/reflect — maintain the learner profile

Follow PROTOCOL.md § "Reflection protocol". Compressed here:

1. Load the raw material: python3 scripts/episodes.py --last 10 --json (the user_code, user_approach, hints_used, and verdict notes are what matters — mock episodes are the richest) and all of memory/profile.json.
2. Existing observations first. For each ACTIVE observation, scan the new episodes for confirming or contradicting behavior and emit confirm / contradict ops with the episode IDs. Be honest about contradictions — two of them auto-retire the observation, and retiring stale beliefs is the point, not a failure. Explicitly retire anything the user has clearly outgrown.
3. New observations, max 2. Each must pass all three gates:
   • specific (a bug shape, a habit, a trigger — not "weak at X");
   • cited (episode IDs; the script rejects uncited adds);
   • actionable (say which skill consumes it: /teach delivery, /test watch-list, or /next pacing — if none would change, don't add it).
4. Write to memory/.staging/profile-ops.json:

   {"ops": [
     {"op": "confirm", "id": "obs-0001", "evidence": ["ep-0012"]},
     {"op": "add", "kind": "implementation",
      "text": "Off-by-one on right-closed boundaries — 3rd occurrence (windows, binary search hi). Watch in /test; flag boundary lines in /teach implementations.",
      "evidence": ["ep-0009", "ep-0012"]}
   ]}

   Run python3 scripts/update_profile.py.
5. Show the user their active profile as "coach's notes" — plain language, no ids — and one sentence on what changes because of it.