#119: heatLoad filter restored to the POD's 28 Hz sample cadence

The clipping coolant-leak voice ("warning... war... warning coo...")
measured to a ReportLeak relaxation oscillation at a drained tank:
draw = zoneDamage x heatLoad hunts across the authored 0.0025/0.003
hysteresis band (constants byte-verified) as the dry sink's conduction
collapses and the tank trickle refills it. State machine, constants,
refill dynamics and the sequencer's chase-and-cut stop (T0 AUDSEQ.cpp)
are all the binary's own -- the ONE divergence was the 15-sample
heatLoad filter: per-Perform in the binary = a 0.536 s window at the
pod's 28 Hz, but only 0.25 s at the port's ~59 fps -- heatLoad twice as
twitchy, the flap at double the pod cadence (the myomer-kinetic dt-less
class, third instance).

UpdateHeatLoad now accrues real time and samples at 28 Hz on any
machine (catch-up capped at the 15-sample window). Per-instance clock
in a static map -- the heat-family layouts are factory-size-locked
(sizeof(Myomers) == 0x358 exact), no new members. Ctor/reset prime
calls sample unconditionally.

A/B (260 s leak-to-empty bench): trigger fires 12 -> 8, transitions
11 -> 7. The residual slow restart at a bone-dry tank is authentic 1995
behavior (same engine sequencer chase-cut); a minimum-retrigger
interval would be an opt-in deviation for the operator to decide.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

game/reconstructed/heat.cpp

@@ -777,7 +777,7 @@ void
 	{
 		heatEnergy         += pendingHeat;
 		currentTemperature  = heatEnergy / thermalMass;
-		UpdateHeatLoad();
+		UpdateHeatLoad(time_slice);						// 28 Hz filter cadence (#119)
 
 		// DIAG census (BT_HEAT_LOG, viewpoint mech): PER-INSTANCE 5-s timers --
 		// the old shared static timer aliased to whichever instance crossed the
@@ -850,8 +850,22 @@ void
 // @004ad7f0 -- recompute the radiated/instantaneous heat and feed it through
 // the running-average filter to produce the smoothed heatLoad reading.
 //
+// POD-CADENCE SAMPLING (#119, 2026-08-02).  The binary adds ONE sample per
+// Perform -- a dt-less per-frame term (the myomer-kinetic class, FUN_0041c018
+// proves Performs ride the frame rate).  15 samples at the pod's 28 Hz = a
+// 0.536 s smoothing window; sampled at the port's ~59 fps the window halves
+// and heatLoad turns 2x as twitchy -- which doubles the cadence of the
+// ReportLeak relaxation oscillation at a drained tank (draw = damage x
+// heatLoad hunting across the authored 0.0025/0.003 hysteresis band), and
+// THAT is the rapidly-clipping "warning coo-- war--" leak voice Oracle
+// reproduced twice in 716.  Sim calls pass their slice and samples accrue at
+// 28 Hz on any machine; ctor/reset prime calls (negative dt) sample
+// unconditionally.  Per-instance clock lives in a static map (the census
+// precedent above) because the heat-family layouts are factory-size-locked
+// (sizeof(Myomers) == 0x358 exact) -- no new members.
+//
 void
-	HeatSink::UpdateHeatLoad()
+	HeatSink::UpdateHeatLoad(Scalar time_slice)
 {
 	radiatedHeat = currentTemperature * coolantLevel;
 
@@ -865,7 +879,31 @@ void
 		sample = HeatLoadMaximum;
 	}
 
-	heatFilter.AddSample(sample);		// FUN_0043ade4
+	if (time_slice >= 0.0f)
+	{
+		static std::map<const void *, Scalar> s_filterClock;
+		Scalar &clock = s_filterClock[this];
+		clock += time_slice;
+		const Scalar kPodTick = 1.0f / 28.0f;			// [T1] pod Perform cadence
+		if (clock < kPodTick)
+		{
+			return;										// heatLoad holds the last average
+		}
+		int catchUp = 0;
+		while (clock >= kPodTick && ++catchUp <= 15)	// >15 samples saturate the window
+		{
+			clock -= kPodTick;
+			heatFilter.AddSample(sample);				// FUN_0043ade4
+		}
+		if (catchUp > 15)
+		{
+			clock = 0.0f;								// hitch: window already saturated
+		}
+		heatLoad = heatFilter.Average();				// FUN_0043ae0b
+		return;
+	}
+
+	heatFilter.AddSample(sample);		// FUN_0043ade4 (prime call)
 	heatLoad = heatFilter.Average();	// FUN_0043ae0b
 }