required_regen_host: fixture supplies its partition rows; decode via wire authority; budget precondition stated as a red

[gunbai-bot]

What was actually failing (attribution, §6b)

cargo test --release -p v1-compiler --lib required_regen_host at main 46b0b2cca1, at #12255's parent and at #12171's parent:

• The 12 named tests, plus 2 in regen_round_cost_tests, fail at all three revisions, so Pkg7d 1b: emitter use-line edges returned typed, admitted against the stage0 crate partition #12171/Required-regen admits emitted edges against the emitted tree's own partition (#12171 bootstrap deadlock) #12255 did not cause them. With no readable cgroup memory bound they refuse HostBudgetUnreadable inside resolution (cli_run::entry_resolve). On srv1 inside systemd-run --user --scope -p MemoryMax=40G (run by neat-boar-16) they pass. That refusal is correct and was never a pass. It is not a code defect.
• Four other tests really are broken, by Pkg7d 1b: emitter use-line edges returned typed, admitted against the stage0 crate partition #12171, and are still broken at main: tests::a_scoped_emission_renders_the_selection_and_nothing_else, an_empty_selection_renders_no_module_and_declares_them_all, the_crate_module_list_and_the_manifest_are_content_independent, the_emitted_roster_is_whole_under_a_scope. They fail with Stage0EmittedEdgesNotCovered for fx_alpha/beta/gamma.

Earliest unjustified boundary: compile_stage0 read its admission inputs (host-shell roster and partition package graph) from the live seed tree inside the function. A witness over a three-module fixture had no way to supply the package graph its population is built under, so it was judged against the seed's rows, which name none of its modules. The admission was right. Its inputs were not at the interface.

Changes

1. One route, parameterized by the supplied package graph. compile_stage0_admitted(sources, selection, host_shell_modules, packages) does the emission and edge admission. Production compile_stage0 is a thin wrapper that calls it with the seed tree's host shell and emitted_tree_packages, so there is no test-only copy of the pipeline. The fixture supplies selection_fixture_packages, which has the seed's shape at fixture scale: v1_rt (rust_runtime_prelude_module) in a foundation crate, and the three modules in a layered core crate that re-exports it. So each prelude edge is admitted by the partition's own layering rule. The admission is unchanged. New RED control: the_fixture_refuses_as_not_covered_without_its_partition_rows (same fixture with no rows → Stage0EmittedEdgesNotCovered naming fx_beta).
2. Follow-up (a): the hand-restated decoder is removed, not recorded. A decode authority exists (value_to_wire_json → serde_json::from_value, as in namespace_baseline decode_environment_value), so emitted_tree_packages decodes the rows, shell name and shell deps into the mirror types through it. GeneratedPartitionCrateRow and the kind variants are no longer restated, and an unrepresentable row refuses as undecodable. Inhabitance claim for the real read: the_emitted_tree_package_graph_decodes_to_the_compiled_rows_at_a_settled_head.
3. Budget boundary (operator pick). require_measurable_host_budget() opens each of the 14 live-corpus claims and the decode claim. With no readable bound it panics NOT MEASURED: host budget unreadable (...): still red, never a pass, now located at the claim. The scope-repair control's formatter arm, Err(_) => return, was a vacuous pass (its comment claimed the opposite); it now panics NOT MEASURED too. New row: gunbc.recurring_failure_mode missing_precondition_reported_as_a_pass.
4. Follow-up (b): justified, no code change. gunbc.stage0_partition_rebuild_scope and gunbc.v1.v1_consumer_census have no Rust mirror and no Rust caller. They run only as interpreted .dag over the evaluated source roots, where stage0_partition_package_dependency_names reads that tree's generated_partition_crate_rows, which are the supplied rows by construction. The compiled-rows defect Required-regen admits emitted edges against the emitted tree's own partition (#12171 bootstrap deadlock) #12255 fixed exists only for a mirror running inside a predecessor binary. If either is ever mirrored and run by a predecessor, it must switch to partition_package_dependency_names_over with supplied rows.

Evidence

• cargo clippy --all-targets -- -D warnings: clean (BuildBuddy).
• required_regen_host on BuildBuddy (no cgroup): 35 pass; 15 red, all NOT MEASURED: host budget unreadable, as designed. All 25 tests:: pass, including the 4 fixed and the RED control.
• srv1 with a cgroup bound: pending (requested from neat-boar-16; to be posted here).

🤖 Generated with Claude Code

[gunbai-bot]

On the one item review 71608 couldn't verify: the derives hold. Stage0EmittedTreePackages (gunbc_stage0_emitted_edge_admission.rs) derives Debug, Clone, PartialEq, Serialize, Deserialize, and GeneratedPartitionCrateRow and GeneratedPartitionCrateKind carry the same emitted derives (the kind with #[serde(tag = "_variant")], the tag policy value_to_wire_json encodes). This is also established by execution: cargo clippy --all-targets -D warnings compiles the serde_json::from_value::<im::Vector<Rc<GeneratedPartitionCrateRow>>> decode, the ..(*selection_fixture_packages()).clone() spread and the assert_eq! on rows, and the tests:: group ran green on it. No change needed. The srv1 run under a cgroup bound is still pending, and it is the first execution of the decode claim itself.

— sent from clever-cat-758

[gunbai-bot]

0f555a2: convergence fixture owns a fresh root. Six regen_convergence_host_instrument_tests failed on srv1 at fc159acd3f, at fs::write(...).unwrap() in fixture_workspace, not in the git call. Cause (confirmed on srv1 by neat-boar-16): the root was keyed only on <pid>-<counter> under a shared /tmp. srv1's /tmp held 3848 such dirs, 3826 owned by ghrunner (the CI runner sharing srv1), so a pid collision wrote into another user's leftover tree and failed on permission. The root now also carries wall-clock nanos and is created with create_dir, so an existing root refuses, located, and is never reused. Clippy --all-targets -D warnings is clean at this head.

Known residue, not fixed here: these fixtures still leak their root when a test fails (remove_dir_all runs only on success), which is how the shared runner accumulated thousands. The fix is a Drop guard owning the root. That changes fixture_workspace's return shape at every call site, so I'm leaving it out of this PR to keep the srv1 qualification of this head valid. It's a small follow-up.

— sent from clever-cat-758

[gunbai-bot]

srv1 qualification at 0f555a28ea (run by neat-boar-16 inside systemd-run --user --scope -p MemoryMax=40G): cargo test --release -p v1-compiler --lib required_regen_host gives 50 passed, 0 failed, 1 ignored. That covers the 4 fixture tests, the NotCovered RED control, the 14 live-corpus claims under a readable budget, the 6 convergence fixtures, and the decode inhabitance claim (the first execution of the wire-decode path). The ignored test is the pre-existing live-corpus dag_artifact_identity_control_holds. Clippy --all-targets -D warnings is clean at this head.

— sent from clever-cat-758