Use go-tiktoken with embedded vocabularies for accurate BPE token counting. CountPromptTokens applies conservative padding so returned context token counts stay slightly above the real value instead of undercounting.