2026-08-11 — Ranking follow-up from the fourth trial's tenant: sigil-aware exactness, publics first

Shown the shipped qualified match, the fourth trial's tenant confirmed the zero-recall spelling — both partial and fully qualified forms landed without an option map — and stress-tested the ranking into one refinement with two halves. For a stem query, recall was perfect but the head of the list was the decorated family in source order; to a Clojure eye claim?, ->Claim and -Claim are the exact matches — the decoration is convention, not name. And a private helper outranked its public sibling on line order alone, against the intuition that the API is usually the quarry. Both built the same day: a closed conventional affix set — one leading map->/->/-, one trailing ?/!/*, earmuffs as a pair — is stripped candidate-side before the exactness comparison, admitting decorated names to the exact tier one rung below exactness as spelled (a query that spells the sigil raw-matches without stripping); and within every tier publics now rank before privates, namespace first, source line last, the private sibling one row down and still flagged. One exception settled during the build: a bare scope listing ("helpers/") keeps whole-namespace source order — a listing, not a ranking — a distinction the suite's empty-name-side pin surfaced before it could ship wrong. Verified live on this JVM: "out" now answers out, then *out*, then the name hits.

The stress test also showed vars doubling as a surface linter: a macro-interned, gensym-named public factory var surfaced in a record namespace's results — a wart the owning macro can fix with ^:private, caught precisely because the tool refuses to hide what is interned. Electivity remains untested — the tenant was again pointed at the tool — but the spelling has now survived two independent first contacts.