Skip to content

CHANGELOG.md

Latest commit

 

History

History
457 lines (398 loc) · 23.6 KB

CHANGELOG.md

File metadata and controls

457 lines (398 loc) · 23.6 KB

Changelog

All notable changes to zwasm are documented here. Format based on Keep a Changelog.

[Unreleased]

Developer / CI infrastructure improvements. No public API or runtime behaviour change for embedders.

Added

  • .github/versions.lock (mirrors flake.nix pins for Windows + CI YAML) replaces the old .github/tool-versions. Single source of truth for Zig, wasm-tools, wasmtime, WASI SDK, Rust pins; comments must live on their own line per file header (the Python reader in ci.yml does a plain split('=', 1)). (#60, D136)
  • .dev/environment.md — developer onboarding doc covering Mac / Linux / Windows native setup, Nix devshell contents, CI ↔ local gate mapping, and the remaining Windows-skipped CI items as a Plan C tracker. (#60)
  • scripts/lib/versions.sh, scripts/sync-versions.sh, scripts/gate-commit.sh, scripts/gate-merge.sh, scripts/run-bench.sh — unified Commit Gate / Merge Gate runners that work identically on macOS, Linux (Nix devshell), and Windows (Git Bash). gate-commit.sh auto-clones wasmtime tests/misc_testsuite into a gitignored .cache/ directory and chains build_all.py && run_compat.py for realworld so a fresh checkout works without manual setup. Auto-skips ffi on Windows to mirror the existing CI guard. (#61)
  • scripts/windows/install-tools.ps1 — provisions Zig + wasm-tools + wasmtime + WASI SDK from versions.lock into %LOCALAPPDATA%\zwasm-tools. Updates user PATH + WASI_SDK_PATH. Auto-installs Microsoft Visual C++ Redistributable via winget when vcruntime140.dll is missing (WASI SDK clang.exe needs it). Idempotent; --Force re-extracts; -OnlyTool selects one. (#61)
  • versions-lock-sync CI job mechanises Merge Gate item #9 — fails the PR if flake.nix and versions.lock disagree on a tool pin. Runs in parallel with the existing test matrix in well under a second. (#62)
  • Memory usage check on Windows via PowerShell Process.PeakWorkingSet64 — first of the eight Windows-skip CI guards removed. Same 4.5 MB budget as the POSIX path. (#64)
  • zig build shared-lib now runs on Windows in CI. Zig produces zwasm.dll + zwasm.lib natively from addLibrary({.linkage = .dynamic}) — the old guard was a no-op. Plan C-a. (#68)
  • zig build static-lib -Dpic=true -Dcompiler-rt=true and test/c_api/run_static_link_test.sh now run on Windows in CI. The C link tests use zig cc (portable across Mac/Linux/Windows) instead of system cc. PIE coverage is preserved on Linux. Plan C-d. (#69)
  • examples/rust cargo run now runs on Windows in CI. build.rs gained a Windows arm: dynamic linking copies zwasm.dll next to the cargo target binary so the OS finds it via the executable directory at runtime (PE has no -Wl,-rpath); static linking uses zwasm.lib and skips the POSIX-only -lc / -lm. Plan C-c. (#71)
  • -Dstrip=true build option in build.zig strips the CLI binary at link time via LLD. Used by the Binary size check and the size-matrix CI jobs so they no longer depend on a host strip tool — portable across ELF / Mach-O / PE. The Binary size check now runs on windows-latest (with a 1.80 MB ceiling reflecting PE overhead; Mac 1.30 MB, Linux 1.60 MB unchanged) and size-matrix is a 3-OS matrix (Ubuntu / macOS / Windows). Plan C-e + C-f. (#70, D137)
  • test/c_api/test_ffi.c ported to Windows: dynamic loading via LoadLibraryA + GetProcAddress, threading via CreateThread + WaitForSingleObject, pipes via _pipe (binary mode). Sources are selected by #ifdef _WIN32 blocks; POSIX path unchanged. The runner switches from system gcc to zig cc so it does not require a host C compiler. The if: runner.os != 'Windows' CI guard on Run FFI tests is dropped accordingly. gate-commit.sh no longer auto-skips ffi on Windows. Plan C-b. (#72)
  • scripts/windows/install-tools.ps1 now provisions Rust (via rustup-init.exe with wasm32-wasip1 target), Go, and TinyGo in addition to the existing core toolchain — i.e. the same set Linux/macOS get from flake.nix. Rust is installed into a self-contained rust-<toolchain>/ with CARGO_HOME / RUSTUP_HOME set in user-scope env so the install does not pollute %USERPROFILE%\.cargo. -OnlyTool accepts rust, go, tinygo. Closes the local-realworld 25/50 → 50/50 gap on Windows. (#74, W52)

