nexora.sh · session #2147 · uptime 412d 06:18 tty/0 · pwd ~/services load 0.34 · mem 41% · gpu idle

$ cat stack.lock

A small toolbox, pinned and reproducible. Nothing here is fashionable for its own sake, each line earns its slot because it makes the result more deterministic, more auditable, or more honest about where the time goes.

root@nexora ~/ $sha256sum stack.lock
4f1c9e…a07b stack.lock · pinned, reproducible, boring on purpose
[01]

stack.lock

pinned · 2026-06-08
rustc 1.84.0# preferred runtime
zig 0.13.0# where rust gets in the way
clang 19# kernel + embedded
linux 6.9# target kernel
nix 2.20# reproducible builds
tla+ 1.8# specs that survive
postgres 17# when a db is fine
perf+ebpf# the truth machine
[02]

why each line

man stack.lock · NOTES

SECTION 1 · languages

rustc · zig · clang

rustc is the default: the borrow checker buys us correctness we'd otherwise pay for in 3am pages. zig comes out when Rust's safety gets in the way of a tight allocator or a freestanding target, comptime and no hidden control flow make it the right tool for the hot, unsafe corners. clang 19 handles the kernel modules and embedded firmware where the toolchain isn't negotiable.

NOTES we read the generated assembly · no language is a religion here

SECTION 2 · platform

linux · nix · postgres

linux 6.9 is the target we test against, we pin the kernel because scheduler and io_uring behavior shifts between releases and we'd rather know. nix gives us builds that are byte-identical on your machine and ours, so "works here" stops being a category of bug. postgres 17 is what we reach for when a real database is genuinely fine, we don't write a custom store to prove a point.

NOTES reproducible or it didn't happen · the simplest store that works wins

SECTION 3 · proof

tla+ · perf · ebpf

tla+ 1.8 specifies the protocols and state machines that have to survive a decade, a model checker catches the race you'd never reproduce in a test. perf and ebpf are the truth machine: when we claim a speedup, it's because the flamegraph and the counters agree, not because the microbenchmark felt faster.

NOTES a spec you can check beats a diagram you can't · measure, don't guess
[03]

environment

uname -a
$ uname -a
Linux build-rig 6.9.0-nexora #1 SMP PREEMPT_DYNAMIC x86_64 GNU/Linux
$ nproc / mem
cores: 16 (AMD Zen4) · ram: 64G ECC · swap: off (we don't page hot paths)
$ toolchain
rustc 1.84.0 · zig 0.13.0 · clang 19.1.0 · nix 2.20.5
$ reproducibility
flake.lock pinned · SOURCE_DATE_EPOCH frozen · CI rebuilds bit-for-bit
$ [wrn] locale set to C.UTF-8, yes, on purpose
$ [ok] environment hash matches lockfile · ready to build