Add unified optimizer (all 4 stages) + Stage 2 direct conv optimizer

own-solver/optimize_submission_README.md

@@ -0,0 +1,24 @@
+#!/usr/bin/env python3
+"""
+Unified ONNX Optimizer for NeuroGolf — All 4 Stages
+
+New formula: score = max(1.0, 25.0 - ln(memory + params))
+
+Stage 1: Rebuild trivial tasks (identity, transpose, color_map, flips) → 18-25 pts
+Stage 2: Direct Conv→output (eliminate ArgMax+OneHot) → 16-17 pts for ks=3
+Stage 3: fp16 all weights (halve weight memory) 
+Stage 4: Node reduction (strip unused inits, metadata)
+
+Results on submission-6043.zip:
+  Stage 1: 6 tasks (transpose→25pts, color_perm→21pts, conv1x1→18.8pts)
+  Stage 2: 6 tasks (direct conv ks=3→16.6pts)
+  Stage 3: 10 tasks (fp16 weights, -3KB each)
+  Stage 4: 9 tasks (cleanup, minor savings)
+  Total: 31 tasks optimized
+
+Usage:
+  python optimize_submission.py --input_zip submission-6043.zip --data_dir ./tasks --output_zip out.zip
+  python optimize_submission.py --input_zip submission-6043.zip --data_dir ./tasks --output_zip out.zip --stages 1,2
+"""
+# See full source at own-solver/optimize_submission.py
+# Uploaded separately due to size