Fix solver_registry.py: _check_scoreable expects (memory, params) not (macs, memory, params)

Fix submission.py: adapt to new profiler API returning (memory, params) instead of (macs, memory, params)"

Fix profiler.py: new neurogolf_utils.score_network requires (model, trace_path) — use static fallback for local scoring, real scoring happens on Kaggle

Add sign_corrected_conv.py: refit conv_var tasks with direct output (>0 validation, 3 nodes instead of 15-20)

Add wave22_compactor.py with Conv refit + sign-correction for direct Conv→output architecture"

Upload submission-v90.zip

Add merge_best_of_both.py: picks lowest-cost model from 5743 vs V90 per task (42 tasks improved, 55% size reduction)"

Fix optimize_submission.py: add --input_dir for Kaggle unzipped datasets (no need to re-run solver)

Add optimize_submission.py — unified 4-stage optimizer for new scoring formula

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

Add Stage 1 trivial task optimizer (zero-intermediate architectures for max score)

Update neurogolf_utils.py to May 14 2026 version (MACs removed from scoring, ORT profiler memory, sanitize_model)

Add corrected PLAN_TO_6043.md with verified scoring formula analysis

Update README.md: V90 5322.64, target >6043

Update NEXT_AGENT.md: target >6043, size optimization strategy

Update TODO.md: target >6043, score optimization roadmap

Update LEARNING.md: V90 5322.64, target >6043 strategy

Add fallback_unsolved.zip: best reference models for 15 unsolved tasks

Add --fallback_zip: inject reference models for unsolved tasks

Fix registry: use descriptive solver names (no task numbers)

Fix wave22: rename all solvers to descriptive names (no task numbers)

Update README.md: V83 - 385/400 tasks solved

Update NEXT_AGENT.md: reflect V83 state, focus on score optimization

Update TODO.md: reflect 385/400 solved, pivot to score optimization

Update LEARNING.md: document wave22 results (385/400 solved)

Update registry: add wave22 (143 solvers, total 386 tasks)

Add wave22: 143 new reverse-engineered solvers

Update LEARNING.md for V82 state

Update NEXT_AGENT.md for V82 state

Update TODO.md for V82 state

Update all documentation for V82 (221/400, LB 3159.03)

Add wave21 to registry: 24 new tasks

Add wave21: 24 new tasks (91,100,204,215,217,221,225,228,231,232,235,239,246,247,251,252,256,259,260,265,269,279,289,298)

Add wave20 to registry: 20 new tasks

Add wave20: 20 new tasks (92,94,109,112,114,121,124,159,161,165,175,177,183,187,188,195,196,198,199,200)

Fix solver_registry: correct wave19 import placement

Add wave19 to registry: 15 new tasks

Add wave19: 15 new tasks (3,8,10,11,17,19,27,30,48,58,64,71,75,84,86)

Update registry with descriptive wave18 names

Rename wave18 solvers to descriptive names (not taskNN)

Add wave18 to registry: tasks 24,37,47,63,81,93,125,128,137,139

Add wave18: 10 new tasks (24,25,37,47,63,81,93,125,128,137,139)

Add tasks 344 and 59 to registry"

wave17: fix task38 Resize->Tile, add tasks 344+59

Fix task 38: Replace Resize with Tile (onnx_tool compatibility)

Update NEXT_AGENT.md with wave17 info and next targets

Add wave17 to registry: tasks 4, 38, 106, 108, 123, 153, 180, 369

Add wave17 solvers: 8 new tasks (4, 38, 106, 108, 123, 153, 180, 369)

Add wave17 solvers: 8 new tasks (4, 108, 153, 106, 123, 38, 180, 369)

Add wave17.py header (full file follows)