shed teach (interactive correction mode)

[CasterlyGit]

What this adds

shed teach — interactive correction mode where the user explicitly tells shed it got something wrong and explains why. The explanation feeds the ranking model.

Spec

1. After any session, shed teach opens a TUI showing:
   • Last 10 prompts + what was injected
   • "Was injection X right? [y/n/explain]"
2. On "explain", user types why the injection was wrong
3. Shed:
   • Logs the corrected pairing to ~/.shed/state/teaching.jsonl
   • Adjusts per-memory injection score (immediate)
   • If the explanation reveals a pattern, files an internal proposal to update ranking weights (queued for morning brief)
4. shed teach --batch for power users — process all uncited injections in one go

Acceptance

• 1 teach session improves next-session injection accuracy by measurable amount
• Teach data feeds into L1 (L1: Closed-loop injection quality (cite-tracking → ranking) #9) and L5 (L5: Self-tuning thresholds for all proposal types #10) — accelerates the threshold tuning
• Teach mode never auto-rejects: user always confirms

Privacy

• User explanations stay local
• Sensitive corrections can be marked --ephemeral (not persisted, just affects current session)