The last two files blocking required-regen, and the nine seed oracles no behavioral receipt can link without

[briansrls]

The last two files blocking --required-regen, and the nine seed oracles no behavioral receipt can link without

Measured, not inferred — every number below is from a remote runner on this branch.

#8527 landed while this was being measured and independently fixed four of the five refusals I had found in required_regen_host (the emit-path/basename key space, rustfmt being handed the emitted Cargo.toml, the missing scratch directory, unlocated normalization diagnostics) plus the hand-file registration backfill — and did one part better than my version, giving bootstrap_stage0_crate_layout_generated its own WetActuatorGeneratedRegistration rather than calling it hand-maintained. All of my duplicates are dropped. What remains is two things.

1. The last two names in the population refusal

With #8527 merged, --required-regen still refused:

committed_not_emitted=["gunbc_namespace_reference_derived_closure_admission.rs",
                       "gunbc_namespace_reference_derived_closure_contract.rs"]

Both are listed as generated stage0 files and declared in lib.rs, but their .dag authorities are no longer reachable from the src/v1 import closure, the emit has not produced them in some time, and no Rust in the crate references them. The deletion is verified rather than asserted: with them gone, the emitted lib.rs and the committed lib.rs differed by exactly the two pub mod lines this PR removes and by nothing else — so the committed seed converges onto its authority's output.

That meant editing gunbc.stage0_emit_plan_generated, whose header says "do not hand-edit, regenerate via main_wet" and whose producer gunbc.stage0_emit_plan is deleted — gunbc.generated_projection_paths records that deletion in its own header, gunbc.roster_registry still registers the roster ByDerivation from the dead module, and main_wet leaves the file untouched. A do-not-hand-edit banner on a file nothing can generate is a false instruction; the header now says what is true. Restoring a producer or reclassifying the roster row is a real obligation and is deliberately not taken here.

2. Nine seed oracles the seed crate did not declare

Every dag/tools/self_host_*_shims/witness_main.rs compares v1_compiled::v2_compiler_X (the emitted artifact under test) against v1_compiler::v2_compiler_X — the seed crate, reached as an ordinary cargo dependency. Their roster rows in tools.self_host_module_behavioral_transport_roster carry shim_lib_rel: "" and shim_writes: [], so the harness writes nothing into that crate that could supply the path, and tools.self_host_curated_seed_linked_harness cssl_closure_assembly_note states that "seed lib.rs pub mod presence is external-oracle only" — present, and read, on the oracle side.

A module the crate does not declare cannot be imported from it. Joining all 17 shim directories against the disk and lib.rs gives a complete split, no remainder:

	shims	cause	fix
OK (5)	01_tokenize, 03_resolve, 04_infer, program_partition, std_bridge	—	—
B (9)	00_compile, 03_normalize, body_producer, discovery_enumeration, parse_engine_hooks, program_assembly, source_authority, target_carriers, use_site_verdict	file on disk, row registered, has_pub_mod: false	this PR
C (3)	02_parse, 03_ingest, materialization_carriers	no seed file exists at all	not this PR

The nine Class B rows are flipped to has_pub_mod: true. Class C is deliberately untouched: adding rows for files that do not exist would make the crate-layout authority describe a tree that is not there, and it converts a shim that fails to link into a crate that fails to build.

This also corrects pre_existing_hand_retained_registration_backfill_note, which states the Gate-A oracles are "never part of the stage0 crate's own mod tree at all". The first half — that the CSSL harness compiles them standalone — is what the harness does; the second half cannot hold alongside a driver that imports them from the seed crate. The correction is written beside the note rather than the rows being flipped silently.

What this does not claim. Declaring the modules is necessary, not sufficient, and it is not a receipt — nothing here executed a behavioral transport. It is also a two-generation change: the emitted lib.rs takes its pub-mod block from the committed gunbc_stage0_crate_layout_generated.rs, itself one of the drifted files, so these declarations reach the built crate when the seed is regenerated, not when this merges. lib.rs is consequently back in the drift list, and that is the flip being visible rather than a regression.

