Bounded vocabulary — the set of names the parser is allowed to intern as atoms.

More precisely: this is the set of source-text names the analyzer/evaluator currently pattern-matches as atom literals. Builtin function names + special forms + bounded namespaces + destructuring modifiers + qualified analyzer keys + short-fn param atoms. Everything else stays as a binary in the AST so the global atom table never grows from user input (issue #953).

Why an explicit allowlist instead of String.to_existing_atom/1: the global VM atom table is non-deterministic — unrelated modules loading later can change how the same source parses. Codex's pushback on the bug thread covers this in detail.

What's in the table

1. Every env-dispatched builtin name from PtcRunner.Lisp.BuiltinNames.env_names/0 — all builtin functions (map, filter, +, str, etc.). These equal the keys of PtcRunner.Lisp.Env.initial/0 but are derived from the compile-time registry so SourceAtoms stays out of the Lisp runtime cycle (issue #1051).
2. Analyzer special forms — let, fn, def, if, case, etc. Only forms that the analyzer currently dispatches on. No aspirational Clojure entries.
3. Bounded keyword modifiers used by for/doseq/destructuring — :else, :keys, :as, :or, etc.
4. Bounded namespaces — data, tool, json, plus Clojure aliases (clojure.string), and fully-qualified Java namespaces from BuiltinNames.java_namespace_atoms/0 (projected from priv/java_interop.exs) (java.time.LocalDate, etc.).
5. Qualified analyzer keys such as JSON member names plus atom-named Java namespace members projected from priv/java_interop.exs.
6. Short-fn param atoms :p1..:p20 synthesized by the short-fn analyzer.

What's NOT in the table

Any spelling not listed above — which in practice means most user-chosen names: let bindings, fn params, def names, custom keywords like :my_kw, namespaced keys like data/foo_42. These stay as binaries in the AST.

Membership is a property of the spelling, not of what the name is used for. A user definition whose spelling happens to be in the table — (defn- parse …), (let [text …] …) — interns as an atom just the same, so a single form routinely mixes atoms and binaries. Anything comparing definition names against reference names must normalise first; comparing raw representations reported such definitions as undefined (#1166).

Cache

Table is built lazily on first call and cached in :persistent_term. Read cost after first call is one :persistent_term.get/1 (no copy).