feat(learnings): redesign as action-gated knowledge base with seven skills

[Ragazoor]

Summary

• Renamed confirm_learning → upvote_learning (DDD alignment: signals quality, not workflow transition)
• Replaced LearningKind::Episodic with LearningKind::Landscape (for codebase/system overviews); migration v44 converts existing episodic rows to convention
• Added seven action-specific skills installed via dispatch setup: /pr-workflow, /codebase-knowledge, /code-conventions, /test-conventions, /dispatch-workflow, /troubleshoot, /improvement — agents invoke these at action checkpoints rather than only at task start
• Replaced the generic dispatch prompt learning instruction with explicit checkpoint references to all seven skills, fixing the root cause of zero agent upvotes (retrieval happened too early and too broadly)
• Renamed the TUI overlay from "Learnings" to "Knowledge Base"; updated the /learnings skill to cover lifecycle management (upvote + record) and defer discovery to the domain skills
• Updated Allium specs (core.allium, learnings.allium) and CLAUDE.md to reflect all renames and the new design

Test plan

• cargo test — all 2288 tests pass
• dispatch setup installs the seven new skill directories under ~/.claude/plugins/local/dispatch/skills/
• Open the TUI, press L — overlay title reads "Knowledge Base"
• Dispatch an agent and verify the prompt includes the seven /skill-name checkpoint references
• record_learning → immediately approved, no more proposed state
• upvote_learning (formerly confirm_learning) still increments confirmed_count

Notes

The root cause diagnosed during brainstorming: agents queried learnings at task start and then never again, so relevant context arrived at the wrong moment and nothing got upvoted. The seven domain skills solve this by putting queries at the decision point (before writing code, before opening a PR, etc.).

Implements #585.