diff --git a/context/decomp-reference.md b/context/decomp-reference.md
index d06c4ce..381be70 100644
--- a/context/decomp-reference.md
+++ b/context/decomp-reference.md
@@ -522,6 +522,7 @@ default-ON (`'0'` disables).
 | `BT_SECTOR_LOG` | SectorDisplay Make/Execute |
 | `BT_NET_TRACE` | `[net-tx]/[net-rx]/[net-upd]` MP tracing |
 | `BT_DEV_GAUGES` | render the 6 pod MFD surfaces in a separate dev window |
+| `BT_GLASS_LAMP_SWEEP` | `0` = restore authentic once-per-gauge-cycle cockpit-lamp updates; default (on) sweeps the lamps EVERY frame in the dev/glass composite path so the lit buttons track the sim even when the throttled background gauge sweep starves under MP load (`L4VB16.cpp` `BTGlassSweepLamps`; the "lighting slow/nonexistent" fix — glass-only, pod serial cadence untouched) |
 | `BT_AIM="x y"` | pin the reticle crosshair (reticle coords) — headless aim harness |
 | `BT_AIM_LOG` | `[pick]` ray/box/hit diagnostics (Mech::PickRayHit) |
 | `BT_FIRE_ARC=<deg>` | OPT-IN external-camera fire-arc clamp (unset = authentic no-arc) |
diff --git a/context/gauges-hud.md b/context/gauges-hud.md
index aaba082..cb1902b 100644
--- a/context/gauges-hud.md
+++ b/context/gauges-hud.md
@@ -111,6 +111,21 @@ viewport), and `gWindowAspect` = the view rect's on-screen aspect (`BTWorldAspec
   the work area). Green tint tunable via `BT_COCKPIT_TINT=RRGGBB` (default `0x27E8`). Labels are a
   lazy GDI-baked MANAGED atlas (survives device reset). Renders in ALL builds; only the PadRIO
   click/lamp seam is BT_GLASS-gated. Full detail: `docs/GAUGE_COMPOSITE.md`.
+- ⚠ **Cockpit-lamp LATENCY under MP load (2026-08-04, glass playtest fix) [T2 runtime-verified].**
+  The lit cockpit buttons are drawn every frame (`BTDrawCockpitPanels`, reading `PadRIO::GetLampState`),
+  but the lamp STATE sweep that fills that store — `LampManager::Update` → `AssertNewLampValue` →
+  `SetLamp` — authentically rides `GaugeRenderer::ExecuteForeground`, which fires only **once per full
+  gauge cycle** (foreground→background→copy). The cycle can't advance to the next foreground turn
+  until the THROTTLED background gauge sweep drains the ~140-instrument active list, and that
+  background task starves under load (issue #45 — "instruments freeze while fps stays healthy"). So on
+  a busy MP mission the lamp sweep ran ~1×/s: **buttons froze / flashes stalled while the 3D view (a
+  separate per-frame foreground render) stayed smooth** — the "lighting slow or nonexistent" reports
+  (all testers on glass surround). Fix: `BTGlassSweepLamps()` runs the lamp sweep EVERY frame on the
+  dev/glass composite path (`L4VB16.cpp`, in `BTDrawGaugeInset`) — cheap (deduped pushes, ~72 lamps),
+  decoupled from the gauge cycle. The pod never enters this path (real gauge hardware), so its
+  bandwidth-paced serial lamp cadence is untouched. Kill switch `BT_GLASS_LAMP_SWEEP=0`. NB the RIO
+  serial/lamp stack itself is byte-identical to pod-proven Red Planet (`L4LAMP.cpp`/`L4SERIAL.cpp`
+  diffed) — the defect was purely the glass update CADENCE, not the lamp wiring.
 
 ## MP gauge-window FREEZE = dangling bindings + permanent SEH disable (Gitea #12, 2026-07-19) [T2 log-convicted]
 The live-MP "dev-gauges window froze entirely mid-session" incident (issue #12) is NOT a
diff --git a/engine/MUNGA_L4/L4VB16.cpp b/engine/MUNGA_L4/L4VB16.cpp
index a14ef9f..05fb0a2 100644
--- a/engine/MUNGA_L4/L4VB16.cpp
+++ b/engine/MUNGA_L4/L4VB16.cpp
@@ -1293,6 +1293,47 @@ void BTCockpitMouseUp(void)
 	}
 }
 
+//
+// (glass lamp responsiveness, 2026-08) Sweep the cockpit lamp STATE every frame.
+//
+// The lamp sweep -- LampManager::Update -> AssertNewLampValue -> RIO/PadRIO::SetLamp,
+// which is what lights the cockpit buttons -- authentically rides the gauge
+// renderer's FOREGROUND turn (GaugeRenderer::ExecuteForeground, gaugrend.cpp), and
+// that turn fires only ONCE PER FULL GAUGE CYCLE (foreground -> background -> copy).
+// The cycle only advances to the next foreground turn after the THROTTLED background
+// gauge sweep drains the whole active-instrument list -- and that background task
+// starves under load (issue #45: "every cockpit instrument freezes at its last paint
+// while the underlying values AND fps stay perfectly healthy").  So on a busy MP
+// mission the lamp sweep ran ~1x/second: the lit buttons froze / flashes stalled while
+// the 3D view (a separate per-frame foreground render) stayed smooth -- exactly the
+// "lighting slow or nonexistent" playtest reports (all on glass desktops).
+//
+// The lamp sweep is CHEAP -- deduped state pushes over ~72 lamps, no raster -- so on
+// the dev/glass composite path we run it EVERY frame here, decoupled from the gauge
+// cycle, and the buttons track the sim regardless of background-sweep slack.  The pod
+// NEVER enters BTDrawGaugeInset (it drives real gauge hardware), so its authentic
+// bandwidth-paced serial lamp cadence is untouched.  BT_GLASS_LAMP_SWEEP=0 restores the
+// authentic once-per-cycle behaviour.
+//
+static void
+	BTGlassSweepLamps()
+{
+	static int s_on = -1;
+	if (s_on < 0)
+	{
+		const char *e = getenv("BT_GLASS_LAMP_SWEEP");
+		s_on = (e != NULL && e[0] == '0') ? 0 : 1;
+	}
+	if (!s_on) return;
+
+	GaugeRenderer *gr = BTResolveGaugeRenderer();
+	if (gr == NULL || application == NULL) return;
+	LampManager *lm = gr->GetLampManager();
+	ModeManager *mm = application->GetModeManager();
+	if (lm != NULL && mm != NULL)
+		lm->Update(mm->GetModeMask());
+}
+
 //
 // Free entry point the MAIN renderer calls (L4VIDEO DPLRenderer::ExecuteImplementation,
 // just before EndScene).  No-op unless BT_DEV_GAUGES is set.  Docks the 6-surface panel
@@ -1301,7 +1342,9 @@ void BTCockpitMouseUp(void)
 //
 void BTDrawGaugeInset(LPDIRECT3DDEVICE9 device)
 {
-	if (!DevGaugeComposite() || !DevGaugeDocked() || device == NULL) return;
+	if (!DevGaugeComposite()) return;
+	BTGlassSweepLamps();	// keep the cockpit lamps tracking the sim EVERY frame (see above)
+	if (!DevGaugeDocked() || device == NULL) return;
 	if (GlassPanelsOwnSurfaces()) return;	// the per-display glass windows draw the surfaces
 	if (gBTGaugeCockpit)
 	{