CLAUDE: chain-rule diagnostic - RMS meas>solved pairs in the pose cycle trace

Per outer cycle: RE-MEASURED pure RMS (prepareLMA, after re-correlation) >
MODEL-PREDICTED pure RMS (runLma, after the solve). Jacobian/chain-rule error
signature (Andrey's hypothesis): solved_n low but meas_{n+1} bounces back up -
the model overpromised and re-correlation takes it back. Healthy damping:
meas_{n+1} ~ solved_n, both creep down. Zero extra computation - both values
already existed inside IntersceneLma.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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

@@ -573,6 +573,13 @@ public class CuasPoseRT {
 		double last_diff_atr = Double.NaN; // By Claude on 07/12/2026
 		final double [] cycle_datr = new double [max_cycles]; // per-cycle ATR-change trace (convergence trajectory) // By Claude on 07/12/2026
 		final int [] cycle_inner = new int [max_cycles]; // inner runLma iterations per outer cycle // By Claude on 07/12/2026
+		// Chain-rule diagnostic (Andrey 07/12/2026): per cycle, the RE-MEASURED pure RMS
+		// (after re-correlation, from prepareLMA) vs the MODEL-PREDICTED pure RMS after
+		// the solve (from runLma). Jacobian/chain-rule error signature: solved_n low but
+		// meas_{n+1} bounces back up (the model overpromised, re-correlation "takes it
+		// back"). Healthy damping signature: meas_{n+1} ~ solved_n, both creep down.
+		final double [] cycle_rms_meas   = new double [max_cycles]; // By Claude on 07/12/2026
+		final double [] cycle_rms_solved = new double [max_cycles]; // By Claude on 07/12/2026
 		int ncyc = 0; // cycles actually executed // By Claude on 07/12/2026
 		// lean v2 (Andrey 07/12/2026): the inner LMA restarts at ilp.ilma_lambda (0.1) every
 		// outer cycle; combined with the 0.1%-improvement inner exit this under-relaxes the
@@ -608,6 +615,7 @@ public class CuasPoseRT {
 					(nlma == 0),                      // first_run
 					null,                             // dbg_prefix
 					clt_parameters.imp.debug_level);  // debug_level
+			cycle_rms_meas[nlma] = intersceneLma.getLastRms()[1]; // re-measured pure RMS at this cycle's pose // By Claude on 07/12/2026
 			// pose_lma_debug >= 2: per-inner-step solver lines; header attributes them to
 			// their outer cycle (the solver lines carry no scene/cycle context).
 			// By Claude on 07/12/2026, from Andrey's request.
@@ -635,6 +643,7 @@ public class CuasPoseRT {
 			last_diff_atr = diffs_atr[0]; // By Claude on 07/12/2026
 			cycle_datr[nlma] = diffs_atr[0]; // By Claude on 07/12/2026
 			cycle_inner[nlma] = lmaResult;   // inner runLma iterations this cycle // By Claude on 07/12/2026
+			cycle_rms_solved[nlma] = intersceneLma.getLastRms()[1]; // model-predicted pure RMS after the solve // By Claude on 07/12/2026
 			ncyc = nlma + 1;                 // By Claude on 07/12/2026
 			if ((fixed_cycles <= 0) && // lean v2: fixed count = no early exit // By Claude on 07/12/2026
 					(diffs_atr[0] < clt_parameters.imp.exit_change_atr) &&
@@ -661,9 +670,15 @@ public class CuasPoseRT {
 			for (int i = 0; i < ncyc; i++) tsb.append(String.format(" %.3g", cycle_datr[i]));
 			final StringBuilder isb = new StringBuilder(); // inner runLma iterations per cycle // By Claude on 07/12/2026
 			for (int i = 0; i < ncyc; i++) isb.append(" "+cycle_inner[i]);
+			// meas>solved pairs: re-measured pure RMS at cycle start > model-predicted after
+			// the solve. Chain-rule error = the NEXT meas bounces back above this solved.
+			// By Claude on 07/12/2026, from Andrey's hypothesis.
+			final StringBuilder rsb = new StringBuilder();
+			for (int i = 0; i < ncyc; i++) rsb.append(String.format(" %.4f>%.4f",
+					cycle_rms_meas[i], cycle_rms_solved[i]));
 			System.out.println("CuasPoseRT cycles: scene "+scene.getImageName()+
 					" ("+ncyc+" of max "+max_cycles+", lambda0 "+lma_lambda+") dATR/cycle ="+tsb+
-					"; innerLMA ="+isb);
+					"; innerLMA ="+isb+"; RMS meas>solved ="+rsb);
 		}
 		if (lma_rms != null) {
 			final double [] last_rms = intersceneLma.getLastRms();