Token-Budgeted Context Packer

A retriever returns fifty passages. The model has room for maybe eight. Something has to choose, and getting it wrong is how a RAG system silently loses the answer it already found.

Implement pack_context(passages, budget):

• passages is a list of (id, text, score), already ordered best first
• The cost of a passage is len(text.split()) tokens
• Greedily include passages in order while they fit in the remaining budget
• Skip a passage that does not fit and keep trying later ones: a single enormous passage must not end the packing
• Return the list of included ids, in the order they were included

The skip-rather-than-stop rule is the whole problem. Stopping at the first passage that does not fit means one 2,000-token document at rank 3 discards ranks 4 through 50, including the passage that actually answers the question.