Add human-in-the-loop approval checkpoints before expensive stages

[lucastononro]

Gap. The orchestrator plans tasks then auto-executes end-to-end; request-clarification exists but is agent-initiated and ad hoc.

Value. Practitioners want to approve consequential decisions — target column, prep/leakage plan, model shortlist — before a long train burns budget. Converts "trust the autopilot" into "co-pilot with a steering wheel."

Suggestion. Add optional approval gates (a session setting or @mention toggle) where the agent posts its decision as an actionable card and blocks on user Approve/Edit, wiring into the existing InlineTasks + clarifications flow.

[lucastononro]

Addressed by #171, now merged into integration branch staging-v0.0.5 (ledger: STAGING-v0.0.5.md). Will land on main via the staging merge — closing.