What problem does it solve? Teams often adopt GPU, cloud, Dask, Polars, or vectorization changes for slow topological data analysis jobs without evidence, wasting effort or silently corrupting results. This Skill enforces a profile-first discipline that proves an acceleration is justified before adoption. ## Core Features & Use Cases - Profile-First Gating: Refuses acceleration proposals until the baseline is profiled and the bottleneck is classified (CPU, memory bandwidth, I/O, GIL, backend, or algorithmic complexity). - Rigorous Acceptance Gate: Requires stage parity on at least 3 representative inputs, material wall-time gains at realistic scale, numerical equivalence within a declared tolerance, passing contracts, and surviving checkpoint/resume behavior. - Worker-Sweep Benchmarking: Mandates worker-count sweeps, execution-locus tables for thread-based candidates, and recorded environment drift and provenance. - Use Case: A TDL job running exact Wasserstein-2 distances is slow at scale. Use this Skill to profile the bottleneck, discover memory-bandwidth contention under loky, and validate a serial process-pool alternative with bit-for-bit output equivalence before accepting it. ## Quick Start Ask the AI to evaluate whether moving a slow TDA job to GPU or a process pool is justified, requiring profiling, a worker-sweep benchmark, and numerical equivalence checks before acceptance.
