Unified approvals, cleaner evidence, stronger run lifecycle

Human judgment stays at the risk points. This release puts HITL prompts, plans, campaigns, and budget checkpoints in one approval card - including in-chat approve / deny.

Run evidence fits the pane. Interrupted runs are marked interrupted, not live. Effort sits beside the model. Coding-agent sessions can attach native screenshots as recorded evidence.

The loop is only useful if the record is honest. This release makes the record honest.