Changed

  • WASI SDK version bumped 25 → 30 to align CI with flake.nix (which was already at 30). Realworld 50/50 PASS verified locally on macOS aarch64 with the new SDK. (#60)
  • CLAUDE.md Commit Gate / Merge Gate sections point at bash scripts/gate-commit.sh / gate-merge.sh as the one-liner entry points; example commands switched to the .py runners that CI exercises. (#61)
  • .github/workflows/ci.yml benchmark job sources HYPERFINE_VERSION from versions.lock instead of hardcoding the version twice. (#65)

Internal

  • D136 (in decisions.md): Nix-as-SSoT design recorded with the Plan B / Plan C scope (unified gate scripts, Nix-based CI, Windows native installer, removal of the remaining Windows-skipped CI steps).
  • D137 (in decisions.md): cross-platform stripping via Zig -Dstrip=true (LLD --strip-all for ELF/Mach-O/PE) plus per-OS size ceilings (Mac 1.30 / Linux 1.60 / Windows 1.80 MB). Records the move away from the host strip tool (which is objcopy --strip-all ELF-only) and from a single global ceiling sized for the largest OS. (#73)
  • Doc alignment sweep: ARCHITECTURE.md, CONTRIBUTING.md, .dev/roadmap.md, and the release skill now agree with CLAUDE.md on Merge-Gate items, runner forms (python3 …py), and per-OS size ceilings. (#75)
  • Removed obsolete .dev/checklist-jit-fuel-timeout.md planning doc (all items shipped: PR #6 timeout merge, fuel-bypass fix, --timeout CLI option). (#76)
  • README front page updated: W52 status (Windows realworld 50/50 via local installer; CI parity is W50) + per-OS size ceilings (Mac 1.30 / Linux 1.60 / Windows 1.80 MB; D137). (#78)

[1.11.0] - 2026-04-26

W46 + W48: re-disable link_libc on the core build and trim the release binary back under the original 1.60 MB ceiling. No public API changes; embedders upgrading from 1.10.0 should see a smaller binary and identical behaviour.

Changed

  • Binary size ceiling: pulled back from the Zig 0.16 transition value 1.80 MB to 1.60 MB. Stripped Mac aarch64 binary is ~1.18 MB; Linux ELF is ~1.56 MB stripped. The original 1.50 MB target remains tracked as W48 Phase 2 (non-blocking).

Fixed

  • WasmModule.loadCore now frees export_fns and cached_fns (and the inner param_types / result_types slices) on the failure path when the Wasm start function traps. The earlier errdefer chain unwound vm/instance/wasi_ctx/module/store/self but skipped the export caches, leaking the slices for any module with a non-empty export and a trapping start function. The fix mirrors deinit so the failure path is symmetric with normal teardown. (Issue #42)
  • OOM test in loadLinked now heap-allocates the FailingAllocator so its address remains stable while the partially-loaded module is held; the prior stack-allocated form was use-after-free under the test's search loop. (PR #56 by @jtakakura)

Internal

  • W46 Phase 1+2 (link_libc=false restoration): build.zig core modules flipped to .link_libc = false while C API targets (shared-lib, static-lib, c-test) keep link_libc=true. WASI path-based ops, cErrnoToWasi, trace.zig stderr writes, and platform.appCacheDir all routed through platform.pfd* helpers so the core no longer pulls in libc on either OS. Linux direct syscalls used where stdlib bindings are missing in 0.16 (std.os.linux.statx, etc.).
  • W48 Phase 1: release binary trimmed via panic handler tightening, segfault handler disabled in ReleaseSafe, and main return type narrowed to u8. Combined effect restores the ceiling back below the original 1.60 MB target.
  • Test-site std.c.{pipe,dup,dup2,read,nanosleep} calls gated for Linux where the bindings are empty (W46 Phase 1c).
  • Credit @notxorand for the initial Zig 0.16.0 migration draft (PR #41, superseded by the v1.10.0 work).

