AGENTS.md never documents the container image: "release.yml then does everything" omits ghcr.io

[plusky]

AGENTS.md's Releases section says:

.github/workflows/release.yml then does everything: hermetic builds for
x86_64-unknown-linux-gnu and aarch64-apple-darwin, an x86_64 .deb and
.rpm, a GitHub release with all of those and their .sha256 files, and
finally the crates.io publish

"does everything" followed by an exhaustive-sounding list, which omits the
container and container-manifest jobs — the multi-arch image published to
ghcr.io.

grep -i 'ghcr\|container\|image' AGENTS.md returns nothing. The contract
file that agents are told to follow does not mention the container artifact at
all, anywhere.

Why it matters

• The image is a shipped, publicly consumable artifact, and a pushed manifest
  cannot be cleanly unshipped — the exact irreversibility that motivated toolchain-drift is not a required status check, so a failing guard does not block a merge #138.
• An agent reading only AGENTS.md would not know the image exists, so it would
  not know to keep the Dockerfile's toolchain, the packaging paths, or the
  release surface consistent with it.
• The Dockerfile and its toolchain-drift guard are described in the repo,
  but nothing connects them to a published artifact.

Scope: a brief mention, nothing more

Narrowed 2026-08-30. The fix is a short mention in AGENTS.md's Releases
section so the contract stops implying the enumerated list is exhaustive —
name the container image and where it is published. That is all.

Explicitly NOT in scope: tag semantics, the multi-arch manifest mechanics, a
container section of its own, or duplicating what release.yml's own comments
already say. AGENTS.md is a contract, not a manual; the detail belongs in the
workflow where it already lives.

One dense line or two, house style. Worth keeping (it is the non-obvious part
and release.yml already encodes it): container is a sibling of publish, not
upstream of it, so a broken image build must not hold back the crates.io release
of a tag whose binaries already shipped.

Found during the adversarial review of #170.