max / makenotwork

Declare what each host is, and check it before building or shipping there Bento and Sando knew a host's name, ssh target, architecture and capabilities and not what the box was, so everything downstream was rediscovered by probing or by a build failing: an ldd guard on the node before the symlink swap, a glibc_check in a recipe, a rebuilt astra regressing silently. ops-core gains `base_image`: a BaseImage (`id/version` from /etc/os-release, parsed not stringly, on Platform's reasoning), a probe, and a check. Bento's Host and Sando's Node gain `base_image` and `libc`. Bento checks in the release preflight barrier beside the commit check; Sando checks before a single byte is pushed, ahead of the rsync and well ahead of the ldd guard. Silence is a skip here, not a refusal, which is where this deliberately departs from Placement::check. macOS and Windows have no /etc/os-release, and base equality is not the compatibility requirement anyway: fw13 is pop/24.04 and prod is ubuntu/24.04, two bases sharing glibc 2.39, and demanding equal bases would refuse every MNW deploy over a problem that does not exist. A host that declares nothing is skipped and logged as skipped. Declarations are the 2026-08-25 measurements: fw13 and astra pop/24.04 glibc 2.39, prod ubuntu/24.04 glibc 2.39, testnot ubuntu/26.04 glibc 2.43. The last is the finding worth reading twice: staging is two Ubuntu releases and four glibc point versions ahead of production. Not done here: comparing a build host's libc against a deploy node's. That is the real compatibility question, it still happens at runtime in the ldd guard, and hoisting it to preflight is filed separately rather than smuggled in.