fix(extract): account for Offsets struct in Linux dataStart

[xav-ie]

Problem

Running extract.mjs on a Bun-compiled Linux binary currently fails with an ERR_INVALID_ARG_VALUE when writing the first extracted module — the parsed "name" is actually a slice of JavaScript source, not a path.

Reproduces on any Linux ELF binary compiled with recent Bun. I hit it on @anthropic-ai/claude-code-linux-x64@2.1.114 (Bun 1.3.13).

Root cause

StringPointer offsets inside each CompiledModuleGraphFile are relative to the raw_bytes region Bun passes to StandaloneModuleGraph.fromBytes. That region contains [data .. module table .. Offsets struct .. trailer], so its total size is byte_count + sizeof(Offsets) + trailer.length.

The byte_count field in the Offsets struct excludes the Offsets struct and the trailer — per the comment in bun/src/StandaloneModuleGraph.zig:

"the length of the module graph with padding, excluding the trailer and offsets"

The Linux/ELF fallback was computing:

dataStart = trailerOffset + trailerBuf.length - byteCount;

which lands 48 bytes (sizeof(Offsets) + trailer.length) past the real start of raw_bytes. Every subsequent lookup reads from the wrong offset, and module names come back as random minified JS.

The Mach-O path derives dataStart from the section header's own length (not byte_count), so it's unaffected.

Fix

dataStart = trailerOffset - byteCount - OFFSETS_SIZE;

Equivalent to trailerEnd - (byteCount + Offsets + trailer), which is the actual start of raw_bytes.

Testing

Tested locally against @anthropic-ai/claude-code-linux-x64@2.1.114 (Bun 1.3.13, 225 MB binary):

• 5 modules extracted with correct names: src/entrypoints/cli.js (12.4 MB), image-processor.js, audio-capture.js, and two .node ELF addons
• cli.js passes node --check
• cli.js starts with expected // @bun @bytecode @bun-cjs wrapper and ends cleanly (}Qt1();}) — no truncation or overrun)
• Extracted .node files parse as ELF 64-bit shared objects
• Downstream resplit.mjs processes the extracted bundle end-to-end: 4,987 module files, 18,954 dependency edges
• Mach-O path unchanged (only the Linux/ELF line, the comment, and a constant extraction differ in the diff)

Known related issue (not addressed here)

MODULE_SIZE = 52 is hardcoded. In Bun 1.3.5 the struct is 36 bytes (no module_info or bytecode_origin_path fields yet); Bun added those somewhere between 1.3.5 and 1.3.13. On older binaries the current code silently reports "Modules: 0". Happy to submit a follow-up if desired.