Rewrite LEARNING.md: only facts, no fluff

Rewrite TODO.md: real roadmap focused on reverse-engineering 6043

Rewrite NEXT_AGENT.md: real instructions, no bullshit

Rewrite docs: honest state, real plan (reverse-engineer 6043 into DSL solvers)

merge_best_of_both.py: use 6043 as-is, no overrides (our cost estimate is unreliable)"

merge_best_of_both.py: use 6043 as base, only override if other source has cost < 50% of 6043"

sign_corrected_conv.py: read from input_zip, only replace if lower cost"

merge_best_of_both.py: validate via neurogolf_utils in subprocess (crash-safe)"

Fix merge_best_of_both.py: skip validation (all sources are Kaggle-validated), prevents ORT segfault crash"

Delete own-solver/merge_submissions.py, own-solver/optimize_submission_README.md

Update merge_best_of_both.py: accept --sub_6043, use shape-inference cost, pick lowest cost from all sources

Fix LEARNING.md: remove version numbers from filenames

Fix NEXT_AGENT.md: remove version numbers from filenames

Add submission.zip: best-of(6043+5743), expected score >=6043

Update optimize_submission.py: multi-stage optimizer (trivial rebuild, direct conv, sign-corrected conv, cleanup)

Add merge_submissions.py: picks lowest-cost model per task from multiple zips

Delete own-solver/merge_best_v5.py, own-solver/optimizer_v4.py, submission-v5-merged.zip

Add optimizer_v4.py: multi-stage optimizer (Stage 1-6)

Update TODO.md with corrected roadmap post-V5"

Update NEXT_AGENT.md with V5 results and corrected strategy

Update LEARNING.md with V5 results, fp16 failure, correct cost model

Add submission-v5-merged.zip: best-of(6043+5743), est. score >6043

Add merge_best_v5.py: combines 6043+5743 for optimal submission

Update README.md: V3 state, correct scoring, optimization tools"

Update NEXT_AGENT.md: correct formula, V3 baseline, clear optimization priorities

Update TODO.md: corrected scoring, new roadmap based on V3 results

Update LEARNING.md: corrected scoring formula, V3 results, new optimization strategies

Restore solver_registry.py (accidentally overwritten)

Fix profiler.py: return 3 values (macs=0, memory, params) for backward compat with solver_registry.py

Fix _check_scoreable: unpack (memory, params) from new profiler.score_network()

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)