iyi 0.1.0

The first release. There is nothing to compare it against, so this says what is
in it rather than what changed, and what a later version will have to keep
faith with.

The language

Four rules, and everything else follows from them (SPEC.md):

• R-1 A module is the unit of compilation. import forms a DAG, and
  compiling a module reads its imports' declarations, never their bodies.
• R-2 Everything a module exports (pub) writes down full parameter and
  return types.
• R-2b using brings exported names into unqualified scope, written by the
  consumer.
• R-3 No open classes. impl Trait for Type lives in the module that
  declares the trait or the one that declares the type.

Traits with defaults and associated types, generic impls with forall, errors
as ordinary union members with ! propagation, defer scoped to the block,
and Crystal's syntax otherwise: union types, nil-safety, blocks, local
inference, macros.

The artifact

.iyimod format v19. A module's declarations, its macros, the bodies that
have to travel, its object code, and a checksum per section. Artifacts are
version-locked: one written by another build of the compiler is refused and
rebuilt, never migrated, so a later version bumping this number is expected
rather than a breakage.

iyi build --emit-iyimod DIR writes them, --use-iyimod DIR builds against
them, and iyi mod dump prints one as text.

The tool

iyi takes build, run, mod, env, clear_cache, tool, version and
help. crystal is the same compiler under its own name, and it still
compiles .cr files. eval is deliberately not on iyi's list: it has no
filename, so it gets Crystal's prelude and Crystal's rules, and a command that
answers in another language is worse than one that says where it went.

Measured

One machine, release compiler, best of seven, seconds. A 30-module,
7,208-line project, rebuilt after changing one line in one module:

	iyi	Crystal	go build
rebuild after one edit	0.13	1.17	0.16

The same edit with every module read from source instead of from artifacts
costs 0.23 s, which is what R-1 is worth on this project. A full build of a
6,900-line program from scratch is 0.24 s against go build's 0.09 s, which is
where iyi loses. python3 bench/incremental.py and python3 bench/build_speed.py print both, and refuse to time programs that do not agree
on their output.

Not in this release

No IO beyond puts. No concurrency: SPEC.md III.4 specifies it and none of it
is built. No package manager, no standard library, no self-hosting. Linux
x86-64 only. derive macros do not cross modules. The prelude is 1,184 lines
and its collections are small, and a[-1] raises rather than indexing from the
end. The formatter is Crystal's and does not know iyi's syntax: iyi tool format says so and leaves .iyi files alone.

Provenance

A fork of Crystal at 1.22.0-dev,
Apache 2.0 with Swift exception, Copyright 2012-2026 Manas Technology
Solutions. The backend, the GC and the type checker are Crystal's work. The
compiler reports itself as Crystal 1.22.0-dev because that is what it is, and
it is bootstrapped by released Crystal 1.21.0, which is the version CI pins
and the one to install if you are building this from source.

Two bugs in Crystal's own compiler were found here and fixed in this fork; they
belong upstream and are separate commits for that reason:

• Crystal.relative_filename chopped the working directory off any path that
  merely began with its name, so a build in /x/crystal with a cache in
  /x/crystal-cache wrote its object files to -cache/….
• The cache cleaner deleted the directory of a build that was still running,
  because it keeps the ten most recently modified directories and a build stops
  looking recent while its units sit in an optimization pass.