Implement streaming enrichment aggregator

You are building a real-time document enrichment aggregator for a Kafka-like event stream. Each incoming event has the following fields: - `document_id`: string - `enricher_id`: string - `enrichment`: any JSON-compatible value Example events: ```json {"document_id": "doc1", "enricher_id": "Sentiment Analysis", "enrichment": "positive"} {"document_id": "doc2", "enricher_id": "Entity Linking", "enrichment": ["AAPL"]} {"document_id": "doc1", "enricher_id": "Topic Classification", "enrichment": ["AUTO", "ANALYST_CHANGE"]} {"document_id": "doc1", "enricher_id": "Entity Linking", "enrichment": ["TSLA"]} {"document_id": "doc3", "enricher_id": "Summarization", "enrichment": "Yesterday the Yankees won against the Mets..."} ``` The aggregator is configured with a required set of enrichers, for example: ```python ["Sentiment Analysis", "Entity Linking"] ``` Implement an in-memory Python aggregator with the following behavior: 1. Consume events one at a time. 2. Ignore events whose `enricher_id` is not in the configured required set. 3. For each `document_id`, keep the latest enrichment value for each required enricher. 4. As soon as a document has received all required enrichments, publish exactly one aggregated output for that document. 5. If 10 seconds have passed since the first relevant event for a document and not all required enrichments have arrived, publish the partial result exactly once. 6. After a document has been published, any later events for that same document should be ignored. 7. After publishing, clean up any in-memory state for that document. 8. Assume there is no background thread. Timeout handling should be done lazily when new events arrive, or through an explicit timeout-check method. The published output should have this form: ```json { "document_id": "doc1", "enricher_ids": ["Sentiment Analysis", "Entity Linking"], "enrichments": { "Sentiment Analysis": "positive", "Entity Linking": ["TSLA"] } } ``` Write the Python implementation and explain any important edge cases you would consider.