Calibrate Knight spell effects from observed casts

[adrunkhuman]

Scope

Use verified #91 casts to build confidence-bounded effect profiles. Semantic role and engine legality stay fixed; observations rank legal actions only.

Track healing by spell, damage by spell and target class when attribution is unambiguous, support duration, mana spent, and cast failures. Start with controller-local profiles and JSONL evidence. Do not add durable learned-state storage unless the progression benchmark requires it.

Acceptance

• Profiles record sample count, bounded range, and conservative expected value.
• Healing samples account for missing health and concurrent damage.
• Damage samples exclude ambiguous melee, other-attacker, and target-loss events.
• Low-confidence profiles retain the Use legal Knight spells in combat and recovery #91 safe policy instead of changing action choice.
• Observations change utility gradually and never bypass engine checks, reserves, or predicted-lethal limits.
• Deterministic fixtures cover healing, damage, support duration, ambiguous observations, and reset behavior.
• Telemetry exposes samples, confidence, ranking input, and rejected evidence.

Related: #91, #92.