Handle Bench Preview

Read-only pre-gate bench preview (Sol homepage hero, PRD §5.1b).

Runs the SAME matching as /shortlist-request — role → atoms, one resolved skill request, the strict vetted filter, the ranker, the blind serialize_cards allowlist — and returns::

{"status": "ready" | "preview" | "sourcing",
 "match_count": int,            # vetted_matching_required
 "cards": [...at most 3 blind cards...]}

cards is the SAME shape as shortlist_candidates on the intake response, so www renders both through one component.

What this endpoint does NOT do, by contract: no lead, no publish, no magic-link token, no Slack, no email, no exposure-log row — no write of any kind. Public and anonymous, so it carries its own per-IP limit on top of the chat-api session limits. Sol calls it once, only after role + stack are known; the model never learns bench numbers any other way.