Drive the artifact axis with a real cargo build and a real digest

[gunbai-bot]

Stacked on #8172 (the pure folds). This is the half that makes the axis derived rather than derivable: before it, nothing in the tree digested a built artifact, so every generation carried ArtifactNotMaterialized and every axis-invalidation control was synthetic.

What lands

• cargo.Build.BuildManifestMessages — cargo build --message-format=json. Its own operation rather than a flag on BuildManifest, because the flag is what makes stdout a message stream instead of human text: it is constitutive of the output contract, not a caller preference.
• materialize_artifact_wet — the only function in the module that reaches the host. Three service calls, no decision the pure folds don't already make, so the hermetic corpus keeps every arm under per-PR execution and the wet cadence only has to prove the real tools reach them.
• A wet execution witness — writes a real minimal crate to a real mktemp dir, runs a real cargo build, decodes cargo's own stream, digests the file cargo named with real sha512sum. Excluded from hermetic discovery, enrolled in bin_witness_wet_entries at function grain.

The digest is taken only for a path cargo reported

The match producing the observation is what makes this structural. The non-reported arms have no executable to hand sha512sum, so they yield a refusal saying exactly that — discarded by the mapper, but true as written, so a later edit that did consult it stays fail-closed instead of naming a file nothing observed.

Evidence

4 wet claims green against a real cargo build (~0.5–0.8s each); the 10 pure folds stay green.

Mutation probe, and it found a real gap first. My initial three wet claims did not discriminate whether the digest was of the reported file — they checked the path the outcome carried, which the mutation didn't move. So I added the_digest_is_of_the_file_cargo_named_not_of_some_other_real_file, then planted the defect: digest manifest_path instead of the reported executable.

Claim	Under the mutation
a_real_cargo_build_materializes_through_the_real_digest	still PASS
the_digest_is_of_the_file_cargo_named_not_of_some_other_real_file	FAIL

That top row is the point. A wrong-but-real file produces a real digest of a real thing that is not the artifact, and a materialization claim cannot tell. Only the identity claim can.

Two receipts kept in-file rather than fixed silently

• shell.Mktemp.Dir projects path: FilePath straight from stdout, so it carries mktemp's trailing newline. Interpolating it puts a newline inside the path and every later write refuses — this witness first failed at the fixture write, nowhere near its subject. Trimmed at the call site with the modeling gap named; the shared operation is consumed raw elsewhere, so changing it is its own change, not a rider.
• \{ is the escape for a literal brace, so "\{root}/Cargo.toml" is the literal text {root}/Cargo.toml. It wrote a file by that name and cargo reported manifest path {root}/Cargo.toml does not exist. Both halves were individually honest.

Explicitly not claimed

Each claim builds its own crate in its own temp dir, so this does not witness one-build-one-artifact — that control is about reusing an artifact across invocations, and every claim here deliberately starts from an empty target dir. Also outstanding: direct invocation via WitnessBin.Run, the warm-marker control, the per-axis invalidation RED owed to #8153, and migrating the four convention-derived callsites.

Lane: SH-C.

🤖 Generated with Claude Code