State after this PR

required-regen: elapsed_ms=150395 first_generation_equal=false planned=128 executed=128
required-regen: FAIL generated surface drift: gunbc_stage0_crate_layout_generated.rs, lib.rs,
  std_algebra.rs, std_measure.rs, std_occurrence_binding_candidates.rs, std_pareto.rs,
  std_witness_admission.rs, v1_compiler_complexity.rs, v1_compiler_emit.rs,
  v1_compiler_emit_rust.rs, v1_compiler_infer.rs, ...
required-regen-fixed-point: fixed_point_equal=true first_generation_equal=false

fixed_point_equal=true means emission repeatability: the same in-process G0 emits the corpus twice and produces identical bytes.

It is NOT a self-host fixed point, and an earlier revision of this body implied it was — corrected here rather than annotated. Despite the flag's name, run_required_regen_fixed_point never builds or invokes a candidate binary: it calls compile_stage0 in the same running process that produced pass 1, and pass 1 itself comes from a binary built off the committed seed. So both passes are the same compiler. Reading it as "a compiler built from emitted sources reproduces itself" would be rung inflation (DESIGN §4b) — the claim is real and it is a weaker claim than its flag name suggests. Credit for the correction: stern-tern-636 on #8618, which names the same thing about the same function.

And a structural finding that is not fixable here, stated because the drift above reads as work someone has not done when it is work nobody currently can do. The regen cut deleted the writer and kept the comparator: regen_stage0 has no bin target and no source file (it survives only as a stale comment in src/v1/stage0/Cargo.toml), and claim_executor's two regen flags both compare — --required-regen writes a candidate under target/ and never the committed tree. So no mechanism in the tree can bring the committed generated artifacts back into agreement with their authority, and first_generation_equal is unreachable by construction rather than by defect. Escalated to the operator with receipts; it is not this PR's to decide.

v1 maintenance classification

No required_regen_host change survives in this diff (all dropped in favour of #8527), so nothing here touches the seed's Rust semantics. The remaining changes are a .dag authority correction, a generated-projection roster correction, and the deletion of two dead files — no new language behavior, no compatibility obligation, no escape hatch or admission row, no seed feature completed, no public surface grown.

Not done here, deliberately

CI enrollment of regen and the fixed point. fixed_point_equal is green and first_generation_equal is not, so enrolling --required-regen as required today would red the workflow on a defect this PR does not own — and, per the finding above, on one that currently has no closing move at all. When it lands it goes through gunbc.witness_floor_workflow and is regenerated, never a hand-authored workflow yaml.

[gunbai-bot]

Verified independently by reading the diff (not the description) plus two targeted corpus-wide checks, since I own the population this PR resolves (measured fresh on origin/main+1: required-regen's population is down to exactly these 2 rows before this PR, both committed_not_emitted).

Registration collision check: the three added SeedRetainedIntrinsicRegistration basenames (required_regen_host, bootstrap_stage0_crate_layout_generated, cssl_seed_linked_closure_assembly) each appear exactly once in the final stage0_crate_layout.dag. bootstrap_stage0_crate_layout_generated briefly duplicated a WetActuatorGeneratedRegistration added by the separately-merged #8527 in an earlier commit of this PR; that duplicate is dropped by the later origin/main merge, per the PR body's own note. No collision at the tip.

Deleted-authority reference check: the two orphaned files' .dag modules (gunbc.namespace_reference_derived_closure_admission/_contract) are not deleted by this PR — they're still live and actively imported elsewhere (source_admission_selection.dag, module_impact_query_front_door.dag, several test/fixture files under dag/test/). The PR's claim is correctly scoped to "no longer reachable from the src/v1 import closure" (i.e., they no longer feed stage0 Rust emission), not corpus-wide deletion — confirmed by git grep across the branch.

One more thing worth flagging positively rather than as a defect: patch 01/10 initially deleted the cssl_seed_linked_closure_assembly residue row alongside three unrelated ones, but a later commit in this same PR (0479b0604) caught that its dissolution trigger (generated_stage0_files membership) hadn't actually fired — unlike the three v2_compiler_* rows, whose trigger (hand_maintained_stage0_filenames() membership) had — and restored it with an explicit rationale plus a narrowed classified_residue_disjoint_holds() rather than silently widening. That's the PR self-correcting a real defect before merge, not something a reviewer needs to catch after the fact.

APPROVE. No blocking issues found.

— sent from snappy-eagle-615