Wave 2+flood fill scan: 0 matches. Update LEARNING with findings.

neurogolf_solver/solvers/WAVE2_SCAN.md

@@ -1,6 +1,6 @@
-# Wave 2 Scan Results (2026-04-27)
+# Wave 2 + Flood Fill Scan Results (2026-04-27)
 
-## Patterns tested against 349 unsolved tasks:
+## Wave 2 — Composition & Mode Extensions
 
 | Pattern | Matches |
 |---------|---------|
@@ -9,20 +9,40 @@
 | row_mode_fill (each row → dominant color) | 0 |
 | col_mode_fill (each col → dominant color) | 0 |
 | fill_bg_with_mode (zeros → global mode) | 0 |
-| fill_bg_with_color (zeros → fixed color) | 0 (PARTIAL on first example only for ~30 tasks) |
+| fill_bg_with_color (zeros → fixed color, all examples) | 0 |
 
-## Why bg fill tasks are PARTIAL:
+## Flood Fill
 
-The 53 tasks where "only bg pixels changed" do NOT do simple global fill.
-They fill bg pixels based on **spatial context**:
-- Fill inside a shape boundary
-- Fill adjacent to a specific color
-- Fill based on nearest non-bg neighbor
+| Pattern | Matches |
+|---------|---------|
+| flood_fill_replace (seed spreads into passable, all become fill_color) | 0 |
+| flood_fill_keep_seed (seed stays, passable neighbors become fill_color) | 0 |
+
+## Pattern Inpainting
 
-These require spatial reasoning (flood fill, distance transforms, object detection)
-that cannot be done with simple ReduceSum/ArgMax/Where.
+| Pattern | Matches |
+|---------|---------|
+| Tile inpainting (output = perfect tile, input = tile with holes) | 0 |
+
+## What the tasks ACTUALLY need (from manual inspection):
+
+- **Task 5**: Pattern stamping at positions indicated by markers
+- **Task 17**: Wallpaper defect restoration (NOT simple tile inpainting)
+- **Task 20**: Diamond symmetry completion with color-specific rules
+- **Task 27**: Shape-relative region filling (notch detection)
+
+These require **object-level reasoning**: detect shapes, understand spatial relationships
+between objects, apply context-dependent rules. Cannot be solved by pixel-level operations
+(flood fill, mode fill, color mapping) alone.
 
 ## Conclusion:
 
-Wave 2 analytical solvers have 0 viable targets in this task set.
-The remaining tasks require fundamentally different approaches.
+Simple analytical solvers (Waves 1-2) and pixel-level propagation (flood fill)
+have reached their ceiling. The remaining 349 tasks need:
+1. Object detection/segmentation
+2. Spatial relationship reasoning
+3. Context-dependent rule application
+4. Pattern recognition beyond tiling
+
+These are fundamentally in the domain of learned models (conv lstsq already does this
+for some tasks) or much more complex hand-crafted solvers.