[feat] AI-generated code quality score — show confidence and diff clarity before applying changes

Problem

When Onlook's AI generates or suggests changes to React components, there's no signal about how confident the model is, or how "clean" the diff is — whether it's a minimal targeted change or a large rewrite that touches many unrelated lines.

As a user, I want to know:

• Is this change high-confidence (model is sure) or exploratory?
• Is the diff surgical (changes only what I asked) or noisy (rewrites unrelated code)?
• Are there parts the model flagged as uncertain or that need a human review?

Proposed Solution

Add a lightweight AI change quality indicator to the diff preview before the user applies a change:

┌─────────────────────────────────────────────────────┐
│  AI Change Preview                                  │
│  ────────────────────────────────────────────────── │
│  Confidence: ████████░░  82%   Scope: Surgical ✓   │
│  Changed: 4 lines   Unchanged: 96%                  │
│  ⚠️  1 line flagged for review (className conflict)  │
└─────────────────────────────────────────────────────┘

Metrics to surface:

• Confidence score: model's self-reported certainty (if using a model API that supports it) or a proxy (diff size vs. requested change scope)
• Scope ratio: lines changed / total lines in component — lower is more surgical
• Flagged lines: lines where the model inserted a comment like // TODO: verify or where a conflict was detected
• Change type: style-only, logic, refactor, or mixed — helps the user know what to review

Why This Matters

Right now, applying an AI suggestion in Onlook is a trust leap. Users can't tell if a change is a 2-line style tweak or a structural refactor that might break behavior. A lightweight confidence/clarity signal before applying would:

• Reduce accidental breaking changes
• Build user trust in AI-generated edits progressively
• Signal to users when to inspect before applying vs. auto-accept

This is especially important for less experienced users who may not be able to read the diff itself.

Implementation Notes

• Most of the metadata is already computable locally from the diff (no extra API calls needed for scope ratio, changed lines, flagged comments).
• Confidence score could be approximated by diff size relative to the user prompt length.
• This could ship as a small info bar above the diff, off by default with an opt-in setting.

Happy to contribute a PR for the UI component + the diff analysis logic if the team is interested.