diff --git a/context/combat-damage.md b/context/combat-damage.md
index c0938ce..1edf23c 100644
--- a/context/combat-damage.md
+++ b/context/combat-damage.md
@@ -324,6 +324,34 @@ per ER-M laser; [[decomp-reference]] §5). Flying projectiles (LRM/autocannon) a
 reconstruction (`BTPushProjectile` — the 2007 Entity is too small for the binary's raw integrator
 offsets). [T2]
 
+## "Mechs die with lightly damaged armor" — AUDITED, no port bug: it's AMMO COOK-OFF (2026-08-02) [T1 chain, T2 measured]
+Field complaint (night 9): mechs die while their panels read nearly pristine, so "the scaling
+must be off". A byte-level audit of the ENTIRE damage chain found **every link faithful**:
+zone stream reads (arcade base ctor `@0041df5c` == engine DAMAGE.cpp, incl. the JointedMover
+skip-5 branch); the BT ctor normalization loop (`@0049ce50` constants 1.0/1e-4/0.5 — and in
+SHIPPED CONTENT the raw cells are EXACTLY `1/armorPts` uniform across all 5 types, so the
+normalize is a no-op); `TakeDamage @0041e4e0` formula+clamps; the override `@0049c690`
+(redirect gate, vital/leg thresholds 1.0/0.5/0.0 `@0049c99c..a4`); `CriticalHit @0049ccc4`
+(fraction 0.5 `@0049ce48`, cap 1.0 `@0049ce4c`, the per-subsystem `applied<damagePercentage`
+cap); the cook-off chain (`stamp @004bc3fc` per-SALVO before the missile divide, `CookOff
+@004bd300` = ammoCount×amount, `DistributeCriticalHit @004ac274` = ÷(zones plugging the bin)
+then one DESIGNATED-zone TakeDamageMessage each).
+**The measured kill route** (death-kinetics bench, 7 deaths): 6 of 7 were **bay-fire
+detonations** — e.g. `AmmoBinLRM15 10 rounds × 50 = 500 (burst 1)` into the bin's ONE authored
+carrier zone (madcat bins are single-zone: LRM15_1=roster[26]→dz_ltorso, LRM15_2=[28]→dz_rtorso,
+AFC100=[21]→{larm,lgun}), which is VITAL (madcat authors **15 of 22 zones vital** — all torsos,
+legs, feet) → one application 0→1.0 → instant death, every other panel bright. The ÷1 is the
+CONTENT's authoring, not a port shortcut — the arithmetic is the binary's own. The one mech that
+died conventionally (300 applications) ended VISIBLY charred: 0.97/0.93/0.92/0.85/0.79 across 13
+zones — mid-fight darkening DOES accumulate when mechs live long enough.
+**Field logs confirm dominance**: Rajel's night-9 log alone has 12 bay-fire events (LRM10
+455-490 pts); ConnMan 10; the "salvo barely dented my armor" reporter was dying to their OWN
+cook-offs. **The real lever is HEAT** (bay fire arms at the bin heat-watcher's FAILURE level;
+the [[heat-and-coolant]]/#96 too-hot economy arms it constantly) plus player training: EJECT
+**hold** purges the burning bay (`DumpAmmo` → "bay fire EXTINGUISHED"); pod players were
+manual-trained to purge ([[pod-hardware]] §Manual), testers are not. Novice sim gates also
+suppress the whole path ([[experience-levels]]).
+
 ## The RAM economy — CLOSED (2026-07-12) [T1 evidence, T3 normalization]
 Three-layer story, measured live + decomp-verified:
 1. **The armor economy is POINTS**: every zone streams `damageScale[5]` — 5 cells indexed by
diff --git a/context/decomp-reference.md b/context/decomp-reference.md
index a58f94e..f9a40ec 100644
--- a/context/decomp-reference.md
+++ b/context/decomp-reference.md
@@ -500,7 +500,9 @@ ex-"clipLoadGuard").  Byte tooling: `scratchpad/dis_4b2980.py` / `dis_range.py`
 `reference/decomp/section_dump.txt`.
 
 **ARMOUR-DARKENING addresses** [T1] (#87; full story: [[rendering]]): `.DZM` res compiler
-`FUN_0041e4e0` (part_002.c:7341, reads the `material` key list per `[dz_*]` section); runtime
+address UNKNOWN — the old citation `FUN_0041e4e0` was WRONG (that function, part_002.c:7337,
+is `DamageZone::TakeDamage`: `level += amount×scale[type]`, byte-matches DAMAGE.cpp:379 —
+verified 2026-08-02, swept here + [[rendering]]); runtime
 consumer inside `BTL4VideoRenderer::MakeMechRenderables` **`FUN_004cef28`** (part_014.c:5190-5272);
 per-zone node ctor `FUN_00455c7c(...,0.04f)`; the per-material watcher
 **`FUN_004573e4(material, &zone->damageLevel, 0.1f)`** (part_007.c:9133, vtable `PTR_LAB_004f1730`,
diff --git a/context/rendering.md b/context/rendering.md
index e12433e..98a26e0 100644
--- a/context/rendering.md
+++ b/context/rendering.md
@@ -335,7 +335,9 @@ the limb swap ("the actual enemy mech in external view is not showing darkened a
 - **`.DZM`** (`VIDEO\<mech>SKIN.DZM`, 40 shipped) lists, per damage zone, the materials that zone
   paints: `[dz_ltorso] material=avaskin:avat2_dz_ltorso_mtl`. One file per SKIN VARIANT — the
   `dzm`/`dzms`/`dzmt`/`dzmo`/`dzma`/`dzmb`/`dzmc`/`dzmd` key family (same suffix set as the
-  `destroyed`…`destroyedd` video variants); the res compiler is `FUN_0041e4e0`.
+  `destroyed`…`destroyedd` video variants); the res-compiler address is UNKNOWN (the old
+  `FUN_0041e4e0` citation was WRONG — that is `DamageZone::TakeDamage`; corrected 2026-08-02,
+  swept with [[decomp-reference]]).
 - Compiled into BTL4.RES and parsed by **MUNGA's own `DamageZone` stream ctor** into
   `materialTable`, keyed by `EntitySegment::SkeletonType` (`engine/MUNGA/DAMAGE.cpp:311-336`,
   `GetMaterialList(skl_type)`) — real engine source [T0]. **This data was loaded and unused in the