diff --git a/context/reconstruction-gotchas.md b/context/reconstruction-gotchas.md
index 3755196..316631a 100644
--- a/context/reconstruction-gotchas.md
+++ b/context/reconstruction-gotchas.md
@@ -795,3 +795,37 @@ grep its helper types for `return 0` bodies FIRST; (b) a bring-up shim must
 either Fail() loudly or log its own emptiness, never silently iterate nothing;
 (c) before shimming an engine-side structure, check whether the engine already
 has it -- the offsets in the decomp comment ARE the lookup key.
+
+## 27. Drawing where the design says DON'T: under-surface buttons, shared planes, and inherited D3D state (#118, 2026-08-03)
+
+Three MFD-corruption incidents in one evening, all mine, all the same class:
+adding a DRAW to a compositor whose layout invariants I had not read.
+
+1. **The eject-wipe gauge** (BitMapInverseWipeScalar) drew into a weapon's ENG
+   graphics port -- but the Eng1-3 planes SHARE composite dest cells with the
+   weapon pages (`ckSlotOf` in L4VB16), so its mis-geometry scribbled striped
+   garbage OVER the weapon MFDs whenever an eject cycle redrew it (= at
+   overheat-jam onset, which the operator correctly correlated with "when the
+   coolant warning starts").
+2. **The "flash overlay"** re-drew flashing buttons' full faces ON TOP of the
+   surfaces.  But the side buttons are BIG rects deliberately tucked UNDER the
+   displays -- the protruding edge IS the lamp (the pod's backlit keys sat
+   beside the CRT; there is no in-display face).  A full-face redraw is wrong
+   BY DESIGN: it covers gauge imagery with button faces.
+3. **The same overlay, drawn after `DrawDevSurface`, inherited its bound
+   texture** -- an un-reset `CkFill` renders a TEXTURED quad sampling the
+   gauge atlas: green diagonal-striped rectangles.  (With the state reset the
+   bug became honest: solid flashing faces covering the MFD -- which is how
+   the operator's "the panel buttons took over the MFD" report finally
+   identified the author.)
+
+Rules: (a) before ADDING a draw to an existing render pass, read the pass's
+layering comments -- "UNDER the surfaces", "masks the part inside", "shares
+its sibling's dest cell" are INVARIANTS, not descriptions; (b) any D3D draw
+inserted after another subsystem's draws must re-assert its own texture/color
+state (CkColorState-equivalent) -- the previous draw's bindings are live;
+(c) when a visual corruption tracks an EVENT ("starts with the coolant
+warning"), list what STARTS DRAWING at that event -- alarm-driven redraws,
+state-change repaints -- before suspecting the event's logic; (d) the
+operator's screenshot is worth ten theories: the red-faces capture identified
+in one frame what three log-side hypotheses missed.