Required drift steps name their true population; record the split generated roster and the floor-at-ceiling finding

[gunbai-bot]

Interim truth-in-labeling fix plus two recorded findings. This PR does not restore stage0 mirror coverage. The operator has ruled that regen must cover all of stage0. Where it runs (per PR or merge_group only) is still waiting on an operator cost ruling, and that change will land separately.

Step names now state their population

generated_artifact_gate main reads only gunbc.generated_artifact committed_generated_artifacts. The ~157 stage0 mirrors are adjudicated by claim_executor's GeneratedArtifact phase, which routes to --required-lane build, and no job has run that lane since #11742. Under its old name, "Generated artifacts match their authorities", the step stayed green over that gap, and std_measure.rs drifted on main (repaired by #12067). Both steps are renamed in their authorities, gunbc.compiler_gate_workflow (drift step and repair regen step) and gunbc.heal_workflow (regen step). witnesses.yml and heal.yml were regenerated with generated_artifact_gate main_wet, using gunbc built from this tree; those two YAMLs are the only generated artifacts that changed.

New recurring_failure_mode rows

• generated_population_split_across_two_unjoined_rosters: two rosters answer "which committed files are generated", and a filename does not tell you which one owns it. The harm is partial coverage that reads as total. The next-rung trigger is one roster: the stage0 emitted population derives its membership into gunbc.generated_artifact. Re-enabling the build lane alone does not retire the row.
• required_floor_runs_at_its_memory_ceiling: in merge-queue run 35716368507 (slot actions-runner@srv4-03), the floor's slot peak was 26847006720 bytes, at its memory.high of 26843545600, with 1109 throttle events. The floor has no headroom left for its own growth, and its wall time is partly throttling. The row is re-derivable from the job's [floor-cgroup] beats. It is recorded here and not ranked.

The separate-process constraint is a NOTE, not a wall

A comment on compiler_gate_drift_step says the stage0 phases fit only as a separate process after the floor exits: sequential peaks (~12.9 GiB after the floor) fit, while additive peaks on top of ~25 GiB would not. Nothing enforces it. claim_executor --required-ci with no --required-lane runs every phase in one process (required_ci_phase_selected(_, None) => true), so the additive form can still be written.

Stale DESIGN prose, not edited here

DESIGN.md "Building & checks" still says CI runs --required-lane build; it has not since #11742. DESIGN is a projection of gunbc.design_document and should be corrected at that authority together with the routing change.

🤖 Generated with Claude Code