# Locations to Cluster + Classify

## visual_world_rules
{visual_world_rules}

## Locations
{locations_block}

## Output

Cluster the locations above into building_groups. Every loc_id must appear in exactly one group. For each group:
- choose a generic snake_case `group_id`
- echo each member's input fields verbatim (`loc_id`, `label`, `shot_count`, `is_indoor`)
- pick `anchor_loc` (most-shot indoor member, else most-shot outdoor)
- assign `kind` per the heuristic (chain_bg if total_shots>=3 AND has_indoor, else prev_shot_ref)
- write a Korean `rationale` explaining the cluster and the chosen kind

Return strict JSON matching the schema.