The paragraph-based chunker needs one additional size guard: a single paragraph can exceed max_tokens, because it is appended after the overflow check and is never split. That matters when the embedding endpoint truncates long inputs without rejecting them.

The stored chunk can then contain a relevant sentence that never contributed to its vector, making the failure look like weak semantic retrieval. I would add a long-paragraph fixture and assert the final token count using the embedding tokenizer, including any contextual prefix. Record truncation explicitly or reject oversized chunks before indexing, so the text you inspect is the text the embedding model actually saw.