Automated reference image generation from text briefs for `img2threejs`.

## Motivation

Right now you must hand the pipeline a reference image. For real projects you often start from a **text brief** or a partial asset list, not photos. An optional front-end that produces candidate reference images would let teams go from "I need a stylized loot chest" to a procedural Three.js model without sourcing art first.

## Scope

- An **opt-in, provider-agnostic** step: given a text prompt or a project/asset spec, generate or source candidate reference images, then feed the best one into the existing `probe → assessment → spec → generate` flow.
- Keep it pluggable — the image source (a text-to-image model, an asset search, or a user folder) should be behind a small adapter, not hard-wired to one provider.
- Surface licensing/provenance of any sourced image so nothing unlicensed slips into a build.

## Acceptance criteria

- `--from-prompt ""` (or equivalent) yields a candidate reference the current pipeline consumes unchanged.
- The adapter interface is documented so a contributor can add a new image source in one file.
- Provenance/licensing is recorded in the assessment artifact.

## Open questions (discussion welcome)

- Which providers to support first, and how to keep the core dependency-free?
- How to rank/choose among multiple candidates automatically vs. asking the user?

**Difficulty:** medium–hard · research-flavored. Comment with design ideas before implementing.