feat: infer a project persona from the hero when there's no audience context

[gabelul]

Small gated fallback for persona sourcing, decided on evidence rather than theory.

The gap

The tailored project persona comes from an audience/brand description captured at setup. When a project has none (no README audience, thin docs), the scan falls back to generic built-ins — even on a landing page whose hero is a clear pitch to a specific user.

The fallback

Step 6a now: if Step 5b produced no project persona and the desktop hero pitches a specific, nameable audience, infer one persona from the hero, tagged as an unconfirmed hypothesis (inferred-visitor). It self-gates — a bare search box, app shell, or generic splash produces nothing, because a hero that isn't pitching an audience can't tell you who it's for.

Priority order: explicit audience text (always wins) → hero inference (this) → built-ins. A guess from pixels never overrides a real audience description.

Why gated, not universal — the evidence

I captured real heroes and had the vision model infer a persona from each, cold:

Hero	Inferred persona	Verdict
Linear ("product development system for teams and agents")	shipping-fast eng/product lead evaluating team fit	Sharp — beats the built-ins
Stripe ("financial infrastructure to grow your revenue")	technical founder scaling payments	Good, but picks one of its two audiences
Airbnb (a search box)	generic "planning a getaway"	No better than built-ins — correctly produces nothing

So the real gate isn't brand-vs-product, it's "is the hero pitch-forward?" The model makes that call itself.

Testing

Full suite green (1121 tests, 0 fail). Fallback wiring guarded in vision-personas.test.js (gated on missing context, pitch-forward hero, hypothesis tag, reads the desktop screenshot).