Two-generation regen cutover + dissolve parked regen_verify_gate plan into executable RegenVerifyGate (cargo-green self-host fixed point, ONE PR)

[briansrls]

Auto-opened by session-dashboard for session eager-heron-723.
Pushing to session/eager-heron-723 advances this PR.

Worker attestation

Before flipping this PR to ready for review, confirm each item:

• 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

TODO: replace this paragraph with one or two sentences naming the change and its motivation. Reviewers read this first.

Test plan

• TODO: list the commands that ran (or "no tests changed; relied on CI") and the outcome.

[gunbai-bot]

Re: gate_spawns_host_compiler(RegenVerifyGate) => false — confirmed correct, the second interpretation in your note is the right one.

The predicate means "the gate body directly invokes the host language compiler on emitted DSL output," not "triggers a host cargo build during scheduling." Precedent: DslCompileCleanGate is also false (ci_floor_plan.dag:143) even though its transport calls ensure_gunbc_built() — same build-the-tool-if-absent prereq pattern as ensure_regen_stage0_built(). EmitHostGate is the lone true because its body runs host rustc/cargo on emitted code (the emit-host smoke). RegenVerifyGate runs regen_stage0 --verify, which does an in-process .dag self-compile + byte-compare of the committed seed; the regen_stage0 binary is already built by the floor's release step (cargo build -p v1-compiler --release --bins), so ensure_regen_stage0_built() is a no-op at gate time, not a host-compile of DSL output.

Also note the value is scheduling-inert here regardless: it's consumed only in gate_serializes_against_corpus = heavy_resolve || spawns_host_compiler (ci_floor_plan.dag:151), and RegenVerifyGate is heavy_resolve=true, so it already serializes against the corpus in batch-2. — sent from eager-heron-723

[gunbai-bot]

Thanks — same note as the earlier thread (issuecomment-4821358842). Keeping it false, deliberately, for consistency rather than dismissal:

The decisive precedent is DslCompileCleanGate — it's false (ci_floor_plan.dag:143) and its transport calls ensure_gunbc_built(), the byte-for-byte same build-the-tool-if-absent pattern as ensure_regen_stage0_built(). So the project's established reading of gate_spawns_host_compiler is "the gate body invokes the host language compiler on emitted DSL output" (true only for EmitHostGate, the emit-host smoke), not "the transport may cargo-build its own tool as a prereq." regen_stage0 --verify does an in-process .dag self-compile + byte-compare; the regen_stage0 binary is already built by the floor's release step, so ensure_regen_stage0_built() is a no-op at gate time.

Under that semantics false is correct and consistent. Under the literal "transport can ever shell cargo" reading, DslCompileCleanGate would be mis-set too — i.e. it's a predicate-rename/reclassification spanning multiple gates, not an asymmetric one-gate flip. That's a fine separate cleanup, but flipping only this gate would introduce the inconsistency, and it's scheduling-inert here (sole consumer ORs with heavy_resolve=true). Leaving it false to match precedent; happy to see a follow-up that reclassifies the predicate across all gates if reading-B is preferred. — sent from eager-heron-723

[gunbai-bot]

Thanks — this is a valid §5 catch, and you've identified a real seed-cement (the same class as the hand-maintained main.rs/v1_compiler_parse.rs carve-outs), not a deliberate scope change.

Root cause: compiler_tests.rs is a GENERATED stage0 file emitted from the .dag authority src/v1/compiler_tests_rust.dag. Both witnesses are present in main's committed compiler_tests.rs seed but absent from that .dag authority (confirmed by git: compiler_tests_rust.dag count = 0 for both on origin/main). They were never grounded in the authority, so main's stale seed was masking them; this PR is the first faithful self-compile, which correctly drops anything the authority doesn't emit. That is exactly what RegenVerifyGate exists to surface.

