Source loading and modules

Before func can be a keyword, it is four bytes from a file. The source stage proves those bytes are safe UTF-8, gives the file an identity, and remembers enough position information to point back to a bad character later.

Preparing one source file

1. Bound the inputReject oversized or binary-looking input before later stages allocate structures proportional to it.
2. Validate encodingAccept valid UTF-8, ignore a leading UTF-8 BOM, and reject NUL or unsupported encodings.
3. Normalize line endingsSupported CRLF becomes LF; a stray carriage return is diagnosed rather than guessed around.
4. Register the fileOwn its bytes, assign a source identity, and build line indexes used by every later diagnostic and trace.

The result is not merely a string. It is text with an identity and a coordinate system. Later stages keep byte spans; the source registry turns those spans into path, line, and column only when a human-facing report needs them.

Loading the module graph

A module is loaded once. Cycles between modules are allowed because Luce has no module initialization phase; a semantic cycle that implies an impossible value layout or recursive constant is rejected later by the stage that understands that meaning.

The host supplies files; the compiler supplies meaning

The compiler asks a compile.Loader for source by logical module request. It never opens a path itself. The command-line app can implement that request with a filesystem and manifest; a test can answer from memory; an editor can compile a buffer that has never been saved.

Positions survive every transformation

Tokens carry source spans. AST nodes combine them. HIR records the span of the checked operation. MIR origins freeze source, line, and column onto emitted instructions. Debug artifacts retain those origins for traps; release artifacts may strip line detail while keeping function names.

Decisions kept here