Self-host receipts link only emitted bytes: delete the stale hand shims behind self_host_03

[gunbai-bot]

What was wrong

self_host_03_normalize_behavioral_receipt_holds was red on main: 13 E0308 at rejected_with_pending(d1.clone(), ..). It was briefed as a seed-emitter value/Rc defect. It was not. At 043c22b the emitted 03_normalize closure (gunbc compile --entry src/v2/compiler/03_normalize.dag, 113 files) passes cargo check cleanly, and a planted type error reds as a control. The emitter writes Diagnostics (None | Some { NonEmptyDiagnostics }) as Option<Rc<NonEmptyDiagnostics>>, the same way at signature and call site.

The witness never built that crate. Its roster row overwrote v2_std_diagnostic.rs with the hand-authored dag/gunbc/instruments/self_host_std_bridge_shims/v2_std_diagnostic.rs, last touched in #10929, which still declared enum Diagnostics and rejected_with_pending(pending: Rc<Diagnostics>, ..). So the earliest unjustified boundary (§6b) was a second, hand-maintained copy of v2.std.diagnostic (a §3 fork) that went stale when the emitter changed. No emitter change and no emitter control were needed. Direction approved by neat-boar-16 (option A: delete-first).

Census: every row that used a shim, and every file in self_host_std_bridge_shims/

For each row: emit the closure, run cssl_assemble (whole-closure assembly), cargo build --release, then run the driver plain and with --inject-fault, with no shim.

row	emitted closure builds	driver without shims
03_normalize	yes	PASS after driver repair (below)
03_body_producer	yes	PASS after fixture repair (below)
parse_engine_hooks	yes	PASS, fault run FAIL, unchanged driver
use_site_verdict	yes	PASS, fault run FAIL, unchanged driver

Every shim passed, so every shim is deleted: the whole std-bridge directory (v2_std_{algebra,collection,diagnostic,integer,node}.rs), the 03_normalize shim set, and three per-row shim lib.rs files. No shim is kept.

The fields that carried them are deleted too (shim_lib_rel, shim_writes, run_closure_assembly, CuratedSeedLinkedShimWrite, the hand-shim fast path). A row can no longer name a shim. The --std-bridge-dir interface was never read (_std_bridge_dir), so it is removed end to end: cssl_assemble, cssl_seed_linked_closure_assembly.rs, the harness, emission_entry_instrument (SeedLinkedAssembly is now nullary), emit_subject_clean_frontier, and product_receipt_stage. The bridge's own dissolution trigger had already fired. So had that of the seed_closed_vocabulary_wildcard_census row about the v2_std_node shim, which is removed.

What the shims were hiding, now fixed

• body_producer fixture was malformed. The Transform body had zero edges. v2.std.node behavior_edges_conform requires count(children) >= 1 for Transform, and the emitted module rejects it correctly. The lax shim had kept it green. The fixture now carries one positional Value edge.

• 03_normalize driver was stale and its oracle was hollow.

  • The driver called normalize_fold_init, which normalize rebuilt the enclosing node once per child; rebuild it once #12014 deleted from the .dag.
  • It compared against the seed's hand-retained v1_compiler::v2_compiler_normalize, whose well_formed returns true.
  • Its fault input (an empty Disj) is accepted by both sides, so the fault arm was red however emitted normalize behaved.

  The driver now checks the .dag contract directly: a well-formed Value node is accepted, and a Value node with an edge is refused. The fault run feeds the malformed tree where the well-formed one is expected.

Deleting the stub src/v1/stage0/src/v2_compiler_normalize.rs

The stub's only consumer was this witness. I proved that by building, not by grep:

• the file, its lib.rs line, and its stage0_crate_layout / seed_retention_frontier rows are removed;
• v1-compiler builds locally (gunbc, claim_executor, cssl_assemble);
• ctrl-build --remote -- cargo clippy --workspace --all-targets -- -D warnings is clean;
• the required floor is left to this PR's CI.

Evidence (local private build, arm64 session)

• Fault-arm truth tables (third commit). Earlier heads had a fault arm that was a contradiction (p && !p): the injected run printed FAIL for every implementation. Found in neat-boar-16's review. Each injected run now asserts only the planted claim. Each cell is the witness's final marker, measured by mutating the emitted crate:

  03_normalize	plain	--inject-fault
  correct	PASS	FAIL (fault detected)
  emitted well_formed forced true (accept-all)	FAIL	PASS, so the harness refuses
  emitted well_formed forced false (reject-all)	FAIL	FAIL

  body_producer	plain	--inject-fault
  correct	PASS	FAIL (fault detected)
  emitted body_producer_dispatch_structured_body accepts everything	FAIL	PASS, so the harness refuses

• Stage0 regen fixed point: claim_executor --required-regen → first_generation_equal=true, 161/161. Generated layout artifacts were regenerated via generated_artifact_gate.dag main_wet, which also covers the derived .gitattributes line.

• Assembler tests: cargo test -p v1-compiler --bin cssl_assemble: 10 passed. The narrow-hand-lib RED control is deleted along with its subject.

• srv1, non-hermetic, at b97b07a (run by neat-boar-16): required-regen is at a FIXED POINT. 03_normalize (receipt and declared_source_refs), parse_engine_hooks and use_site_verdict PASS. body_producer FAILED there only on its stale claim-run path structured_body_dispatch.dag, a main defect. The second commit fixes it (the same one-line fix as MQ-1: list literals lower whole as their free-monoid construction, through the one value reader #12208), and the renamed claim PASSes locally under gunbc run --claim-run.

Known main reds this PR does not touch (pre-existing)

In the same srv1 run, 8 other self_host_*_behavioral rows FAIL. None of them ever carried a shim (their shim_writes were already empty), so their assembly path is unchanged apart from the unused --std-bridge-dir argument going away. The only seed module this PR removes is v2_compiler_normalize.

• 02_parse, 03_ingest, materialization_carriers. E0432 on v1_compiler::v2_compiler_{parse,ingest,materialization_carriers}. These seed modules are absent from src/v1/stage0/src/lib.rs on origin/main as well, so each driver names a seed oracle that no longer exists.
• 04_infer. The driver calls infer_arrow_domain_binding_heuristic_note, which the emitted module no longer carries: a stale driver.
• program_partition, discovery_enumeration, target_carriers. Emitted-crate rustc errors (E0308, E0223, im::Vector::Empty). These are emitter-side defects to be chain-derived.
• 03_resolve, program_assembly, source_authority. Red with no rustc error in the summary; cause not read.

Trigger for all of them: the rfm row's next rung, where the oracle is the interpreter over the same .dag rather than a hand seed module. Owner: to be dispatched by neat-boar-16.

Recurring failure mode

This PR adds gunbc.recurring_failure_mode hand_shim_shadows_an_emitted_module. For the roster, the class is now structurally impossible, because no field exists to name a shim. The residue is the seed's other hand-retained oracles (v2_compiler_body_producer, parse_engine_hooks, use_site_verdict, target_carriers, discovery_enumeration). Next-rung trigger: the behavioral oracle is the v1 interpreter's evaluation of the same .dag entry.

🤖 Generated with Claude Code