Resolution (in progress, not a silent drop): both witnesses will be restored by grounding them into compiler_tests_rust.dag so the regen emits them — the §5 'make the regen produce it' fix, not a hand-maintain cement. Notably contracts_sidecar_wired_into_emit_scope is the discriminating witness for the contracts-sidecar emit-scope wiring, which is precisely a separately-identified emitter alias-resolution-scope defect currently under operator review; once grounded it will go red on that defect until its fix lands, so it is being bundled with that fix. The grounding lands before this PR's final single re-regen (the behavioral bar is by-execution at that regen, since a .dag test-emission change can't take effect in the seed-built suite without it). I'll update the PR carrier with the dissolution note when grounded.

— sent from eager-heron-723

[gunbai-bot]

Update: sole_constructor_fieldless_newtype_witness is now grounded in the .dag authority (src/v1/compiler_tests_rust.dag, new ct_sole_constructor_fieldless_witness_test wired into compiler_tests_source), so the regen emits it rather than dropping it — the §5-correct restore, not a hand-maintain. Verified by execution: interpreting the authority fn directly (reads the edited source, no regen needed) emits the witness as valid Rust byte-for-byte; its behavioral pass rides this PR's single re-regen. contracts_sidecar_wired_into_emit_scope is being grounded in the same way, bundled with the operator-gated emitter alias-resolution-scope fix it discriminates (it lands red-until-that-fix-green; nothing red ships since the cutover doesn't merge until then).

Meta-point worth surfacing: this is precisely the class that RegenVerifyGate — the gate this PR adds — exists to catch. An ungrounded hand-edit to a GENERATED file makes committed-seed ≠ regen-fixed-point, which the gate flags RED at the introducing PR. So your catch here is exactly the gate doing its job one PR early; once both witnesses are grounded and the re-regen is green I'll re-request review to clear this.

— sent from eager-heron-723

[gunbai-bot]

Good question, and your hypothesis is exactly right — false is correct, and the resource-scheduling concern is already covered by a different flag.

In ci_floor_plan.dag, scheduling serialization is: gate_serializes_against_corpus(g) = gate_is_heavy_resolve(g) || gate_spawns_host_compiler(g) (line 152). RegenVerifyGate sets gate_is_heavy_resolve => true (line 129), so it already serializes against the corpus as emit-heavy work — the --verify self-compile is resource-scheduled regardless of the host-compiler flag.

gate_spawns_host_compiler is the narrower 'brings up the in-process host compiler' axis — only EmitHostGate is true. RegenVerifyGate shells out to the regen_stage0 child process through its ShellProgram/serialize_bash transport (it does not bring up the in-process host), so false is the semantically correct value — exactly the 'in-process host bring-up vs. child process' distinction you flagged. So no change needed; the emit-heavy weight is carried by gate_is_heavy_resolve => true.

— sent from eager-heron-723

[gunbai-bot]

Both findings fixed on head b39cc7db23 (commit regenerates ROADMAP.md from its authority).

F1 (roadmap done-state out of sync) — FIXED. dsl/gunbc/roadmap_authority.dag 5-regen-verify set done: true (the gate is wired into the executable CI floor and green on the merged tree), and ROADMAP.md regenerated from the authority via main_wet — line 223 now [x]. No hand-edit to the generated file; generated_artifact_drift_gate is clean (verified by execution, RC=0).

F2 (scope overstatement) — FIXED. Tightened the wording to match --verify's actual scope: "enforces no hand-edits to generated stage0 files (GENERATED_STAGE0_FILES); HAND_MAINTAINED_STAGE0_FILES are copied through, excluded by design." This matches regen_stage0.rs (only GENERATED files are byte-compared) and the gate's accurate failure text.

Note: this head also folds in origin/main (#5891 variant-owner grounding touched the same 04_infer.dag/v1_compiler_infer.rs region) + a 2-generation re-regen so the committed seed matches a fresh emit from the merged authority — the prior CI red was benign merge-drift (the gate correctly caught a stale seed vs main), now resolved (merged-tree CI run 28328259121 was fully green; full test suite incl. both #5891's variant_owner tests and the wire/variant tests passed, confirming the merged inference is coherent).

— sent from eager-heron-723