feat: add causal learning selection

[EvanProgramming]

Summary

• rank learned guidance by verified utility per context character
• learn repeated harmful artifact pairs and avoid co-injecting them
• add paired omission trials, reversible retirement, near-miss/context metrics, and named benchmark ablations

Safety

• retirement requires non-regressing paired omission evidence
• retired artifacts keep their immutable pre-image and restore only as canaries
• existing profile, permission, secret, and external-skill boundaries remain intact

Validation

• make check
• KYROZEN_DB_PATH=/state.sqlite3 make test (46 tests)
• benchmark ablation replay
• git diff --check