When Leaderboards Move: A Pre-Specified Two-Axis Audit Protocol for Code LLM Rankings under Equivalent Prompts and Stronger Test Suites

Code-generation leaderboards compress a multi-stage measurement process into a single score, even though the result depends on both the wording of a programming task and the tests used to decide functional correctness. This paper specifies a controlled two-axis audit that varies those choices within one paired design. HumanEval+ and MBPP+ define the task and oracle universe; each eligible task is represented by three fixed ReCode formulations (nominal, synonym substitution, and back-translation), and every generated program is judged under the original benchmark tests and the stronger EvalPlus tests. Four open base code-model families in the 1.3B–3B range are frozen at exact revisions and evaluated with deterministic, unquantized FP16 direct completion. The primary estimand is the prompt-by-oracle interaction: whether measured sensitivity to equivalent task formulations changes when correctness is tested more rigorously. Secondary analyses cover oracle disagreement, rank displacement, Kendall rank stability, pairwise reversals, and task-level volatility. Exclusions, compatibility gates, 10,000 task-level bootstrap replicates, a 32-test exact McNemar family with Holm correction, and interpretation rules are fixed before execution. The contribution is a reproducible measurement protocol for auditing the stability of code-generation benchmark conclusions.