P1 epic from the principal-engineer review (TASK-1.1 through 1.4):
- ship `bully` console entrypoint (`bully validate / lint / doctor /
  show-resolved-config`); hook.sh continues to call pipeline.py by
  absolute path so the hook env doesn't depend on $PATH
- skills + docs no longer hardcode $HOME/.bully/pipeline/pipeline.py;
  primary path is `bully` with documented plugin/manual fallbacks
- top-level `skip:` in .bully.yml + ~/.bully-ignore merge with built-in
  SKIP_PATTERNS so multi-language monorepos don't have to patch source
- --explain flag surfaces per-rule verdict (fire/pass/skipped <reason>/
  dispatched); run_pipeline gains include_skipped=True; hook-mode shape
  unchanged

Trust-gate docs in SECURITY.md; ast-engine + trust-gate tests land for
the prior P0 work.

Dogfood ourselves on the new tooling:
- two engine: ast rules in .bully.yml (prefer-perf-counter,
  no-shell-true-subprocess) scoped to pipeline source; the legitimate
  shell=True in execute_script_rule carries an inline bully-disable
- scripts/dogfood.sh prefers `bully lint` when on PATH and runs
  `bully validate` as a sanity preamble
- CI installs ast-grep-cli so ast rules actually fire instead of
  silently skipping; runs on every push (not just main)

Tests: 226 -> 243.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>