Budget first, then fill it
Most systems retrieve a fixed k and hope it fits. That produces prompts that vary in size by a factor of three depending on chunk length, which makes cost unpredictable and occasionally overruns the window.
Packing to a budget inverts it: decide what you can afford, then fill that space with the best available chunks. Cost becomes a constant you chose rather than an outcome you observe.
The score threshold does more than the budget
A chunk scoring 0.22 is not weakly relevant, it is irrelevant — and including it makes the answer worse as well as more expensive. A minimum score is the cheapest quality improvement in a RAG pipeline.
If the threshold is dropping chunks you needed, the problem is upstream in retrieval or chunking. Packing cannot fix bad candidates.