[1.10.0] - 2026-04-24

Toolchain bump from Zig 0.15.2 → Zig 0.16.0 ("I/O as an Interface"). No public API removals or signature changes; downstream source stays compatible but consumers must upgrade to Zig 0.16.0 to build.

Changed

  • Zig toolchain: 0.15.2 → 0.16.0. Flake pins and all CI workflows updated.
  • WasmModule.Config gained io: ?std.Io = null and Vm gained io: std.Io — when Config.io is null, loadCore/loadLinked stand up a private std.Io.Threaded owned by the module (see D135). Existing embedders pass nothing and get the default behaviour.
  • LEB128 decoding pinned to the pre-0.16 stdlib algorithm. 0.16's std.Io.Reader.takeLeb128 does not enforce WASM's "integer too large" overshoot rule; the zwasm decoder was rewritten inline (40 lines) with the 0.15 @shlWithOverflow-based algorithm so spec test binary-leb128.77/78 continue to reject malformed 10-byte i64 encodings.

Fixed

  • Cross-platform fstat: on Linux, std.c.fstat / std.c.fstatat / std.c.Stat are all unavailable (empty bindings) and std.posix.Stat is void. path_filestat_get now dispatches to std.os.linux.statx on Linux (decoded to a neutral FileStat) and std.c.fstatat on Darwin. Test helpers that only needed the file size moved to lseek(SEEK_END).
  • build.zig modules explicitly set link_libc = true. Mac's Zig toolchain was lenient about extern "c" without explicit libc linkage; Ubuntu 0.16 rejects it hard.

Internal

  • Vm struct: io: std.Io = undefined field (set by loader).
  • WasmModule.owned_io: holds the auto-constructed std.Io.Threaded when the embedder did not supply one.
  • main(init: std.process.Init) on entry points (CLI, e2e_runner) so they can grab init.io / init.gpa / args from the runtime's start.zig.
  • WASI handlers: use std.c.* with file.handle for the POSIX ops that std.posix dropped (fsync/mkdirat/unlinkat/renameat/ftruncate/futimens/ pread/pwrite/dup/readlinkat/symlinkat/linkat/close/pipe/getenv). Errno → Errno via a single local cErrnoToWasi() helper.
  • @Vector runtime indexing was rejected by 0.16's compiler; SIMD extract/replace_lane and lane-memory ops rewritten to use [N]T arrays with @bitCast at push time.
  • Closes PR #41 (notxorand's migration draft) as superseded.

[1.9.1] - 2026-04-24

Changed

  • invoke/invokeInterpreterOnly now return error.ModuleNotFullyLoaded if the underlying VM is uninitialized (e.g., after OOM in loadLinked). This is a new error variant in the public API. Embedders matching on specific errors should handle this case. See API docs for details. (PR #40 by @jtakakura, closes #39)

Fixed

  • WasmModule.cancel() no longer segfaults on a partially-loaded module (vm == null after OOM in loadLinked). Matches the C API contract that documents cancel as a no-op on idle modules.

[1.9.0] - 2026-04-24

Added

  • Asynchronous execution cancellation (PR #28 by @jtakakura, closes #27). A host thread can now abort a running invocation:
    • Zig API: WasmModule.cancel() / Vm.cancel() — thread-safe, returns error.Canceled from invoke() at the next ~1024-instruction checkpoint (or JIT fuel interval).
    • C API: void zwasm_module_cancel(zwasm_module_t *) and void zwasm_config_set_cancellable(zwasm_config_t *, bool).
    • CLI: reports execution canceled when the runtime returns error.Canceled.
  • WasmModule.Config.cancellable: ?bool — opt-out of periodic cancellation checks for peak JIT throughput when the host never cancels.

Changed

  • By default, JIT-compiled loops now fire the fuel-check helper every DEADLINE_JIT_INTERVAL iterations even when no fuel/deadline is set, so cancellation takes effect without host instrumentation. Pass cancellable = false to restore the pre-v1.9.0 unconditional jit_fuel = maxInt(i64) behaviour.

