Add Task 076 progress notes (202/266 best, key insights documented)

medal-solvers/TASK076_PROGRESS.md

@@ -0,0 +1,31 @@
+# Task 076 Progress
+
+## Best: 202/266 (v11 - 8-connected BFS for decoration detection)
+
+## What works:
+- 4-connected component detection for bars (scipy.ndimage.label)
+- 8-connected BFS from bars to find decorations (1, 2, 3 cells)
+- 8 transforms (4 rotations + 4 reflections) matching rotated template to incomplete bars
+- Using the 2-marker as the reference point for relative positioning
+- Applying rotated decorations at matching positions
+
+## What fails (64 examples):
+- ~20: No matching transform found (bar shape relative to 2 differs due to 2 being at different position along bar)
+- ~33: Template detection fails (BFS over-connects nearby objects)
+- ~11: Unknown
+
+## Key insight discovered:
+The 2-marker can be at DIFFERENT positions along the bar in different objects of the same example.
+E.g., template has 2 at position 1 of 5 along bar, but incomplete has 2 at position 3 of 5.
+This means the relative coordinates of bar cells differ between template and incomplete
+even when the correct rotation is applied.
+
+## What to try next:
+1. Match bars by centroid-offset (align centroids, not 2-positions) for finding rotation
+2. Then compute the parallel offset (how much the 2 shifted along the bar after rotation)
+3. Apply that parallel offset to decorations (shift decorations along the bar axis)
+4. This requires knowing the bar axis direction after rotation to shift correctly
+
+## Files:
+- task076_solver_194.py (uploaded earlier, uses simple 8-transform on 4-adj decos)
+- v11 (202/266) is the best - uses 8-BFS from bar for decoration detection