Positioning: answer "why not just prompt it", and separate verified from taste

[codeswithroh]

Closes #11.

Why

Two fair criticisms from the feedback round, both about how the project presents itself rather than what it does:

You can do this with just a conversation and "ok write these decisions as a design bible and use it as an anchor as we move forward."

The only verified thing in the repo is the contrast script, which tbh has nothing to do with taste, that's just accessibility best practice.

Both land. The README was leading with the weakest part (five presets) and letting an accessibility check stand in as proof of taste. This fixes the framing rather than the code.

Changes to README.md

1. Leads with the check that actually runs. The core ideas went from three to four, and now open with the contrast contract from #8. A rule that executes is the one thing a prompt genuinely cannot reproduce, so it should be the first thing a reader sees, not the presets.

2. New section: "Can I not just tell the AI to write the decisions down?" Answers the skeptic directly, and concedes the true part first: for keeping three screens consistent inside one chat, a conversation works and you do not need this skill. Then the four things it cannot do:

• survive the session (the lock and profile are files on disk, not context)
• run a check (a written preference is a judgment you can talk yourself out of; the matrix is a computation)
• read pixels (a conversation turns a reference into a description, then rebuilds from the description)
• enumerate legal combinations (a written bible lists colors, it does not say which may touch which)

3. New section: "What is verified, and what is judgment." Draws the line the feedback asked for:

• Verified: contrast. A computation, explicitly accessibility and not aesthetics. Says outright that a palette clearing every ratio can still be ugly, and that the reason it earns its place is that it catches a failure eyes cannot run.
• Judgment: reference extraction, mood matching, the profile, the anti-slop checklist. Informed defaults and accumulated preference, not proof, and the README now says so.

It also admits the project had been blurring the two, which seemed better than quietly fixing it.

4. Features table clarified: "verified" now states it means ratios were computed, not that taste was proven, and the contrast row describes the matrix rather than a one time check.

Verification

• Zero em-dashes (project copy style).
• All nav anchors validated against actual headings, including the two new sections.
• Rendered and read through in full.