[1.8.0] - 2026-04-21

Added

  • WasmModule.Config — unified configuration struct for module loading (PR #30 by @jtakakura)
  • WasmModule.loadWithOptions(allocator, wasm_bytes, config) — new primary entry point that consolidates all load variants
  • C API configuration setters: zwasm_config_set_fuel, zwasm_config_set_timeout, zwasm_config_set_max_memory, zwasm_config_set_force_interpreter

Changed

  • Resource limits (fuel, timeout, max_memory_bytes, force_interpreter) now apply during the start function. Previously the CLI applied them post-load, leaving the start function unconstrained.
  • Fuel budget is now decremented across successive invoke() calls, matching the pre-existing /// Persistent fuel budget doc comment. Previously each invoke reset fuel to the originally-configured value.
  • CLI --link fallback retry is now scoped to error.ImportNotFound. Previously any error on the imports-enabled load attempt triggered a retry without imports.
  • loadWithImports / loadWasiWithImports accept ?[]const ImportEntry (source-compatible with existing []const ImportEntry callers via Zig optional coercion).

Fixed

  • loadLinked: store.deinit() is now called on decode/instantiate failure paths (resource leak fix).
  • loadCore: errdefer allocator.destroy(self.vm) protects against VM leak on post-allocation failures.

Notes

  • All existing load* helpers (load, loadWithFuel, loadFromWat, loadWasi, loadWasiWithOptions, loadWithImports, loadWasiWithImports) are retained as thin wrappers over loadWithOptions, so embedders do not need to update call sites.

[1.7.2] - 2026-04-21

Fixed

  • ARM64 JIT: i32.rem_s/u and i64.rem_s/u produced wrong remainders when the destination register aliased the divisor (rd == rs2). UDIV/SDIV clobbered the divisor before MSUB could read it, so MSUB computed dividend - quotient * quotient instead of dividend - quotient * divisor. Triggered by TinyGo-compiled gcd (IR: r3 = r0 % r3), causing infinite loops after JIT compilation (HOT_THRESHOLD). The prior fix in v1.7.1 only covered rd == rs1. This commit preserves whichever operand rd aliases before the divide.

[1.7.1] - 2026-04-21

Added

  • -Dpic and -Dcompiler-rt build options for static library consumers (PR #24)

Fixed

  • Preserve caller-set vm.* settings (fuel, deadline, max_memory_bytes, force_interpreter) across invoke() and invokeInterpreterOnly() — previously reset to defaults between invocations (PR #31, #32)
  • setup-orbstack.md: broken Zig download URL (PR #35)

Changed

  • Spec testsuite bumped to f9c743a (from 072bd0d)

[1.7.0] - 2026-04-03

Added

  • SIMD JIT: ARM64 NEON (253/256 native) and x86_64 SSE (244/256 native) (Phase 13)
  • SIMD JIT optimizations: v128 base address cache (W43), Q-reg/XMM register class (W44), loop persistence (W45), guard page bounds check elimination (W46), FMLA/FMLS instruction fusion (W47)
  • C API: FD-based WASI stdio and preopen configuration (zwasm_wasi_config_set_stdin_fd etc.) (D133)
  • C API: zwasm_module_invoke args marked const (PR #16)
  • JIT fuel check at back-edges: enables timeout support for JIT-compiled code (Phase 19.2)
  • Epoch-based JIT timeout (D131): cooperative timeout via shared fuel counter
  • Multi-value return support in RegIR (OP_RETURN_MULTI)
  • --interp CLI flag for interpreter-only execution (debugging/differential testing)
  • Wide-arithmetic validation support (i64.add128, i64.sub128, i64.mul_wide_s/u)
  • 5 real-world SIMD benchmarks (C -msimd128): grayscale, box_blur, sum_reduce, byte_freq, nbody
  • 3 additional SIMD benchmarks: mandelbrot, matmul (C, wasi-sdk), blake2b_simd (Rust), simd_chain
  • SIMD benchmark comparison infrastructure (bench/run_simd_bench.sh, bench/simd_comparison.yaml)
  • Rust FFI example using zwasm C API
  • FFI tests in CI and commit/merge gate checklists

