_PL1_W:-28}"
+COOL_PL2_W="${COOL_PL2_W:-50}"
+SAVED_PL1="/tmp/saved_rapl_pl1.txt"
+SAVED_PL2="/tmp/saved_rapl_pl2.txt"
+RAPL_DIR="/sys/class/powercap/intel-rapl:0"
 
 # Display help
 show_help() {
@@ -91,6 +100,11 @@ run_diagnostics() {
     echo "Min performance percentage:"
     cat /sys/devices/system/cpu/intel_pstate/min_perf_pct 2>/dev/null
   fi
+
+  if [ -d "$RAPL_DIR" ]; then
+    echo "Package power limits (spec for i7-10875H: PL1 45W):"
+    echo "PL1 $(( $(cat $RAPL_DIR/constraint_0_power_limit_uw) / 1000000 ))W sustained / PL2 $(( $(cat $RAPL_DIR/constraint_1_power_limit_uw) / 1000000 ))W burst"
+  fi
   
   # Check for excessive disk I/O
   echo -e "\n=== DISK I/O STATUS ==="
@@ -171,6 +185,17 @@ run_diagnostics() {
   echo "4. Run 'sudo $0 --extreme' before performance to apply maximum optimizations"
 }
 
+# Cap package power via RAPL (PL1 sustained / PL2 burst, in watts). Saves the
+# current values once so --stop can restore them.
+set_power_caps() {
+  [ -d "$RAPL_DIR" ] || return 0
+  [ -f "$SAVED_PL1" ] || cat $RAPL_DIR/constraint_0_power_limit_uw > $SAVED_PL1 2>/dev/null
+  [ -f "$SAVED_PL2" ] || cat $RAPL_DIR/constraint_1_power_limit_uw > $SAVED_PL2 2>/dev/null
+  echo $(($1 * 1000000)) > $RAPL_DIR/constraint_0_power_limit_uw 2>/dev/null
+  echo $(($2 * 1000000)) > $RAPL_DIR/constraint_1_power_limit_uw 2>/dev/null
+  echo "✓ Package power capped: PL1 ${1}W sustained / PL2 ${2}W burst"
+}
+
 # Function to apply standard performance optimizations
 optimize_standard() {
   check_root "--optimize"
@@ -204,24 +229,28 @@ optimize_standard() {
     echo "✓ CPU Turbo enabled"
   fi
   
+  # Even at full performance, sustained power must stay at the 45W spec —
+  # the BIOS's 135W setting is what produced the 91°C thermal runaway.
+  set_power_caps 45 90
+
   # Apply IRQ thread priorities
   echo "Setting IRQ thread priorities..."
   for irq in $(find /proc/irq/*/smp_affinity -type f 2>/dev/null | grep -v self); do
     irq_num=$(echo $irq | cut -d '/' -f 4)
     name=$(cat /proc/irq/$irq_num/name 2>/dev/null)
-    
+
     # Check if this is an audio-related IRQ
     if [[ "$name" == *"snd"* ]] || [[ "$name" == *"audio"* ]] || [[ "$name" == *"hda"* ]]; then
       echo "Setting audio IRQ $irq_num ($name) affinity"
       echo 1 > /proc/irq/$irq_num/smp_affinity 2>/dev/null
     fi
   done
-  
+
   # Reduce kernel printk level to minimize logging impact
   echo "Reducing kernel logging level..."
   echo 1 > /proc/sys/kernel/printk
   echo "✓ Kernel logging reduced"
-  
+
   # Set process priorities with standard mode
   set_priorities standard
   
@@ -283,7 +312,12 @@ optimize_extreme() {
     echo 100 > /sys/devices/system/cpu/intel_pstate/min_perf_pct 2>/dev/null
     echo "✓ Intel CPU min/max performance pinned to 100% (no downclocking, turbo reachable)"
   fi
-  
+
+  # Extreme pins clocks but must still bound sustained power at the 45W spec:
+  # with the BIOS's 135W limit the only brake left is the 100°C thermal
+  # limiter (measured 91°C mid-set 2026-07-18). PL2 90W keeps bursts snappy.
+  set_power_caps 45 90
+
   # Disable NMI watchdog (can cause stutters)
   echo "Disabling NMI watchdog..."
   echo 0 > /proc/sys/kernel/nmi_watchdog
@@ -378,6 +412,10 @@ optimize_cool() {
     echo "✓ Peak clock capped to ${COOL_MAX_PCT}% (turbo bounded, idle stays low)"
   fi
 
+  # The real heat lever: bound sustained package power in watts. Tune with
+  # COOL_PL1_W=NN (lower = cooler; 28W ≈ all-core ~2.5GHz on this chip).
+  set_power_caps "$COOL_PL1_W" "$COOL_PL2_W"
+
   # Light VM tuning: avoid swap-driven stalls during a set without forcing anything.
   echo 10 > /proc/sys/vm/swappiness 2>/dev/null
   echo "✓ Swappiness lowered to 10"
@@ -638,6 +676,13 @@ reset_system() {
     rm -f "$SAVED_EPP"
     echo "✓ EPP restored"
   fi
+
+  # Restore RAPL package power limits if any mode capped them
+  if [ -d "$RAPL_DIR" ]; then
+    [ -f "$SAVED_PL1" ] && cat "$SAVED_PL1" > $RAPL_DIR/constraint_0_power_limit_uw 2>/dev/null && rm -f "$SAVED_PL1"
+    [ -f "$SAVED_PL2" ] && cat "$SAVED_PL2" > $RAPL_DIR/constraint_1_power_limit_uw 2>/dev/null && rm -f "$SAVED_PL2"
+    echo "✓ RAPL power limits restored"
+  fi
   
   # Restore NMI watchdog
   echo 1 > /proc/sys/kernel/nmi_watchdog