feat: built-in blessed-patterns layer for common AI coding agent patterns

[jamessoubry]

Problem

Claude Code and other AI agents have predictable, safe command patterns that clawband repeatedly asks about — building binaries, running test suites, jq/python3 pipelines, heredoc config writes. These are standard AI-agent idioms, not risky user commands.

Proposed behaviour

Add a builtin_allow() tier (alongside builtin_deny and builtin_ask) of patterns that are known-safe in an AI coding agent context and should pass through silently without consulting user allow.patterns.

Example candidates:

• cargo build, cargo test, cargo clippy, cargo fmt
• python3 -c "import sys,json; ..." (visible inline code)
• jq '.field' <<< '$VAR'
• gh pr, gh issue, gh release commands
• git log, git diff, git status, git fetch

Notes

• These should be conservative and specific — broad patterns risk suppressing genuinely risky commands
• User can override with their own deny.patterns / ask.patterns
• Distinct from user allow.patterns — these ship with the binary and represent Anthropic/maintainer judgment about AI agent safety