Learns which tools are reliable from execution feedback.

• Tracks success rate and average score per tool
• Produces preferred_tools and discouraged_tools lists
• Configurable: success_threshold=0.6, min_examples=1

Turns policy-related feedback into concrete recommendations.

• Detects tools that are frequently blocked or risky
• Recommends deny, require_approval, or relax per tool
• Compares against baseline policy

Solidifies failure root causes into permanent constraints (Ψ.constraints).

• Extracts (tool, root_cause) pairs from failed feedback entries
• Adds new constraints to prevent repeat failures
• Ratchet-risk mitigation: marks constraints as stale after stale_after entries with no violations — prevents capability decay from over-constraining

Detects when a task domain causes persistent early_stop events and recommends spawning a specialist sub-agent.

• Tracks early_stop ratio per domain
• Triggers recommendation when task_ratio(domain) > split_threshold (default 0.4)
• Requires min_samples=3 before recommending

Only E2, no E3	Gets better at tasks, but repeats the same mistakes
Only E3, no audit	Gets more cautious until capability atrophies
Only E4, no protocol	Single-agent chaos becomes distributed chaos

All four strategies together keep capability, constraints, and structure evolving in balance.

Code: loom/evolution/strategies.py