D026 - The tooling is Rust, in one binary

decided · 2026-08-19

tool/ builds obscene-tool. The Python scripts are being replaced by subcommands, starting with the NID chain.

The stated reason for scripts had already expired. The argument for Python was that obSCEne should need nothing but clang - a real property for a conformance probe, since one that needs an unusual toolchain is one most people cannot build. But make check already required Python 3 and a POSIX shell, so the property was gone and the scripts were being defended on grounds that no longer held.

Given that, compiled is better on every axis that matters here:

• Most of this work is binary formats - offsets, sizes, packed structures, byte order. struct.pack format strings and untyped integers are liabilities in exactly that work, and the failure mode is a file that is accepted and subtly wrong.
• It ships as a binary, so a user needs no runtime at all - strictly better than requiring Python.
• cargo test instead of a shell script asserting exit codes.
• It matches the rest of the ecosystem, so it is maintainable by the person who owns it.

The deciding argument was sequencing, not cleanliness. The largest piece - mkmodule, which builds a vendor dynamic segment - is unwritten. The real question was never "should the existing scripts be ported" but "what language should the next five hundred lines be in", and porting afterwards is pure waste.

It does not depend on the emulator

orbistoun-nid implements the same algorithm, and this deliberately does not use it. A probe that depends on the emulator it measures is measuring that emulator's opinion of itself, and the independence is worth more than the duplication costs.

The duplication is made safe by a shared fixture, not by avoiding it. Today's byte-order confusion happened because there were two implementations and no test vector - one documented big-endian, the other implemented little-endian, and nothing could tell them apart. Both now check sceKernelLoadStartModule -> wzvqT4UqKX8, so they cannot drift without a test going red. Two implementations with one fixture is a sound arrangement; two with none is how this went wrong.

Lints

Borrowed from the emulator side: clippy::all denied, pedantic on, plus arithmetic_side_effects and indexing_slicing - because silent wrapping and out-of-range indexing are the two ways this particular tool would produce a corrupt module rather than an error.