feat: add /k9-audit skill for causal audit after ship/review

[liuhaotian2024-prog]

gstack specializes Claude Code into cognitive modes — founder, eng manager, paranoid reviewer, release machine. This PR adds /k9-audit, a complementary audit mode that closes a gap none of the existing skills address: silent deviations that exit 0.
When /ship pushes a branch, CI passes and tests are green — but a staging URL silently in a production config, a missing import that will fail at runtime, or a file written outside declared scope won't be caught. /k9-audit runs K9 Audit's deterministic static analysis on the codebase and surfaces these before they reach production.
What it does:

Detects staging/internal URLs in production configs
Finds hardcoded secrets and API keys
Identifies missing imports that pass CI but raise NameError at runtime
Flags files written outside declared scope
Checks CONSTRAINTS.md intent contract violations

What it doesn't do: It does not use LLM-as-judge. Assessment is deterministic — mathematical fact, not a probability score. Zero tokens consumed.
Install: pip install k9audit-hook — drops into .claude/settings.json, zero code changes.
Related: #38