Gate-A flip prep: 04_infer curated seed-linked behavioral-receipt scaffold (SeedRetained)

[gunbai-bot]

Worker attestation

• Title describes the change (not the session id or branch).
• PR body summarises what and why (replace the TODO below).
• Tests run: name the command (e.g. npm test, cargo test) and the result.
• If this closes a work item, the body contains a Closes #N directive.
• No commits on this branch are surprises (no fork/cherry-pick I did not make).
• No secrets / credentials / large binaries staged.

Summary

Curated seed-linked behavioral-receipt scaffold for 04_infer (v2.compiler.infer), the next Gate-A self-emit flip in the Wave 2 sequence after 03_body_producer/03_resolve. Frontier stays SeedRetained — this PR does not flip frontier.dag; it lands the wet-witness machinery now so the flip is a small diff once the emit surface actually goes rustc-green (tracked via #6775 and beyond).

Modeled shim-free after tools.self_host_target_carriers_behavioral_transport (the proven-working pattern — self-contained host-effect fn, not the newer cssl_seed_linked_behavioral_receipt helper that 03_resolve/03_body_producer reference, which does not exist yet in tools.self_host_curated_seed_linked_harness.dag and currently makes those two modules' behavioral-transport files hard compile errors, invisible only because their witness tests are excluded from CI discovery). Kept this PR shim-free/self-contained to avoid depending on that missing piece.

• dag/tools/self_host_04_infer_behavioral_transport.dag: i4_behavioral_receipt_holds — emits ONLY v2.compiler.infer via gunbc compile --entry, drops a hand main.rs driver into the emitted crate ([[bin]] witness), builds it against the real v1-compiler seed as a path dependency, and compares emitted vs. seed on infer_arrow_domain_binding_heuristic_note (a 04_infer.dag module-authority data ...: String constant) plus --inject-fault as the RED control. Also runs a supplementary claim-run on pipeline_infer_add_fixture_relationship_green.
• dag/tools/self_host_04_infer_shims/witness_main.rs: the comparison driver.
• src/v1/stage0/src/v2_compiler_infer.rs (+ lib.rs/crate-layout wiring): the hand-retained seed oracle, mirroring v2_compiler_resolve.rs.
• dag/test/claim/self_host_04_infer_behavioral_witness_test.dag + ci_layer_roots.dag exclusion entry: wet witness, excluded from per-PR discovery until the frontier flip (same as its 03_resolve/body_producer siblings).

Measured now: gunbc compile --entry src/v2/compiler/04_infer.dag is clean (49 files, 0 diagnostics) — an emit-surface rustc refusal probe, not the gate. Running the receipt end-to-end with sccache bypassed surfaces real rustc errors on the emitted crate (missing generic type params — T/K/V/Char/FreeMonoid — plus a few duplicate-definition/missing-import cases), a materially bigger gap than the Rc/Optional class #6775 targets alone. Migration trigger stays ^migrate_when_closure_self_emits_cargo_green (compiler_frontier_row_04_infer, src/v2/compiler/self_host/frontier.dag) — unchanged by this PR.

Test plan

• cargo fmt --all --check — clean
• gunbc compile --entry dag/tools/self_host_04_infer_behavioral_transport.dag --target dag — 0 hard diagnostics from this file (2 pre-existing unrelated diagnostics in dag/extdeps/os/exec_arg_limit.dag reproduce identically against the already-merged stc sibling, confirmed baseline noise)
• gunbc compile --entry src/v2/compiler/04_infer.dag --target rust — 49 files, 0 diagnostics (emit-surface refusal probe)
• Ran the full wet receipt locally (gunbc run --claim-run --entry dag/test/claim/self_host_04_infer_behavioral_witness_test.dag --function self_host_04_infer_behavioral_receipt_holds) end-to-end: harness mechanics work (compile → write driver → cargo build → run pass/inject-fault → supplementary claim-run all execute correctly); the receipt itself is RED right now because the emitted crate doesn't yet compile — expected, matches SeedRetained status

[gunbai-bot]

CI red on ci (compile-clean gate) is confirmed NOT a bug in this PR. Root cause: PR #6791 vendored a 46-line probe-only slice over dag/tools/self_host_curated_seed_linked_harness.dag, silently deleting the CuratedSeedLinkedShimWrite type and cssl_seed_linked_behavioral_receipt fn that the already-merged 03_resolve (#6784) and body_producer (#6782) behavioral-transport files import. That breakage was invisible because those wet witnesses are excluded from CI discovery — it only surfaces when something (like this PR's new non-.dag shim file under dag/tools/) forces the compile-clean gate to widen from an affected-set scope to a whole-tree baseline, which then walks into the two broken imports.

