docs: Add Gemini notes on NaN disparity handling upon correlation failure

docs/TODO_BUG_REPORTS_AI.md

@@ -19,3 +19,16 @@ Verify if `execCorr2D_normalize` (or the downstream CPU equivalent if GPU falls
 
 ---
 
+
+### [2026-04-24] Missing Peak Fallback (Gemini)
+**Location:** `CuasRanging.java`, method: `rangeSingleTarget()`
+
+**Observation:** 
+If the integrated correlation fails (e.g., peak drowned out by the potential double-normalization bug), `disp_diff` evaluates to `NaN`. Because `use_non_lma` is `false`, it doesn't fall back to Center-of-Mass (CM) disparity. The method returns `null`, abruptly aborting the refinement loop for that segment. 
+
+This causes the outer calling function to potentially log or retain `0.0` or `NaN` for `RSLT_GDISPARITY` (Global Disparity), rather than properly recording a failed integration or utilizing the un-refined initial disparity.
+
+**Action Required:**
+Check if returning `null` on failed refinement is the intended behavior, or if the system should persist the initial (single-frame) disparity into the `RSLT_GDISPARITY` field rather than zeroing it out when integration fails.
+
+---

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

@@ -1037,6 +1037,12 @@ public class CuasRanging {
 					disp_diff = disparity_map[ImageDtt.DISPARITY_INDEX_CM][ref_tile];
 					str = disparity_map[ImageDtt.DISPARITY_INDEX_CM+1][ref_tile];
 				}
+				/*
+				 * [Gemini Note, 2026-04-24]: If correlation fails (e.g. due to double-normalization destroying the peak),
+				 * disp_diff is NaN. If use_non_lma is false, disp_diff remains NaN, and the function returns null here,
+				 * meaning the loop exits. This causes target[CuasMotionLMA.RSLT_GDISPARITY] to likely retain a 
+				 * zero/NaN value depending on caller initialization.
+				 */
 				if (Double.isNaN(disp_diff)) {
 					System.out.println ("rangeSingleTarget(): Failed to range target #"+target_id+" for sequence "+ first_seq+" ("+ref_tileX+":"+ref_tileY+").");
 					return null;