nix-bundle-lgx writes darwin-amd64. This package manager computes
darwin-x86_64. The alias table bridged only the four linux- rows, so
darwin and windows fell through with no alias at all:
Error: Package does not contain variant for platform:
darwin-x86_64-dev (package provides: darwin-amd64-dev)
Fail-closed, at install, for every package the standard bundler builds.
Reproduced against real binaries with matched controls — the same payload
spelled darwin-x86_64 installs, and the same shape on linux installs
because that row happens to be bridged.
The two producers disagree with each other while the SECOND CONSUMES THE
FIRST: nix-bundle-logos-module-install bundles a .lgx that nix-bundle-lgx
named, then calls lgpm with the arch spelled its own way. That is why the
disagreement is absorbed here, in the consumer, rather than by picking a
winner in one producer and breaking every .lgx already published with the
other spelling.
The canonical vocabulary is the one logos-module-builder's resolvePlatforms
just pinned — os in {linux, darwin, windows}, architecture in
{x86_64, aarch64} — so the first entry of each row is canonical and the
rest are legacy spellings a producer actually writes. Reusing that
vocabulary rather than inventing a third table is the point: a third table
is how this bug got here.
Only the ARCHITECTURE half is aliased. The OS half stays verbatim on
purpose — aliasing it would weaken the fail-closed check that stops a
Windows package installing as a macOS one.
Why no test caught it: every doctest hand-rolls the LGPM spelling and none
uses the bundler's, so the two vocabularies never met in CI.
Worth recording — the published catalog carries zero darwin-amd64 and zero
darwin-x86_64 across all 15 packages and 49 versions. No Intel-Mac package
has ever shipped, which is consistent with one being dead on arrival.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
