Fix evalbuff signal quality and add edit history to doc writer

[jahooma]

Summary

• Commit pre-copied docs in test repos so they don't leak into the agent's diff — previously judges penalized agents for "creating unnecessary documentation" that was actually pre-loaded context
• Isolate Claude calls (prompt generator + doc writer) by setting cwd=tmpDir so they don't read the repo's CLAUDE.md/AGENTS.md and get confused
• Filter lockfiles (bun.lock, package-lock.json, etc.) from diffs and file lists to reduce noise
• Add score comparison threshold (0.3 points minimum) to avoid accepting docs based on noise
• Pass edit history to the doc writer so it knows which prior docs were accepted/rejected (with scores), avoiding repeated failures and building on successes
• Cap improvement loop at 5 iterations to prevent runaway cost

Test plan

• Tested on freebuff commits: rate-limit commit went from 7/10 → 9/10, login-URL commit now scores correctly
• Run full evalbuff loop on a batch of commits to verify end-to-end

🤖 Generated with Claude Code