Episodes: improve Anthropic JSON mode with explicit schema instruction

[abbudjoe]

From PR #196 review (Issue #11).

The LLM client uses a generic "Respond with valid JSON only." instruction for Anthropic's Claude models. Claude responds better to explicit JSON schema instructions.

Current:

if json_mode:
    prompt += "\n\nRespond with valid JSON only."

Recommended:

if json_mode:
    prompt += "\n\nYour response must be valid JSON matching this exact format:\n{...schema...}"

Provider-specific JSON handling would improve classification reliability.

Part of Episode Memories (#190).