Integrate Dan 1:1 (2026-05-28) takeaways into research context

Fold the Dan Mossing 1:1 (2026-05-28) takeaways and action items into the living research context. Source notes live in docs/mentor_updates/2026-05-28.md (the weekly update Thomas sent Dan; captures Dan's pushes and the agreed next steps). Why it matters: keeps research_ideas.md Topics and the Phase Tracker aligned with what the mentor actually steered toward, so daily work tracks the agreed priorities rather than drifting. Concrete next steps: (1) the two directions Dan pushed (behavior-leakage B to B', tasks #404/#411; geometry-predicts-generalization, #406) should be reflected in research_ideas.md (Localization/Propagation + the User Modeling Topic) and the Phase Tracker rationale column. (2) Dan's X vs (X-Y) spurious-correlation caveat on selectivity (#383/#397) should be noted as an open question. (3) The deprioritize-make-evil-dumb decision (per Dan's RL-incentive argument) should be recorded. Starting point: read docs/mentor_updates/2026-05-28.md sections 3 (open questions) and 4 (next steps), then diff against research_ideas.md to find what is missing.