Fixed

  • memory64: u64 offset decoding and i64 address handling across decoder, validator, predecoder, and VM — fixes wasm-tools 1.246.1 compatibility
  • JIT correctness sweep (Phase 20): 8 real-world compat programs fixed (45→50/50)
    • ARM64 fuel check clobbering vreg 20 (x0) at loop back-edges
    • written_vregs pre-scan: stale reload in loops after call sites
    • Void self-call result clobber (n_results vs result_count)
    • Void-call reloadVreg clobbering live local variables
    • ARM64 ABI register clobbering in emitMemFill/emitMemCopy/emitMemGrow
    • Remainder register aliasing when rd == rs1 (MSUB on ARM64)
    • Stale scratch cache in signed division overflow check
    • x86 i32 div/rem signed edge case: 32-bit CMP for -1 check
  • JIT memory64 bounds check and custom-page-sizes (CI failure since 3/24)
  • JIT memory_grow64 truncation and cross-module call instance
  • JIT extract_lane encoding and back-edge poisoning
  • ARM64 JIT simd_v128 sync in MOV/CONST for SIMD correctness
  • Q-cache stale read in extract_lane upper-half fallback
  • ARM64 JIT emitGlobalSet ABI clobber: vreg r21 overwritten before reading value (W35)
  • x86_64 wide-arithmetic E2E: 19 tests fixed (validator + Debug i128 build mode)
  • i32.store16 access size in interpreter
  • --interp flag incomplete for doCallDirectIR path
  • Windows C API: intptr_t for host fd, cross-platform File.close for stdio handles
  • Shared library segfault on Linux x86_64 (PR #11)

Changed

  • SIMD operations now JIT-compiled (was: stack interpreter only, ~22x slower)
  • SIMD microbenchmarks: image_blend 4.7x, matrix_mul 1.6x (beats wasmtime)
  • HOT_THRESHOLD lowered from 10 to 3 for earlier JIT compilation
  • Contiguous v128 storage layout for SIMD JIT (W37)
  • Real-world compat: Mac 50/50, Ubuntu 50/50 (was 42/50)
  • E2E tests: 795/795 on Mac (was 773/792 on Ubuntu)
  • Spec tests: 62,263/62,263 (100%, 0 skip)
  • Binary size: 1.29 MB stripped. Memory: ~3.5 MB RSS.
  • wasm-tools bumped to 1.246.1
  • E2E runner built with ReleaseSafe (fixes x86_64 Debug i128 issues)

[1.3.0] - 2026-03-02

Added

  • Module cache for predecoded IR (Phase 1.2): reuse compiled module state across invocations
  • Cache-enabled benchmark variants for all bench scripts (--cache flag)

Changed

  • Benchmark suite expanded with cached variants for real-world programs
  • CI: pinned wasmtime version to avoid broken install script on macOS
  • Release process: migrated Ubuntu testing from SSH to OrbStack
  • Project docs reorganized for post-v1.2.0 phase planning

[1.2.0] - 2026-02-27

Added

  • Real-world compatibility tests: 30 programs (C, C++, Go, Rust) verified against wasmtime
  • On-Stack Replacement (OSR) for back-edge JIT: enter at loop body, bypass prologue side effects
  • Spec test strict mode: --strict flag exits non-zero on skips, enforced in CI
  • 11 new unit tests (521 total)
  • E2E tests expanded: 792/792 assertions (from 356)

Fixed

  • Spec validation: 87 previously skipped tests now enforced (GC typed references, subtyping, table types, exception handling catch clauses, rec group boundaries)
  • Spec infra: 18 assert_exception tests now handled by spec runner
  • E2E memory leak: errdefer in void-returning function was a no-op in Zig 0.15
  • JIT self-call stack overflow use-after-free (E2E segfault on aarch64)
  • Back-edge JIT restart corrupting Go WASI state machines (side-effect detection)
  • ARM64 OSR prologue: push FP callee-saved d8-d15 (stack corruption fix)
  • x86_64 OSR prologue: load physically-mapped vregs (stale register fix)
  • x86_64 select aliasing: val2 clobbered when rd == val2_idx
  • JIT IR instruction limit (MAX_JIT_IR_INSTRS=1500) prevents miscompilation of large functions
  • Go state machine detection: br_table boundary inclusive of target instruction

Changed

  • Spec tests: 62,263/62,263 passed (100.0%, 0 skip — up from 62,158 with 105 skips)
  • Benchmark results: 20/29 match or beat wasmtime, 27/29 within 1.5x (up from 14/23)
  • st_sieve: 0.97x wasmtime (restored from 30.82x regression via OSR)
  • GC benchmarks: gc_alloc 0.50x, gc_tree 0.73x wasmtime (JIT for struct ops)
  • nbody: 0.97x wasmtime (FP cache D2-D15 expansion)
  • Nightly CI: spec tests now run ReleaseSafe (eliminates Debug tail-call timeouts)