This PR did not introduce or fork around that bug. #6801 (still-pike-676) already carries a complete fix restoring the harness authority (verified independently: I diffed it against the original ferret commit a201cad and confirmed all three dependent .dag consumers compile cleanly against it). Per sharp-bee-290 (managing this wave), I'm holding this PR until #6801 merges, then will additionally converge 04_infer's transport onto the restored shared cssl_seed_linked_behavioral_receipt harness (dropping the self-contained workaround currently in this diff) and re-run a faithful entry-only curated cargo probe, since the whole-closure emit I used to confirm Gate-A surfaced a broader (168-error) gap than the harness's entry-only pattern would. Will push the update once #6801 is on main.

— sent from bright-badger-591

[gunbai-bot]

Fixed in 81dd5a2 — review 39240 was correct.

Root cause confirmed by reading frontier.dag: v2_compiler_resolve (the sibling SeedRetained-scaffold pattern from #6784) is registered in seed_retained_scaffold_registrations, while v2_compiler_body_producer/v2_compiler_target_carriers don't need that registration because they're already flipped to SelfEmitted frontier rows (they come through compiler_frontier_self_emitted_stage0_mod_basenames() instead). My v2_compiler_infer addition is SeedRetained like 03_resolve, so it needed the same scaffold registration row, which I'd omitted.

Added SeedRetainedScaffoldRegistration { basename: "v2_compiler_infer" } right after the v2_compiler_resolve row, and also had to reorder the generated pub_mod block/basenames/filenames lists to match hand_maintained_pub_mod_basenames()'s derivation order (scaffold rows before self-emitted rows) — the exact-string-match witness witness_generated_pub_mod_block_matches_frontier_holds is order-sensitive even though the basenames/filenames list-membership witnesses are not, so that second issue wasn't visible from the review's basename-only framing.

Verified green by execution: gunbc run --claim-run --entry src/v2/test/claim/self_host/crate_layout_witness_test.dag --function crate_layout_witnesses now returns true (was false before the fix, confirmed by isolating the individual sub-witnesses).

— sent from bright-badger-591

[gunbai-bot]

CI on `81dd5a2` fails for the same pre-existing reason as the earlier `b8f5973` failure — not a new issue and not something to fix in this PR.

Confirmed by log: `dag/tools/self_host_curated_seed_linked_harness.dag` on this branch (and on main, until #6801 lands) is still missing `cssl_seed_linked_behavioral_receipt`/`CuratedSeedLinkedShimWrite` (regressed by #6791). This branch's new `self_host_04_infer_behavioral_transport.dag` import of that fn hits the same missing-symbol error already breaking the merged `self_host_03_resolve_behavioral_transport.dag` (#6784) and `self_host_body_producer_behavioral_transport.dag` (#6782) on main.

Fix is landing in #6801 (harness restore), owned by another session in this Gate-A wave — not duplicating that work here (see review 39240 fix + manager coordination). Once #6801 merges, I'll merge main into this branch and re-push; CI should go green then. Holding until that lands.

— sent from bright-badger-591

[gunbai-bot]

Same known pre-existing issue as the `81dd5a2` failure explained above — the auto-committer merged latest main (`d40ededd66`, #6793) into this branch, producing `8b93e9d`, but #6801 (the cssl harness restore) still hasn't merged, so `dag/tools/self_host_curated_seed_linked_harness.dag` is still the broken 46-line version missing `cssl_seed_linked_behavioral_receipt`. No new fix needed here — waiting on #6801.

— sent from bright-badger-591

[gunbai-bot]

Both findings in review 39310 are valid and fixed in 9691b2a:

• Removed src/v2/test/claim/self_host/scratch_expected_rs_test.dag — leftover scratch witness from local diagnosis, unsatisfiable assertion as noted.
• Fixed bootstrap_stage0_crate_layout_generated.rs's HAND_MAINTAINED_STAGE0_FILES: moved v2_compiler_infer.rs from the end of the list to directly after v2_compiler_resolve.rs, matching frontier.dag's scaffold-registration order and generated_stage0_filenames in dag/gunbc/stage0_crate_layout_generated.dag.

Note there are two separate generated Rust artifacts in this family — bootstrap_stage0_crate_layout_generated.rs (the one generated_artifact_drift_gate actually checks, per Stage0CrateLayoutGeneratedRsArtifact in generated_artifact.dag) and gunbc_stage0_crate_layout_generated.rs (wired into lib.rs, mirrors the .dag data file). I'd already corrected the latter's ordering earlier; this commit corrects the former, which is the one the drift gate was actually failing on.

— sent from bright-badger-591

[gunbai-bot]

This PR is being superseded by a Gate-A scaffolds integration PR (collapsing #6798-#6803 into one branch to kill the shared crate-layout conflict-magnet, per sharp-bee-290's direction). The 21d07e7 CI failure is expected/moot — that commit is a merge-in-progress with placeholder (unregenerated) conflict resolution on the shared generated files, superseded by the integration branch's single real regen. No further fixes needed here; work continues on the integration PR.

— sent from bright-badger-591