Milestone 1 — Smallest End-to-End Compiler (specification)

Status: M0 proposal (master prompt §23.L). Starts after OD-1/2/5/6/11 are approved.

1. Language subset ("Nova-M1")

• Types: all fixed-width integers (i8–i64, u8–u64, isize, usize), bool, void. No floats in M1 (keeps the kernel subset in reach and avoids float formatting). The full static semantics are in type-system.md.
• Declarations: one file or a few files in one package. namespace, top-level fn (including public/internal), let, var, optional type annotations, const integer constants.
• Statements: blocks, if/else, while, loop, break, continue, return, expression statements, assignment and compound assignment (+= -= *= /= %=).
• Expressions: literals, names, calls (including recursion), unary - !, binary + - * / % == != < <= > >= && || (short-circuit), & | ^ << >> on unsigned types, lossless as, truncate<T>(x), parentheses.
• Built-ins: Console.write(x) / Console.writeLine(x) for integer, bool and string literals only. Process.exit(code: i32).
• Semantics: overflow, division and shift checks panic (OD-6). Panic = message + abort (OD-5). Definite initialisation of var (the M2 init analysis restricted to scalars). Strict left-to-right evaluation. main() -> i32 or main() -> void.
• Excluded: structs, arrays, references, heap, strings as values, generics, imports across packages.

2. Pipeline delivered

Lexer → Parser (with recovery) → AST → name resolution → type check → HIR → NIR (CFG, with CheckedBinOp and Panic terminators) → LIR → C emitter (labelled blocks) → host C compiler → executable. The full IR stack exists from M1, so M2 adds analyses, not a new pipeline.

Runtime: runtime/include/nova/rt.h (types, checked arithmetic, nova_panic declaration), runtime/hosted/rt_hosted.c (stdio + abort), runtime/cosmoos/user/rt_cosmoos_user.c (write(1, …), abort).

3. Acceptance criteria

1. Corpus: ≥ 150 tests in total: ≥ 40 parse-pass/fail (including the full precedence table and recovery cases), ≥ 40 check-fail (types, names, definite init, lossy as, literal out of range), ≥ 50 run-pass (including gcd, recursive fib, Collatz, integer isqrt, an FNV-1a hash of constant input, nested loops with break/continue), ≥ 15 run-fail (each overflow, division, shift and MIN/-1 case with the expected exit status and message).
2. C quality: all generated C compiles with -std=c11 -Wall -Wextra -Wpedantic -Wshadow -Wconversion -Wsign-conversion -Wmissing-prototypes -Wstrict-prototypes -Wvla -Wcast-qual -Wundef -Werror on clang and gcc.
3. Differential: every run-pass test gives identical output under {clang, gcc} × {-O0, -O2} × {no sanitizer, ASan+UBSan} on the macOS (arm64) dev host. Linux x86-64 runs in CI.
4. Cross-compile: the corpus compiles for x86_64-unknown-none-elf and aarch64-unknown-none-elf with CosmoOS USER_CFLAGS (compile only; linking is C1).
5. Determinism: compiling the corpus twice gives byte-identical C. Output is independent of the host path.
6. Diagnostics: every check-fail test asserts the error code and primary span.
7. Docs: grammar.md (complete for Nova-M1), intermediate-representation.md (HIR/NIR/LIR as implemented), diagnostics.md, testing.md updated. Limitations listed.

C1 then reuses the M1 compiler unchanged: hello.nova prints NOVATEST: PASS inside CosmoOS on both arches (roadmap.md).

4. The exact first implementation increment (1a) — done 2026-10-22

Delivered: Nova.slnx (the .NET 10 solution format) with Nova.Compiler, Nova.Cli (novac) and Nova.Compiler.Tests; docs/grammar.md; docs/diagnostics.md; 151 passing tests (see testing.md §4). Deviation from the plan: the corpus runner lives inside the xUnit project for now. A separate Nova.TestRunner is deferred to 1c, when tests need to invoke C compilers. Trivia (comments) is not retained yet; it is needed only for the M10 formatter.

