docs: Add Gemini notes on potential double-normalization bug in TD accumulation

docs/TODO_BUG_REPORTS_AI.md

+# AI Agent Bug Suspicions & TODOs
+
+This file tracks potential issues, structural anomalies, and mathematical mismatches identified by AI agents during code analysis. Items here require human verification before invasive code changes are made.
+
+---
+
+### [2026-04-24] Potential Double Normalization in Ranging Accumulation (Gemini)
+**Location:** `CuasRanging.java`, method: `refineSingleTargetDisparity()` around line 1946.
+
+**Observation:** 
+The method accumulates transform-domain (TD) correlations across multiple scenes to increase SNR for deep disparity. 
+1. It calls `OpticalFlow.accumulateCorrelationsAcOnly(accum_weights, fcorr_td_acc)`. This method loops over the array and divides `fcorr_td_acc` values by `accum_weights`.
+2. It then calls `image_dtt.clt_process_tl_correlations(...)`, passing *both* the normalized `fcorr_td_acc` and the `accum_weights`.
+
+If the downstream GPU call (`gpuQuad.execCorr2D_normalize`) uses `accum_weights` to divide the values again, the integrated signal is divided by $N^2$ instead of $N$. For a sequence of 200 frames, this suppresses the signal by a factor of 40,000, effectively destroying the correlation peak.
+
+**Action Required:**
+Verify if `execCorr2D_normalize` (or the downstream CPU equivalent if GPU falls back) applies division by the passed weights. If so, remove the `accumulateCorrelationsAcOnly` call in `refineSingleTargetDisparity`.
+
+---
+

src/main/java/com/elphel/imagej/cuas/CuasRanging.java

@@ -1947,6 +1947,14 @@ public class CuasRanging {
 		} // for (int nseq = 0; nseq < num_seq; nseq++) {
 		
 		// Normalize accumulated correlations
+		/* 
+		 * [Gemini Note, 2026-04-24]: Potential Double-Normalization Bug.
+		 * accumulateCorrelationsAcOnly explicitly divides the fcorr_td_acc arrays by accum_weights.
+		 * Immediately after, image_dtt.clt_process_tl_correlations is called and also receives accum_weights.
+		 * Inside the GPU pipeline (execCorr2D_normalize), the weights might be applied again.
+		 * If so, the signal is divided by N^2 instead of N, causing massive signal loss on long integrations.
+		 * Requires verification of GPU behavior before removing.
+		 */
 		OpticalFlow.accumulateCorrelationsAcOnly( 
 				accum_weights,       // final float [][][]     num_acc,          // number of accumulated tiles [tilesY][tilesX][pair]
 				fcorr_td_acc); // final float [][][][] fcorr_td_acc      // [tilesY][tilesX][pair][256] sparse transform domain representation of corr pairs