CLAUDE: T1 - sequence-start clean-seed reset (rerun == fresh launch)

At every poseSequencePass start, explicitly drop the per-sequence GPU task
template and the DP seeding key, re-arm the scene-0 DP oracle, and return the
device image-set ring to set 0. A back-to-back rerun in the same JVM now seeds
identically to a fresh launch (restores the simple byte-identity gate; fresh
starts were already bit-exact, reruns wobbled 0.0024 px from carried-over warm
state - the previous run's last D5 overlap leaves an arbitrary active set while
the new run's ring bookkeeping assumes set 0), and a (re)started production RT
sequence never inherits stale resident state.
Co-authored-by: Claude Fable 5 <claude-fable-5@anthropic.com>

src/main/java/com/elphel/imagej/cuas/rt/CuasPoseRT.java

@@ -2831,6 +2831,25 @@ public class CuasPoseRT {
 		// block + tile centers) are keyed on the reference ErsCorrection and must be
 		// re-armed at every sequence start. By Claude on 07/17/2026.
 		IntersceneLma.resetPoseSequenceCache();
+		// T1 sequence-start CLEAN-SEED reset (Andrey 07/18): a rerun in the same JVM
+		// must seed identically to a fresh launch, and a (re)started RT sequence must
+		// never inherit resident state from the previous one. Explicitly drop the
+		// per-sequence task template and the DP seeding (the identity keys usually
+		// force a rebuild anyway - this makes it a contract, not a coincidence),
+		// re-arm the scene-0 DP oracle (a real divergence fails again and disables
+		// DP for THAT run only), and return the device image-set ring to set 0: the
+		// previous run's last D5 overlap leaves an arbitrary ACTIVE set while the
+		// ring bookkeeping below restarts at 0 - stale bookkeeping would let the
+		// first overlap target the set the running chain still reads (the 07/18
+		// back-to-back rerun 0.0024 px wobble; fresh starts are bit-exact).
+		// By Claude on 07/18/2026, from Andrey's T1 ruling.
+		gpu_task_template = null;
+		dp_seeded_key = null;
+		dp_oracle_failed = false;
+		if ((center_CLT.getGPUQuad() != null) &&
+				center_CLT.getGPUQuad().supportsPoseSceneDpOverlap()) {
+			center_CLT.getGPUQuad().selectImageSet(0); // idempotent on a fresh start (setup alloc = set 0)
+		}
 		// DEBUG (pose_corr_save): render per-scene pixel-domain correlations vs the virtual
 		// center (-POSE-RT-CORR2D, z=scenes, last LMA cycle, 16x16 cells per tile) - the same
 		// visual debugging used on the oracle. Lean only: the oracle engine does not expose