Question: the surface is sugar over the semantics (formatter, linter, pseudoinstructions). The principle is set: never add a kernel primitive for something expressible as a composite of existing primitives (invariant #5) — instead provide aliasing + metaprogramming that expands to primitive composites (like assembly pseudo-ops). Open: the mechanism — how macros/aliasing work, and how much syntactic extensibility the surface offers.

Why it matters: this is "write your own constructs" from the vision, and the discipline that keeps the kernel at five primitives as the surface grows. Get it right and new paradigms/notations are libraries; get it wrong and the kernel bloats or the surface fragments.

Detail: levels of extensibility — (a) plain aliasing (a name for a composite, no new syntax); (b) hygienic macros that expand to core terms before lowering (Lean 4 elaboration, Racket define-syntax, Scheme); (c) full user-defined notation / reader extension (custom operators, mixfix — Lean notation, Agda mixfix). Hygiene (capture-avoidance) interacts with ADR-0006/0020 (no implicit capture; de Bruijn). The semantic DSL mechanism already exists (effects + handlers = a little language per effect); this Q is about syntactic extension on top.

Options: (1) elaboration-style hygienic macros expanding to core Comp (recommended; Lean 4 model — composes with the existing lowering pass in Bang/Frontend/Surface.lean); (2) aliasing only (no new syntax — minimal, may be too weak for ergonomic DSLs); (3) full reader/notation extension (most powerful, most rope). The five-primitive invariant + "no new primitive if composite" is the constraint; the mechanism is the choice.

Blocked on: nothing now — a surface-layer concern (liquid); meaningful once the surface grows past the rung-0/1 toy parser.

Revisit signal: the surface accumulating repeated composite patterns that want a name; or building the first user-defined construct/notation.