diff --git a/context/reconstruction-gotchas.md b/context/reconstruction-gotchas.md
index dc11e27..51ca2b7 100644
--- a/context/reconstruction-gotchas.md
+++ b/context/reconstruction-gotchas.md
@@ -433,6 +433,53 @@ VehicleDead handler's disasm before changing). [T2]
 
 ---
 
+## 21. Port fixed-array too small for the real data → heap overflow (the DEATH-CRASH, Gitea #12)
+
+**Symptom class (Gitea #12 death-crash):** a SILENT hard crash a few seconds AFTER a mech DEATH
+(no WER, no crash markers) — solo on a killed dummy, on the local player's own MP death, and on a
+surviving peer when a player disconnects (PEER_DOWN). Combat runs fine for minutes; the crash is
+deterministic *on the first kill*. Under cdb the stack is always in the **audio-entity teardown**
+of the death Explosion: `FryDeathRow → Explosion::~Explosion → Entity::~Entity →
+…NotifyOfEntityDestruction → RendererOrigin::RemoveUninterestingEntity →
+L4AudioRenderer::NotifyOfBecomingUninterestingEntity → DestroyEntityAudioObjects →
+SocketIterator::DeletePlugs → ~Dynamic3DPatchSource → operator delete` = **HEAP CORRUPTION
+DETECTED after Normal block** (Debug); or a garbage vtable dispatch in `AudioControlSequence::
+RunSequence → AudioControlEvent::Send` (`call [eax+0x1c]`, vtable == 0x16) (Release). Both are the
+same corruption manifesting at the next free / next virtual call.
+
+**Cause:** a WinTesla-port fixed array sized for an assumption the real gamedata violates.
+`L4AUDHDW.h SourceSet { int count; ALuint sources[5]; }` — the OpenAL source array was **5** wide
+(an "AWE 4-voice" assumption), but `channelSet.count = GetAudioVoiceCount()` = the streamed
+per-preset `voiceCount`, and the **AllExplosion preset (bank2 p125 — the death boom) authors 25
+layered zones** (`MAX_PRESET_SAMPLES = 25`, L4AUDLVL.h, one OpenAL source per zone). On channel
+acquisition `RequestAudioChannels` runs `for (i<count) alGenSources(1, sources+i)` → writes
+`sources[5..24]` **past the end of the 5-element array**, clobbering the adjacent heap object's
+vtable / the debug heap guard bytes. Nothing faults until that neighbour is next *used* (the
+running sequence's virtual `Send` — Release) or *freed* (the death Explosion's audio teardown —
+Debug), seconds later. The `Playing hit an error: 40961` (`AL_INVALID_NAME`) burst right after the
+wreck swap is the same corruption (the source-id array trashed). This is NOT the held #16
+DPLEyeRenderable/boresight work (exonerated — never on the stack), and NOT the `rev=-64` mppr
+value (a `ControlsButton` is "negative for release" by design, CONTROLS.h:26 — benign; the
+`< 1` test at mechmppr.cpp:870 reads it correctly).
+
+**Fix (2026-07-20):** size `SourceSet.sources[]` to `AUDIO_SOURCESET_CAPACITY = 25`
+(== `MAX_PRESET_SAMPLES`; L4AUDLVL.h `static_assert`s the two in lockstep), so the largest
+legitimate preset fits. Plus defence-in-depth clamps (a malformed stream can't overflow):
+`channelSet.count` in the `L4AudioSource` ctor and `requested` in `RequestAudioChannels`. The
+teardown path is generic to ANY entity leaving the interest set, so the one fix covers all three
+#12 death flavors (solo kill, MP self-death, MP peer PEER_DOWN — all fire the same 25-voice
+explosion preset through the same teardown). Files: `engine/MUNGA_L4/L4AUDHDW.h`,
+`L4AUDLVL.h`, `L4AUDIO.cpp`, `L4AUDRND.cpp`. [T2 — solo stack-confirmed + fix survives 10+ kills
+under cdb; MP flavors explained-by-mechanism, live-MP repro pending]
+
+**Rule:** a port fixed-size array fed from streamed gamedata must be sized to the DATA's real
+maximum (dump it — here the `MAX_PRESET_SAMPLES` comment already named the 25-zone worst case),
+not a guessed hardware limit; and index loops driven by a streamed `count` must clamp to capacity.
+Grep the port for sibling `[N]` arrays filled from a resource-derived count (channel/voice/zone/
+segment tables) — the same class of latent overflow.
+
+---
+
 ## Diagnostic recipe (the standard loop)
 1. Read the RAW decomp `reference/decomp/all/part_*.c` for the `FUN_xxxx`.
 2. Map `FUN_`/`DAT_`/`this+0xNN` to engine symbols via BT headers + WinTesla MUNGA source +
diff --git a/engine/MUNGA_L4/L4AUDHDW.h b/engine/MUNGA_L4/L4AUDHDW.h
index 5ee0f38..fb2155c 100644
--- a/engine/MUNGA_L4/L4AUDHDW.h
+++ b/engine/MUNGA_L4/L4AUDHDW.h
@@ -6,10 +6,24 @@
 #include "..\munga\audio.h"
 #include "openal/al.h"
 
