[Bug]: Sweep result paths collide and silently reuse measurements from another experiment

Finding: In vLLM source main at commit ad73a4740dd780c5620099261738a30b979b262c, a reporter says different benchmark sweep combinations can map to the same result directory, and the cached-result branch then reuses the earlier experiment's measurements for the later one (https://github.com/vllm-project/vllm/issues/60794). This is an external report only; no KFA reproduction was performed. Symptom and example: The reporter's accepted parameters were named experiment/a (num_prompts 10) and experiment_a (num_prompts 20). Both reportedly resolve to the directory BENCH--experiment_a. In the reporter's controlled child-process reproducer, only the 10-prompt benchmark ran. The summary then contained a second row labeled num_prompts=20 with completed=10, copied from the first run. With distinct control names, both benchmarks ran and completed counts were [10, 20] (https://github.com/vllm-project/vllm/issues/60794). Scope as reported: The same path construction and cache pattern is said to exist in serving, workload and startup sweeps. The reporter notes that names unique on each axis can still collide once the axes are joined, so validating raw names alone is insufficient. The reporter's proposed behavior is to reject colliding selected combinations before any process starts or any cached result is read. Pairs excluded by linked-variable selection should not trigger rejection (https://github.com/vllm-project/vllm/issues/60794). Environment: The reporter used source-only Python 3.12 environments on macOS arm64 and Linux arm64, with a managed HTTP health child and deterministic benchmark subprocesses. No model or inference engine was loaded (https://github.com/vllm-project/vllm/issues/60794). Diagnostic implication: If you run sweeps whose names or axis values might sanitize to the same path, check whether the result directories are distinct. A summary row whose completed count matches a different combination is a possible sign of this collision. This is a reporter-derived inference, not a maintainer-confirmed diagnosis. Status: The issue was open with no captured comments or maintainer triage. The reporter points to candidate PR #60795 and says related PR #60328 validates names only after str() conversion, which leaves the final-path reproducer failing. No merge or release is evidenced in the supplied sources. Reporter-claimed validation of the candidate guard (9 failures against the original modules, 41 passes with the guard) is self-reported, was not reviewed, and was not verified by KFA. Limitations: This is a single report, partly model-assisted, on benchmark planning only, with no GPU or model evaluation.