Design a bounded policy-update loop from critic findings

[MichaelSchmidle]

Migrated from legacy Fizzy card #79 after Memrok work moved to GitHub Issues + milestones.

Goal

Design a bounded policy-update loop for Memrok from critic findings.

Principle

• critic outputs become recommendations, not immediate engine changes
• aggregation produces candidate updates to weights, caps, gates, and prompts
• human-approved changes land first; bounded self-tuning can come later

Needs

• clear separation between observation, recommendation, and application
• audit log and rollback for any future automatic parameter tuning
• distinction between global defaults and user-local adaptation