feat: the agent's machine carries its own Rust, built rather than bor…

[CesarManzoCode]

…rowed

On 2026-08-30 a paid benchmark run watched Claude, inside the machine, choose exactly the right primitive — ask what a symbol is, then rename it — and be answered:

context: source: index, resolution: matched, analyzer_starts: 0
rename:  ok: false, error: unresolved,
         message: there is no `cargo` on this machine

Everything after that in the transcript is a consequence. The preflight had said READY, and it was right about what it asked: the machine was alive and holding the right tree. Neither is the capability a Rust task needs, and there was nothing between the money and finding that out.

The first fix attempted was to copy the host's ~/.rustup onto the store. Cesar stopped it and he was right: Filosofia-Fundacional.md says Thalyx is the whole system, so a programming face that only works because Fedora has rustup installed is a face that belongs to Fedora. Move the disk to another x86_64 box and the semantic provider would vanish.

So the runtime is Thalyx's own artifact. dev/build-rust-runtime.sh builds it from digest-checked upstream tarballs — Rust's official musl host tools, whose sha256 come from Rust's own channel manifest — plus the two files Rust does not publish: musl's loader, compiled here from musl's release tarball, and libgcc_s.so.1, linked out of the libunwind.a that ships inside rust-std itself. Nothing is copied from the machine that builds it. 644 MB, on the store, never in the initramfs — make count still says the kernel and one program, and there is now a test that arms the archive and counts it.

Why musl and not the ordinary toolchain was a measurement taken before anything was written: the GNU one needs glibc's loader, libc, libm, libdl, librt, libpthread, libgcc_s and libz from the host plus a separate 191 MB libLLVM; the musl one needs exactly two files, and LLVM is inside librustc_driver. Two missing files is a problem a person can close.

That the unwinder-only libgcc_s is enough is also a measurement: of the 883 undefined symbols across cargo, rustc, rust-analyzer, the proc-macro server and librustc_driver, everything resolves against musl and librustc_driver except 29, and the only ones of those that are not weak are the fifteen Unwind*, all of which libunwind.a defines. Then it was run rather than argued.

What is checked, and how, because ldd is the wrong instrument: it starts the real loader against the real /lib, so it says what this Fedora would resolve. An artifact missing a library that happens to live in the builder's /usr/lib looks perfect on the builder and is a directory of dead ELF files inside Thalyx, whose /lib is empty — and the failure surfaces as ENOENT from execve, which reads as "there is no cargo on this machine" about a cargo that is right there. So thalyx dev rust-runtime reads PT_INTERP and every DT_NEEDED out of the headers and asks whether they are inside the artifact.

Also here:

• discovery puts Thalyx's own runtime second, behind only a variable that names a file. When Thalyx carries a compiler, that is the compiler;
• PID 1 makes /lib/ld-musl-x86_64.so.1 point at the artifact's loader after mounting the store, which is what the kernel reads out of PT_INTERP;
• a read-only toolchain verb that runs cargo --version and rust-analyzer --version inside the machine and reads the workspace's manifests, and thalyx-mcp --preflight --needs-rust, which refuses to be READY without it. The benchmark asks for it whenever the project has a Cargo.toml — derived from the tree, because a flag somebody has to remember is the flag that was missing on the run that needed it;
• make -C image agent PROJECT=… stages it automatically for a Cargo workspace, prints where it came from and how much, and refuses — deleting what it copied — if the artifact is not closed.

Proven physically in this container, not just in tests: inside a chroot holding the artifact, /proc and three device nodes — no shell, no /usr, no /lib64 — cargo, rustc and rust-analyzer start, cargo metadata describes a synthetic workspace, and a full rust-analyzer session resolved a definition and returned four real rename edits. What is not proven here is a booted Thalyx doing it; that needs the hardware, and dev/verify-agent-rust.sh is the one command for it.

Two rules learned and written into Estrategia-de-Pruebas.md: a check that asks the host answers about the host, and musl resolves $ORIGIN for the main program by reading /proc/self/exe — the same binary starts with /proc mounted and dies without it, complaining about a library that is right where it says.

Claude-Session: https://claude.ai/code/session_01CvoEad9oWgX1Y594uirzbn