iterate-fat-llama

Before doing anything else, read `.claude/skills/rules/iterate-fat-llama.md` in full — it defines the cycle cap, score formula, naming conventions, version-bump policy, and remote-action confirmation requirement referenced throughout the steps below, and may be updated over time without this file changing. Also read `.claude/rules/scope-and-safety.md` for the filesystem write scope and safety boundaries every skill/agent follows, and `.claude/rules/project-mission.md` for what fat_llama actually is and how it works. Also read `README.md` at the repo root in full, right now, before Step 0 — it's fat_llama's own description of its purpose and method (iterative soft thresholding of FFT data to upscale compressed audio across supported formats, tested and built primarily against the MP3→FLAC outcome, CUDA-only, deliberately without AI/ML-based upscaling). Carry that context — condensed per `.claude/rules/project-mission.md` — into every `generate-code` dispatch in step (d) below and into the docs/changelog/PR steps that follow, so nothing drifts toward an out-of-scope "fix". `args` is optional free-text directives on what to fix or not fix (e.g. "focus on bitrate accuracy, don't touch the CLI"). Handle it per the DIRECTIVES rule in `.claude/skills/rules/iterate-fat-llama.md`. This skill coordinates other skills/subagents — it does not review, test, or fix anything itself. It also touches git branches, tags, and (at the very end, with confirmation) the GitHub remote — read Step...