fix(build): compile x64 binaries against Bun's baseline runtime

[elucid]

Fixes #454.

Problem

Every prebuilt x64 binary dies before running a line of Hunk code on CPUs without AVX2 — pre-2013 hardware and VMs with a conservative CPU model. bun build --compile with no --target embeds Bun's default x64 runtime, which is built for Haswell (AVX2/BMI2). Reported against 0.15.3; 0.19.0 is identical.

Reproduced on a real Nehalem KVM VM and on Windows 11 masked to SandyBridge:

	today	with this PR
Linux x64	Illegal instruction (core dumped), exit 132	0.19.0, full TUI with syntax highlighting
Windows x64	Bun panic, Features: no_avx2, exit 3	0.19.0

Fix

Compile x64 hosts against Bun's baseline runtime, which needs only x86-64-v2 (SSE4.2/POPCNT, Nehalem 2008+). arm64 keeps Bun's host default. CI and the release workflow now run the built linux-x64 artifact under an emulated pre-AVX CPU, so dropping the target fails the build instead of shipping a binary that cannot start.

Performance tradeoff

The baseline runtime is compiled without AVX2-dependent codegen, so this trades a small amount of optimization for running on every x86-64 CPU since 2008. Measured, that tradeoff is not material. Both Bun runtimes were run against the repo's benchmark suite with per-sample interleaving on Skylake Linux and Windows hosts:

	metrics	median delta	metrics with self-noise ≤3%
Linux (full suite, 15 samples/runtime)	58	−0.10%	within ±1%
Linux (interleaved re-run, 9 samples)	18	−0.29%	within ±1%
Windows (5 scripts, 7 paired samples)	41	+0.84%	median +0.07%

Every metric that looked like a regression turned out to be noise: the release gate's own thresholds are exceeded by same-runtime control comparisons on these hosts, and the one metric with a consistent early signal (scroll_tick_median_ms, baseline slower in 5 of 7 pairs) reversed at 17 pairs to 4.6% faster, slower in 9 of 17.

At the primitive level only Bun.stringWidth moved (+19.5% CJK, +9.1% ASCII). OpenTUI calls it, but instrumenting real renders shows ~20 calls totaling 0.06–0.08 ms per render — Hunk's own width path is plain JS and unchanged — so it costs roughly 0.015 ms per frame.

Distribution

No asset or package names change, so npm, mise/aqua (hunkdiff-{{.OS}}-{{.Arch}}.tar.gz), and the GitHub release tarballs are unaffected. Homebrew builds from source via bun run build:bin, so its x86_64 bottles pick this up automatically. nix/package.nix invokes bun build --compile directly and is unchanged — --target fetches a runtime, which the Nix sandbox cannot do.

Not included

• Launcher fallback in bin/hunk.cjs. Today it exits 1 with no output when the prebuilt binary dies. Worth fixing separately; note it cannot rescue Windows, where the failure is a normal exit code rather than a signal.
• Nix, which still requires AVX2 to build and to run.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
hunk-web	Ignored	Preview	Aug 18, 2026 3:09pm

[greptile-apps]

Greptile Summary

The PR compiles x64 binaries against Bun’s baseline runtimes and adds pre-AVX2 startup checks to CI and release builds.

• Selects OS- and libc-specific baseline targets for x64 hosts while retaining Bun’s default runtime on arm64.
• Adds unit coverage for target selection and QEMU-based Linux compatibility gates.
• Documents the new x86-64 CPU floor and records the patch-level release change.

Confidence Score: 4/5

The PR appears safe to merge after the non-blocking test-file indentation inconsistency is corrected.

The baseline target selection and Linux compatibility gates align with the supported native build matrix; the only accepted issue is formatting in the newly added unit test.

Files Needing Attention: scripts/build-bin.test.ts

Important Files Changed

Filename	Overview
scripts/build-bin.ts	Adds host-aware Bun baseline target selection and passes the selected target to compiled binary builds.
scripts/build-bin.test.ts	Covers x64 platform targets, Linux musl selection, arm64 defaults, and unsupported platforms; formatting needs adjustment.
.github/workflows/ci.yml	Adds a QEMU Nehalem smoke test for the locally built Linux x64 binary.
.github/workflows/release-prebuilt-npm.yml	Adds the same pre-AVX2 startup gate to the released Linux x64 artifact build.
README.md	Documents the SSE4.2 CPU requirement for x86-64 installations.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[build:bin] --> B{Host architecture}
    B -->|arm64 or unsupported| C[Bun host-default runtime]
    B -->|x64| D{Host platform}
    D -->|macOS| E[bun-darwin-x64-baseline]
    D -->|Windows| F[bun-windows-x64-baseline]
    D -->|Linux glibc| G[bun-linux-x64-baseline]
    D -->|Linux musl| H[bun-linux-x64-musl-baseline]
    E --> I[Compile dist binary]
    F --> I
    G --> I
    H --> I
    C --> I
    I --> J[Linux x64 QEMU Nehalem startup check]

Loading

Prompt To Fix All With AI

### Issue 1
scripts/build-bin.test.ts:5
**Test indentation breaks convention**

The newly added test bodies use two-space indentation instead of the repository's required four-space indentation, leaving the file inconsistent with the required formatting convention.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "fix(build): compile x64 binaries against..." | Re-trigger Greptile}