preflight catches PyTorch training bugs before run

Silent training failures are one of the most painful failure modes in ML, and preflight-ml is a simple, opinionated answer to a gap that bigger tools like Deepchecks don't neatly fill.

• –Ten checks across fatal/warn/info severity tiers catch the silent bugs (NaNs, label leakage, wrong channel ordering) that waste GPU hours without ever throwing an error
• –CI integration via exit codes is the right call — this belongs in the pre-training pipeline, not as a post-hoc debugging tool
• –Operates on sampled batches (~30 seconds), so it's fast enough to be a default step without adding meaningful overhead
• –No direct competitor occupies this exact niche: Deepchecks is heavier, Great Expectations is data-only, PyTorch Lightning embeds similar checks but only inside its own framework
• –v0.1.1 alpha with a clear roadmap (drift detection, auto-patching, domain plugins) — early but genuinely useful even now