Heretic 1.3 adds reproducibility, benchmarks

Heretic is evolving from a clever model-editing tool into a more serious research pipeline: you can now track what produced a result, measure whether it damaged the base model, and fit larger architectures with less memory overhead.

• –The new `reproduce` directory is the biggest upgrade because it captures the environment details needed for byte-for-byte reruns, which matters a lot for GPU-dependent tensor ops
• –Built-in lm-eval-style benchmarking removes a lot of friction when deciding whether a trial is publishable or worth iterating on
• –Peak VRAM reductions are practical, not cosmetic; they let more users run larger models on the same hardware
• –Broader layer/module handling is what keeps the project relevant as model families keep changing underneath it
• –The optional upload flow for reproducibility data is a good trust signal because it keeps publishing under user control