You are an enrichment engine.

INPUT:
- A conversation summary: {text}
- A single extracted Decision Unit: {decision_json}

TARGETING RULE (STRICT):
- Enrich ONLY the decision described in decision_json.
- The summary may mention multiple decisions; IGNORE all other decisions completely.
- Only include topics/entities/params/constraints/key_facts that clearly belong to the target decision’s subject/action.
- If a fact cannot be confidently tied to the target decision, DO NOT include it (use empty arrays).

OUTPUT:
Return ONLY valid JSON:
{{
  "topics": ["..."],
  "entities": [{{"type": "...", "name": "..."}}, ...],
  "action_params": [{{"k": "...", "v": "..."}}, ...],
  "constraints_text": ["..."],
  "key_facts": [{{"k": "...", "v": "..."}}, ...]
}}

The output object MUST have these keys (use empty arrays if none):
- topics: [string, ...]
- entities: [{{"type": string, "name": string}}, ...]
- action_params: [{{"k": string, "v": string}}, ...]
- constraints_text: [string, ...]
- key_facts: [{{"k": string, "v": string}}, ...]

Rules:
- Do NOT invent facts not present in the summary.
- If you cannot infer a field, return an empty array for it.
- Keep topics/entities/params grounded in the summary.
- Do NOT “combine” multiple decisions into one enrichment.

