Add TRM solver skill for agent guidance

trm_solver/SKILL.md

@@ -0,0 +1,76 @@
+# TRM Solver Skill — NeuroGolf 2026
+
+## Purpose
+Use Kilo (headless DeepSeek CLI) to analyze ARC-AGI tasks and drive a tiny NN executor that exports ONNX models for NeuroGolf submission.
+
+## Pipeline
+```
+ARC task JSON → render as image → kilo run → markdown output → parse TransformSpec → create NN → export ONNX
+```
+
+## Key files
+- `trm_solver/executor.py` — 14 transform NN implementations + ONNX export
+- `trm_solver/kilo_bridge.py` — full pipeline: render → Kilo → parse → NN → ONNX
+- `trm_solver/README.md` — usage and transform catalog
+
+## How to run
+
+### Quick test (no LLM needed)
+```bash
+python trm_solver/kilo_bridge.py --dry-run
+```
+
+### Single task
+```bash
+python trm_solver/kilo_bridge.py --task data/training/007bbfb7.json --output-dir onnx_models
+```
+
+### Batch (overnight for full 400)
+```bash
+python trm_solver/kilo_bridge.py --data-dir data/training --output-dir onnx_models
+```
+
+### From Kaggle (SDK tunnel)
+```bash
+python trm_solver/kilo_bridge.py --data-dir data/training --use-sdk --server-url http://localhost:8765
+```
+
+## Kilo prompt format
+The prompt is in `kilo_bridge.py` as `KILO_PROMPT_TEMPLATE`. It instructs DeepSeek to output:
+
+```markdown
+## Transform
+name: <transform_name>
+
+## Parameters
+- param1: value1
+```
+
+Available transforms: identity, color_map, flip, transpose, rotate, upscale, kron_self_similar, tile_repeat, concat_patterns, pos_color_lut, spatial_gather, onehot_conv, onehot_linear.
+
+## What to do when Kilo gets it wrong
+1. Edit the `_spec.json` file in the output directory
+2. Re-export manually:
+```python
+from trm_solver.executor import TransformSpec, create_transform_nn, export_to_onnx
+spec = TransformSpec(name="correct_name", params={...})
+model = create_transform_nn(spec)
+export_to_onnx(model, (H, W), "task_id.onnx")
+```
+
+## ONNX size targets
+- Most transforms: 0-1000 params (< 10 KB ONNX)
+- onehot_conv (3×3): 900 params (~12 KB)  
+- onehot_conv (5×5): 2500 params (~30 KB)
+- Target: keep all models under 50 KB for NeuroGolf
+
+## Verification
+```bash
+# Check ONNX model works
+python -c "
+import onnxruntime as ort
+import numpy as np
+sess = ort.InferenceSession('onnx_models/007bbfb7.onnx')
+out = sess.run(None, {'input': np.zeros((1,1,3,3), dtype=np.float32)})
+print(out[0].shape)
+"