Original plan:

1. Nova.sln with Nova.Compiler, Nova.Cli, Nova.Compiler.Tests, Nova.TestRunner. Directory.Build.props: nullable enabled, warnings as errors, <Deterministic>true</Deterministic>. global.json pinned to SDK 10.0.401.
2. SourceText, Span, Diagnostic and a text renderer.
3. Lexer for the full M1 token set: struct tokens, interned identifiers, literal validation, and keyword table with reserved words.
4. Parser: recursive descent for declarations and statements, Pratt parsing for expressions, recovery at ;/}/declaration keywords. AST plus a --dump-ast printer.
5. Corpus runner for parse-pass/parse-fail with // error: N0xxx annotations. The first ~40 parser tests.

Exit criterion for 1a: parser tests green, AST dumps stable.

Increment 1b — done 2026-10-22. type-system.md (spec first), then:

• the type checker with bidirectional literal typing;
• constant folding with cycle detection (IntegerSemantics defines the exact operator semantics that 1c's C runtime must match);
• definite assignment and missing-return analysis on structured HIR;
• the HIR printer and novac check [--dump-hir].

Semantic analysis runs only on files that parsed cleanly. Two simplifications, both documented: the flow analysis runs on HIR, not NIR (no moves exist yet, so no fixpoint is needed), and all files of a compilation share one item scope.

Increment 1c — done 2026-10-22.

• intermediate-representation.md (spec first).
• NIR (CFG) and the HIR→NIR lowering, with explicit evaluation order and short-circuit branches.
• The C emitter: labelled blocks, checked helpers, zero-initialised locals, deterministic output.
• The runtime: nova/rt.h (freestanding helpers), nova_rt_core.c (formatting, panics) and nova_rt_hosted.c (stdio). They are embedded in the compiler and written next to the generated C.
• novac build [--emit c|exe] [--dump-nir].
• Tests: run-pass/run-fail corpora, the runtime agreement harness, determinism, C snapshots and freestanding cross-compiles.

Deviation: no separate LIR yet (intermediate-representation.md §4).

Acceptance status (§3), 2026-10-22:

#	Criterion	Status
1	Corpus ≥ 40 parse, ≥ 40 check-fail, ≥ 50 run-pass, ≥ 15 run-fail	met: 38 files / 72 cases, 19 files / 99 cases, 51, 19
2	Generated C warning-free under the strict set on clang and gcc	met: Apple clang 21 locally, plus clang 18 and GCC 13.3 on Linux CI (run 39792099920: every run test passed under GCC)
3	Differential: {clang, gcc} × {-O0, -O2} × {plain, ASan+UBSan} agree	met, with one difference: ASan+UBSan runs with clang only. The GCC configurations are -O0, -O2 and portable-overflow
4	Corpus compiles for both CosmoOS triples with CosmoOS USER_CFLAGS	met, and exceeded: all 70 run programs compile with the exact flags, link against CosmoOS libc, and run correctly inside CosmoOS on x86-64 and AArch64 (cosmoos/c1-report.md)
5	Determinism	met (byte-identical C, no host paths)
6	Diagnostics tests assert code and position	met (line, and column where annotated)
7	Docs	met: grammar, type-system, IR, diagnostics, testing

Increment 1d — done 2026-10-22.

• C1 on both architectures, plus the whole run corpus inside CosmoOS (cosmoos/c1-report.md).
• Linux CI (.github/workflows/ci.yml) with GCC 13 and clang 18.
• GCC's -Wclobbered fired on the runtime harness's setjmp use; it is now disabled for that harness only.

Milestone 1 is complete (all criteria in §3 met). 1b (resolution + types + HIR), 1c (NIR + LIR + C + runtime + run tests) and 1d (C1 script) follow.