Stage0 mirrors are adjudicated on merge_group: the build lane as a step of the floor job

[gunbai-bot]

Adds full stage0 regen coverage (the build lane, both generations) as a merge_group-only step. The sign-off is operator decision msg_6e36a6a2 (2026-09-23). It supersedes msg_4a9045fc (generation 1 only) and the ~20 min pricing of msg_f1ade7cf-77e4-4896-a8b3-476747692c37. It covers exactly this step on this event, and nothing wider.

Land AFTER #12157 (the stage0 mirror resync). Until #12157 lands, this step is red on main's drifted mirrors.

The first live merge_group run caught real drift. In its own merge_group run 35973487586, this PR's step refused the composed revision with stage0-mirrors FAIL generated surface drift: v1_compiler_emit_rust.rs. Main had drifted again after #12157, because nothing enforced the mirrors yet, and the queue dropped the PR (failed_checks). The chicken-and-egg is resolved by carrying that resync in this PR: v1_compiler_emit_rust.rs is regenerated by claim_executor --regen-round-cost (changed_paths=1, one convergence stage) on origin/main 90bd54965cd. On the merged tree, the step's exact command claim_executor --required-ci --source-root dag --source-root src/v2 --required-lane build then gives stage0-mirrors first_generation_equal=true, generated-artifact OK, fixed_point_equal=true, phases_failed=0. The source commits touching the emitter since #12157 are #12180 (3e36b84e7b1) and 36dd24afc38; I did not bisect which one moved the mirror.

What changed

• gunbc.compiler_gate_workflow compiler_gate_stage0_regen_step: one step on the existing floor job, running the build lane's own command (gunbc.witness_floor_workflow build_lane_run_script → claim_executor --required-ci --source-root dag --source-root src/v2 --required-lane build). The binary's exhaustive partition decides which phases run: GeneratedArtifact adjudicates every stage0 mirror against the first generation, and RegenFixedPoint compares a second generation against it. No job had run this lane since CI: required witnesses check builds only the compiler, on a hosted runner #11742.
• The step's condition is !cancelled() && github.event_name == 'merge_group' && steps.build_witness_fold.outcome == 'success'. It runs as a separate process after the floor exits, which is the measured memory constraint. A floor red cannot disarm it.
• Not a new job: floor already checks out the tree and builds claim_executor. The rationale is recorded at witness_floor_lane_jobs. No timeout was raised.
• witnesses.yml and DESIGN.md were regenerated with generated_artifact_gate main_wet, not hand-edited. The stale DESIGN "Building & checks" sentence is corrected at gunbc.design_document.
• New claim test.claim.compiler_gate_workflow_witness_test w_RED_the_stage0_mirrors_are_adjudicated_on_merge_group_only_after_the_floor. It checks that the step runs --required-lane with no measurement receipt, is gated on merge_group with !cancelled(), and comes after the floor's producer. Locally it returns true. It went RED on the real floor path (run 35890770775) on a branch whose step ran a different command, so it discriminates.

Evidence (fleet runs; throwaway drafts widened the event gate to pull_request through the authority, and all are closed)

control	run	stage0 step	detail
clean tree = green (#12104 + #12157, full lane)	35890770775	success, 22.5 min	first_generation_equal=true, registry 49/49 match, fixed_point_equal=true, phases_failed=0; the floor job took ~58 min against its 90-min timeout
planted drift = red (+1 comment line in std_measure.rs)	35902516565	failure, 20.8 min	FAIL generated surface drift: std_measure.rs (only that file), phases_failed=1
PR runs skip it (this PR's own pull_request run)	35810834019	skipped	floor succeeded

The ~30 min first generation measured earlier was inflated by the field-name re-export leak that #12100 fixed.

Along the way, the first attempt found 32 drifted mirrors on main. That drift traced to the #12089 emitter defect (fixed by #12100) and to an installer gap for the crate-less manifest (fixed by #12146). #12157 resyncs the remainder.

Rung drops: none retired

• floor_cut_regen_second_generation_agreement: its trigger is second-generation refusal for EVERY phase that regenerates a committed population, shown by a red on a generation-unstable fixture. RegenFixedPoint compares a second generation for the stage0 mirrors only (it reads target/stage0-regen-receipt.json). The registry projections get no second generation, and no unstable fixture was run. Not met.
• rust_unit_tests_off_the_merge_path (runner supply for the --lib tests) and witness_floor_off_the_required_gate (the whole witness population on the landing revision): not touched, not met.

🤖 Generated with Claude Code

[gunbai-bot]

Control results: this PR must not land as it stands. Converted to draft.

Positive control #12102 (clean tree, run 35810815149, floor job 107021735980):

• required-regen: elapsed_ms=1805940 ... planned=157 executed=157 adjudicated=157. The generated-artifact phase FAILED on a clean tree: 32 stage0 mirrors on main already drift (extdeps_languages_rust_representation.rs, std_algebra.rs, std_types.rs, v1_compiler_{compile,emit,emit_rust,infer*,parse,tokenize}.rs, v1_std_core.rs, …). This is the drift that went unobserved after CI: required witnesses check builds only the compiler, on a hosted runner #11742. Until the mirrors are re-synced, this step would red every merge_group run.
• Cost: generation 1 took 30 min (compile.emit 26 min). The regen-fixed-point phase then started a second emit and was cancelled at the 90-min job timeout. The floor job's steps before the new step took ~41 min, so the lane is ~60 min, not ~20.

Negative control #12103 was cancelled the same way, so it discriminates nothing beyond the positive control's red.

Event gate (the parent's point (b)): on this PR's own pull_request run 35810834019, the floor job succeeded and Stage0 mirrors match… concluded skipped. The emitted if: is !cancelled() && github.event_name == 'merge_group' && ….

Blocked on: (1) re-syncing the 32 mirrors, and (2) a cost decision, since the operator's merge_group ruling was priced at ~20 min.

[gunbai-bot]

On the note in review 70794 about transcribed measurements: fair under DESIGN §6. Each figure in the rationale sits beside the run that produced it (35890770775 for the 22.5 min lane and ~58 min job; 35716368507 for the memory.high peak), so the producer is named and a reader can re-derive the figure there. I'm not re-pushing for it now, because a new head would restart CI right before re-queue. If you want the bare numbers removed and only the run IDs kept, it's a comment-only follow-up.

— sent from neat-carp-265