[1.1.0] - 2026-02-18

Added

  • WAT parser spec parity: GC type annotations, memory64 syntax, typed select, NaN payloads
  • WAT roundtrip 100%: 62,156/62,156 spec test modules parse and re-encode correctly
  • SIMD interpreter fast-path: predecoded IR dispatch for v128 ops (~2x improvement)
  • SIMD performance analysis and 3-phase optimization roadmap
  • Japanese documentation (book/ja/) with language switcher
  • GC benchmarks in runtime comparison recording

Changed

  • Binary size: 1.31 MB (ReleaseSafe, from 1.28 MB — WAT parser improvements)
  • Runtime memory: 3.44 MB RSS (fib benchmark)
  • Benchmark results: 14/23 match or beat wasmtime (up from 13/21, at v1.1.0 time)
  • Updated all documentation with fresh benchmark data

[1.0.0] - 2026-02-17

First stable release. API frozen under SemVer guarantees.

Added

  • mdBook documentation site with 12 chapters
  • GitHub Pages deployment for book
  • CI benchmark regression detection (20% threshold)
  • CI binary size check (1.5 MB limit), ReleaseSafe build verification
  • E2E tests in CI (wasmtime misc_testsuite, 356 assertions)
  • Nightly sanitizer job (Debug build) and fuzz campaign (60 min)
  • CI caching for Zig build artifacts
  • Overnight fuzz infrastructure (test/fuzz/fuzz_overnight.sh)
  • API boundary documentation (docs/api-boundary.md)
  • CONTRIBUTING.md, CODE_OF_CONDUCT.md, issue templates
  • Release automation: tag-triggered cross-platform builds
  • Install script (install.sh), Homebrew formula template
  • Host function and WASI examples (5 examples total)
  • CHANGELOG.md
  • SemVer versioning policy, deprecation guarantees

Changed

  • BREAKING: loadWasi() defaults to cli_default capabilities (stdio, clock, random, proc_exit). Use loadWasiWithOptions(.{ .caps = .all }) for full access.
  • --env KEY=VALUE injected variables now accessible without --allow-env
  • Error messages now use human-readable format (30 error variants)
  • README: badges, install section, documentation links

Added (security)

  • --sandbox CLI flag: deny-all capabilities + fuel 1B + memory 256MB
  • Capabilities.sandbox preset for library API
  • Restrictive library API defaults for safe embedding

[0.3.0] - 2026-02-15

Added

  • GC proposal: 31 opcodes (struct, array, i31, cast operations)
  • Threads: 79 atomic operations (load/store/RMW/cmpxchg), shared memory, wait/notify
  • Exception handling: throw, throw_ref, try_table
  • Function references: call_ref, br_on_null/non_null, ref.as_non_null
  • Wide arithmetic: add128, sub128, mul_wide_s/u
  • Custom page sizes proposal
  • Multi-memory proposal
  • JIT optimizations: inline self-call, smart spill, direct call, depth guard caching
  • x86_64 JIT backend
  • Fuel-based execution limits
  • Max memory limits
  • Resource limit enforcement (section counts, locals, nesting depth)
  • Security audit (.dev/archive/audit-36.md, docs/security.md, SECURITY.md)
  • Fuzz testing infrastructure with 25K+ corpus

Changed

  • Spec test coverage: 62,158/62,158 (100%)
  • E2E tests: 356/356 from wasmtime misc_testsuite
  • Binary size: 1.28 MB (ReleaseSafe)

[0.2.0] - 2026-02-10

Added

  • Component Model: WIT parser, binary decoder, Canonical ABI, WASI P2 adapter
  • WAT text format parser (zwasm run file.wat)
  • WASI Preview 1: 46/46 syscalls (100%)
  • Capability-based WASI security model
  • Module linking (--link name=file)
  • Host function imports
  • Memory read/write API
  • Batch mode (--batch)
  • Inspect and validate commands
  • ARM64 JIT backend
  • Register IR with register allocation
  • SIMD: 236 v128 opcodes + 20 relaxed SIMD

[0.1.0] - 2026-02-08

Added

  • Initial release
  • WebAssembly MVP: 172 core opcodes
  • Stack-based interpreter
  • Basic CLI (zwasm run)
  • Zig library API (WasmModule.load, invoke)