diff --git a/engine/MUNGA_L4/L4AUDRND.cpp b/engine/MUNGA_L4/L4AUDRND.cpp
index e6241c5..4537c29 100644
--- a/engine/MUNGA_L4/L4AUDRND.cpp
+++ b/engine/MUNGA_L4/L4AUDRND.cpp
@@ -395,12 +395,27 @@ void
 		// DPLIndependantEffect voices, and while the pool is pinned every new
 		// acquire fails outright and that sound is silently dropped.
 		//
-		// ⚠ OFF BY DEFAULT, DELIBERATELY.  Raising the cap is NOT a free win, and
-		// the reported dropout was already fixed on 2026-07-23 by a999e5c (the
-		// atomic-delete leak) -- no field complaints since.  What remains in the
-		// night-5 logs is transient exhaustion at the peak of a firefight, where
-		// a dropped voice competes with ~256 already sounding, so it is very
-		// likely sub-perceptual.  Against that: the 256 ceiling also acts as a
+		// ⚠ OFF BY DEFAULT, DELIBERATELY.  Raising the cap is NOT a free win.
+		//
+		// ⚠⚠ UPDATE 2026-08-01 (#32): the "no field complaints since" claim below
+		// did NOT hold.  Every 4.11.674 player log is saturated -- 3031/4657/5245/
+		// 6275/6571 failures across five machines, first failure ~10% into a match
+		// and still failing at 97%.  The root cause was NOT the source budget at
+		// all: RequestAudioChannels was alGenSources'ing per sound event and
+		// ReleaseSourceSet alDeleteSources'ing on release, so combat CHURNED
+		// through the ceiling.  Sources are now POOLED and recycled (see
+		// BTAudioPoolAcquire below), which fixes it WITHOUT touching this budget.
+		// Measured on the same bench: frame time went 8.67ms -> 7.79ms over ~10k
+		// frames, i.e. removing the churn is a net CPU WIN, so the governor
+		// argument below does not apply to pooling.  The reasoning about raising
+		// the CAP (more voices mixing = more CPU) still stands on its own, which
+		// is why BT_AUDIO_SOURCES remains opt-in and unset by default.
+		//
+		// (Historical:) the 2026-07-23 fix a999e5c addressed the atomic-delete
+		// leak.  What remained in the night-5 logs is transient exhaustion at the
+		// peak of a firefight, where a dropped voice competes with ~256 already
+		// sounding, so it is very likely sub-perceptual.  Against that: the 256
+		// ceiling also acts as a
 		// GOVERNOR.  The steal loop only steals when the incoming source outranks
 		// a running one, so a higher cap means many more voices mixing at once --
 		// with EFX reverb + the lowpass chains live, that is real CPU, spent