v0.2.0 — Prompts and scoring

The prompt is how you tell Carl what you actually want, and this release is what makes it count.

Everything a source returns is embedded and ranked against your prompt, a cheap model scores what survives, and a more expensive one re-scores the best of those and writes the relevance note explaining why each finding earned its place. Each step costs more and handles fewer resources, so a brew spends where it matters.

Carl's own prompts are versioned markdown rather than strings buried in code, with tracing on every model call.