v0.19.5 — CLI honesty + safe prune

A fourteenth adversarial review, a thorough 2nd pass of the ~4500-line Typer CLI. Most command groups reviewed clean (doctor/init don't leak secrets, migrate/skills-retire are dry-run + reversible, solve/drift exit non-zero correctly). Five fixes:

Fixed

• CLI gates now exit non-zero on failure. chimera workflow and chimera lifecycle (and the fusion-bench/cascade-bench/skillcard-bench verdicts) printed a red "failed"/"REGRESSION" but exited 0 — so chimera workflow triage.yaml && ./deploy.sh would deploy on a failed workflow, and a bench regression stayed green in CI. They now exit non-zero on the failing branch.
• memory prune no longer silently deletes data. It ran immediately and hard-deleted everything below the budget — including persona/profile facts. It is now dry-run by default (shows the count; --apply to delete), and persona facts are never pruned.
• solve HITL flags are validated — more than one of --approve/--respond/--edit/--deny is refused, and --edit requires --answer (it would otherwise finalize the paused run with an empty answer).
• .env writes are atomic (temp + os.replace) so a crash mid-write can't truncate your secrets/config.

Gate: 1458 passed / 3 skipped, mypy + ruff green.

Upgrade: `pip install --upgrade 'chimera-agent[full]'`