+// SourceSet capacity (Gitea #12 death-crash fix): the OpenAL sources array
+// must hold as many voices as a preset can carry.  channelSet.count is set
+// from GetAudioVoiceCount() (the streamed voiceCount), and a full-zone preset
+// legitimately carries up to MAX_PRESET_SAMPLES (== 25, L4AUDLVL.h) sample
+// zones -- one OpenAL source each (RequestAudioChannels alGenSources into
+// sources[0..count-1], SetupPatch/PlayNote index sources[i] parallel to the
+// samples[] zones).  This array was 5 wide, so the death-boom preset (>5
+// voices) overwrote adjacent heap on channel acquisition -- corrupting the
+// next AudioComponent's vtable / the heap guard bytes and crashing seconds
+// later at the death Explosion's audio-entity teardown (DestroyEntityAudio-
+// Objects) or the next sequenced virtual dispatch.  L4AUDLVL.h static_asserts
+// MAX_PRESET_SAMPLES <= AUDIO_SOURCESET_CAPACITY to keep the two in lockstep.
+const int AUDIO_SOURCESET_CAPACITY = 25;
+
 struct SourceSet
 {
 	int count;
-	ALuint sources[5];
+	ALuint sources[AUDIO_SOURCESET_CAPACITY];
 };
 
 //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ MIDI types ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
diff --git a/engine/MUNGA_L4/L4AUDIO.cpp b/engine/MUNGA_L4/L4AUDIO.cpp
index 942dc41..15fd021 100644
--- a/engine/MUNGA_L4/L4AUDIO.cpp
+++ b/engine/MUNGA_L4/L4AUDIO.cpp
@@ -672,7 +672,20 @@ L4AudioSource::L4AudioSource(
 ):
 	AudioSource(stream, entity)
 {
-	channelSet.count = GetAudioVoiceCount();
+	// Gitea #12: never let the voice count exceed the SourceSet capacity -- a
+	// malformed/oversized preset stream would otherwise overflow sources[] on
+	// channel acquisition and corrupt the heap (the death-crash class).  The
+	// array is now sized to the largest legitimate preset (AUDIO_SOURCESET_
+	// CAPACITY == MAX_PRESET_SAMPLES), so this clamp is defence-in-depth.
+	int voices = GetAudioVoiceCount();
+	if (voices > AUDIO_SOURCESET_CAPACITY)
+	{
+		DEBUG_STREAM << "[audio] WARNING: voiceCount " << voices
+			<< " exceeds SourceSet capacity " << AUDIO_SOURCESET_CAPACITY
+			<< " -- clamped (Gitea #12)\n" << std::flush;
+		voices = AUDIO_SOURCESET_CAPACITY;
+	}
+	channelSet.count = voices;
 	L4AudioSourceX();
 }
 
diff --git a/engine/MUNGA_L4/L4AUDLVL.h b/engine/MUNGA_L4/L4AUDLVL.h
index 934aea8..d7a9df8 100644
--- a/engine/MUNGA_L4/L4AUDLVL.h
+++ b/engine/MUNGA_L4/L4AUDLVL.h
@@ -40,6 +40,13 @@ struct SAMPLEINFO
 // AllExplosion (bank2 p125) authors 25 layered zones -- the largest preset.
 const int MAX_PRESET_SAMPLES = 25;
 
+// Gitea #12: the SourceSet OpenAL-source array must hold one source per zone,
+// so its capacity must cover the largest preset.  If MAX_PRESET_SAMPLES ever
+// grows past the SourceSet array, channel acquisition would overflow the heap
+// (the death-crash class) -- lock the two together at compile time.
+static_assert(MAX_PRESET_SAMPLES <= AUDIO_SOURCESET_CAPACITY,
+	"SourceSet.sources[] too small for MAX_PRESET_SAMPLES voices (Gitea #12)");
+
 struct PRESETINFO
 {
 	int sampleNum;
diff --git a/engine/MUNGA_L4/L4AUDRND.cpp b/engine/MUNGA_L4/L4AUDRND.cpp
index 8077135..8e3415b 100644
--- a/engine/MUNGA_L4/L4AUDRND.cpp
+++ b/engine/MUNGA_L4/L4AUDRND.cpp
@@ -1294,6 +1294,13 @@ Logical
 	//Do we have enough?
 	int requested = source_request->count;
 
+	// Gitea #12: hard cap at the SourceSet capacity so alGenSources can never
+	// write past sources[] (the death-crash heap overflow).  count is set from
+	// the streamed voiceCount and is normally already clamped in the
+	// L4AudioSource ctor; this guards the acquisition site itself.
+	if (requested > AUDIO_SOURCESET_CAPACITY)
+		requested = AUDIO_SOURCESET_CAPACITY;
+
 	bool failed = true;
 
 	alGetError();