Skip to content

module-prelude.md

Latest commit

 

History

History
32 lines (22 loc) · 4.21 KB

module-prelude.md

File metadata and controls

32 lines (22 loc) · 4.21 KB

< All Decisions

Module Prelude — Design Rationale

Panel Deliberation

Five panelists (systems, web/scripting, PLT, DevOps/tooling, AI/ML) voted independently on 3 questions. Each expert ran as a separate agent and returned votes without seeing other experts' reasoning.

Q1: Which types belong in the prelude (3-1-1 for types + ADTs + traits)

  • Systems: Full set including ConversionError, Range, Handler (Option D). Every type in A through D is already compiler-intrinsic. The compiler must know their layout, vtable shape, and monomorphization behavior. Forcing imports for items the compiler already knows about adds zero information and wastes tokens. (dissent — wanted D)
  • Web/Scripting: Types + ADTs only, no traits (Option B). Every mainstream language makes primitives and core collections available without imports. But traits in the prelude go too far — they're an intermediate concept newcomers shouldn't need to encounter in the prelude. (dissent — wanted B)
  • PLT: Types + ADTs + traits (Option C). == desugars to Eq.eq, < to Ord.cmp, for through IntoIterator/Iterator, interpolation through Display, @derive references traits by name. If these traits aren't in scope, every file needs boilerplate imports for basic language features. Stops at C because ConversionError, Range, Handler are used by specific subsystems, not core typing rules.
  • DevOps: Types + ADTs + traits (Option C). LSP autocomplete for x. needs all compiler-known traits in scope to show methods. Error messages referencing Eq or Ord should never say "did you forget to import Eq?" — that's a broken diagnostic experience. ConversionError/Range/Handler are importable without degrading day-to-day tooling.
  • AI/ML: Types + ADTs + traits (Option C). Every import is a failure point for LLMs. Primitive types + core traits have near-100% usage in training data without imports. Forcing imports for Int or Eq would create a pattern with zero training data. ConversionError/Range/Handler are niche enough to import.

Q2: true/false — keywords or prelude values (5-0 for keywords)

  • Systems: Keywords. The compiler emits direct boolean instructions — no vtable, no trait resolution, no name lookup. Making them keywords reflects what the hardware sees.
  • Web/Scripting: Keywords. In every language web devs know, boolean literals are keywords. let true = 5 should give "cannot use keyword as identifier," not a confusing shadowing error.
  • PLT: Keywords. true and false are introduction forms for Bool, like 42 for Int. They participate in exhaustiveness analysis. The lowercase spelling already signals "keyword, not constructor" — Blink constructors are PascalCase.
  • DevOps: Keywords. Keywords get syntax highlighting for free. No "unknown identifier" errors. No import needed. Clean diagnostic story.
  • AI/ML: Keywords. true/false are keywords in 95%+ of training data. Zero ambiguity, zero decision points.

Q3: Test builtins — auto-available or require import (5-0 for auto-available)

  • Systems: Auto-available. Test builtins are compiler intrinsics — assert_eq needs source locations and display values at compile time. They're not library functions. Test code is dead-code-eliminated in release builds regardless.
  • Web/Scripting: Auto-available. Jest, pytest, Go testing — all have built-in assertions. Requiring imports adds ceremony to the simplest tests. Cap auto-available set at these four; richer testing utilities require imports.
  • PLT: Auto-available. test blocks are a special syntactic form with their own scoping rules. assert/assert_eq/assert_ne are observation forms of the testing judgment — they need no import, same as match needs no import for pattern elimination.
  • DevOps: Auto-available. The compiler already knows about test blocks. Built-in assertions should produce compile errors when used outside test blocks — this requires compiler awareness, not library code. LSP can show "only available in test blocks" hints.
  • AI/ML: Auto-available. In training data, test assertions are never imported. LLMs would forget the import at high rates. Zero-friction test authoring is critical for adoption.