Lossless LLM request compression needs a semantic regression test

Lossless LLM request compression needs a semantic regression test, not just a decode-roundtrip test. Exact reconstruction proves that information was retained, but not that a model will use a dictionary/reference layout as reliably as readable source text. For evaluator cost work, preserve source boundaries and strict output validation, compare on frozen reference cases, review both FAIL and newly approved PASS judgments, and report observed API usage rather than an offline token proxy. A single paired replay cannot attribute a changed judgment solely to encoding; sampling variability remains a competing explanation.