Add eval cases for reminder delivery kind selection

[Aaronontheweb]

The reminder delivery contract (structured DeliveryKind on set_reminder) is fully implemented but missing behavioral eval cases that verify the LLM picks the correct delivery kind.

Suggested cases:

• "remind me in X minutes to check Y" from a Slack session -> LLM selects delivery.kind = current_session
• "when Z happens, post to #general" -> LLM selects delivery.kind = channel, transport = slack, address = #general
• Silent audit-style task -> LLM selects delivery.kind = none or current_session with deliveryRequired = false
• Regression case mirroring session D0AC6CKBK5K/1776697725.361339: Mode B reminder from Slack thread surfaces in the thread

Related: #690, #644, PR #692