nx_artifact_id.nx

nx_artifact_id.nx -- name a built artifact by its CONTENT so build and deploy can agree on WHICH BYTES. THE RACE THIS CLOSES (measured 2026-07-30, and it shipped foreign bytes to production): nx_sov_build_run writes _build/<t>.sov.elf. A separate step later copies that file to a staged name and promotes it. ANY sibling session that rebuilds the same target in between OVERWRITES _build/<t>.sov.elf. I built nx_mgmt_api at 578865, verified it, and by deploy time _build held a sibling's 579263 -- WHICH IS WHAT WENT LIVE. I verified one artifact and shipped another, and nothing anywhere detected it. It is a TOCTOU on the build output. The `build-<target>` lease does not cover it: that lease is released when the COMPILE ends, so the whole build->deploy handoff is unprotected by design. THE FIX IS THE CONTAINER-ECOSYSTEM ONE: DEPLOY A DIGEST, NOT A TAG. A path is a mutable tag; a (size, content-hash) pair is an immutable name. Bind them and the race becomes DETECTABLE instead of silent -- `verify` REFUSES rather than shipping whatever happens to be sitting at the path. COMPOSES THE CANONICAL HASH. nx_fnv.nx's own header carries the cardinal law: "All other primitives needing FNV-1a MUST import nx_fnv.nx ... never re-implement a hash inline." A tree-wide grep today found ~25 hand-rolled FNV variants ignoring exactly that, so this organ composes rather than adding a 26th. FNV-1a is a FINGERPRINT, not a cryptographic seal -- it detects an ACCIDENTAL swap by a sibling build, which is the measured failure. It is NOT a defence against a deliberately crafted collision; that rung is sha256 via the signing path, and this file must not be described as more than it is. VERBS nx_artifact_id id <path> -> `ARTIFACT size=<n> fnv=<16hex> path=<p>`, exit 0 nx_artifact_id verify <path> <size> <fnvhex> -> exit 0 MATCH / exit 3 MISMATCH (prints BOTH) / exit 4 unreadable license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0