feat: add kaggle_competition_skills tool — domain knowledge from 20+ competitions

[topprismdata]

Summary

Adds a new tool (kaggle_competition_skills) to MLZero's tool registry, containing crystallized domain knowledge from 20+ real Kaggle competitions (6 Gold medals, 14× speedup).

This tool gives MLZero agents the ability to:

• Avoid known dead ends (30+ confirmed failure patterns)
• Apply competition-specific strategies (autogluon-first, OOF/LB gap awareness, TrueSkill simulation strategy)
• Transfer knowledge across tasks (3-layer architecture: core → domain → universal principles)

What's Added

tool.json

10 prompt_template rules that guide MLZero's code generation:

• "Always run AutoGluon as FIRST baseline"
• "ONE variable per experiment"
• "Always compare OOF vs LB"
• "Model correlation > 0.93 → ensembling yields < 0.001"
• "For time series: use TimeSeriesPredictor, Chronos-2 supports known_covariates"
• "3× <0.0001 improvement → pivot"

tutorials/

File	Content
kaggle-sop.md	Standard experiment operating procedure (Phase 0-4)
cv-lb-gap.md	OOF vs LB gap — when it's large, decision rules
competition-types.md	Strategy per competition type (tabular/code/sim/timeseries)
cross-domain-principles.md	16 universal principles validated across competitions

condensed_tutorials/

Quick decision tree + top 5 rules + top 5 dead ends.

Validation

These skills are validated by:

• 6 Gold medals (Jigsaw Toxic, Text Normalization, Leaf Classification, TPS Dec 2021, Spaceship Titanic, Denoising)
• Top 4.7% in PTCG AI Battle (simulation, TrueSkill)
• 14× speedup in time-to-Top-10% (2 weeks → <1 day through knowledge accumulation)
• 37+ crystallized skills extracted from real competition outcomes
• 30+ confirmed dead ends that prevent wasted exploration

Source: cultivating-ml-agent

How This Complements MLZero

MLZero	This Tool
MCTS explores solution space blindly	Skills guide MCTS toward promising directions
Starts from scratch each task	Skills activate from prior competition experience
Discovers dead ends through trial	Dead-end list prevents wasted iterations
Doesn't model OOF/LB gap	OOF/LB gap awareness prevents overfitting CV

Test Plan

• Run MLZero on Store Sales competition WITH this tool → should choose TimeSeriesPredictor (not TabularPredictor)
• Run MLZero on tabular competition WITHOUT this tool → baseline comparison
• Verify ToolSelectorAgent correctly selects kaggle_competition_skills for competition tasks
• Verify prompt_template rules appear in generated code