Upload 3 files

Delete medal-solvers/optimized/task023_fp16.onnx

Delete medal-solvers/optimized/task023_fp16_fused.onnx

Delete medal-solvers/optimized/task023_h48_hardened.onnx

Delete medal-solvers/optimized/task023_minH32.onnx

task023 H48 hardened score=12.893 real=266/266 gate=4977/5000

task023 conv7 minimal H=32 score=13.284 real=266/266 gate=4959/5000

task023 fp16 + dedup/dead-elim fusion (266/266 exact)

task023 fp16 lossless shrink (266/266 exact match)

Delete medal-solvers/optimized/task023_5iter_fp16.onnx

Delete medal-solvers/optimized/task023_6iter.onnx

Delete medal-solvers/optimized/task023_v2.onnx

task023: fp16 surgery on ORIGINAL 370-node model. 266/266 + 5000 random PASS. Score 11.67 vs base 11.06 (+0.61)

task023: 5-iter fp16, 266/266 pass, score 11.840

task023 v2: delta approach (no Concat with repeated zeros), 6 iter, 266/266

Add task023 reduced ONNX (6 iterations, 361 nodes, 266/266 + 2500 random pass)

Task023: Conv7x7+ReLU+Conv1x1 F32, C=32, 266/266, 3 nodes, score~12.6

Task023: 2-Conv fp16 (7x7+1x1+ReLU), 266/266 pass, 3 nodes, ~13+ pts

Add task023 fp16 ONNX (266/266 pass, score 11.639, gain +0.578 over 11.061)

Delete medal-solvers/optimized/task023.onnx

Add task023 trained Conv model (score 11.78, +0.72 from 11.06)

Update task285.onnx: V4 propagation-based (score 10.696, +1.86 pts)

Task092 v4: score 13.245 (mem=127380)

Task092 v3: score 13.231 (memory=129180, +0.160 over base)

Add optimized task092.onnx (memory=143040, score=13.129 vs base 13.071)

Add task161 optimized ONNX (score 13.603, +0.429 gain over base 13.174)

Delete medal-solvers/optimized/task219.onnx

Delete medal-solvers/optimized/task219_nosim.onnx

Add task202 optimized ONNX (score 13.466, +0.208 gain over base 13.258)

Fix task219 ONNX v2: correct tie-break uses first-occurrence position (0 reproducible failures in 60K tests)

Fix task219 ONNX: 0/10000 failures on random inputs (was 6.4%)

task219: upload NON-simplified model (onnxsim may cause Kaggle issues)

Fix task219 ONNX: use opset 13 + ReduceMax/Min axes-as-attribute (Kaggle compatible)

Upload task219 non-simplified variant (869 nodes, backup)

Task 219 ONNX: 265/265, score 12.759 (was 8.422, +4.337)

task370 v6: 266/266, score 13.39 (+1.2 gain), 107 nodes, 18KB, 107K memory. Where-based paint, no clips on binary convs, eliminated [1,9,20,20].

task370 v5: 266/266, 124 nodes, 20KB, score 12.954 (+0.77 vs base 12.187). Single combined 11x11 kernel, 4 iterations.

Add task370 ONNX: 266/266 correct, 142 nodes, 17KB, static score 12.613 vs base 12.187 (+0.43). Needs Kaggle profiling.

Remove task041,045,098,356 from optimized/ — Kaggle profiling shows they score WORSE than base

Add task098 ONNX (boundary detection, 266/266 pass)

Add task356 ONNX model (fill-between rule, all examples pass)

Add task045 ONNX model (fill-between rule, all examples pass)

Add task041 ONNX model (fill-between rule, all examples pass)

Update task243 ONNX: 28 iter (score 12.996 vs 12.946, +0.05)

Add hand-crafted task243 ONNX model (flood fill, +1.979 pts)

Remove 11 onnxsim models from optimized/ - they hurt score vs 6043 base

Add onnxsim task363 to optimized (validated 100% correct, beats 6043 static)

Add onnxsim task330 to optimized (validated 100% correct, beats 6043 static)

Add onnxsim task324 to optimized (validated 100% correct, beats 6043 static)

Add onnxsim task233 to optimized (validated 100% correct, beats 6043 static)