CI fail-fast: cheap-gate early batch (SIMPLE declared membership + dissolution trigger) + event-scoped stop policy

[gunbai-bot]

Summary

• Cheap-gate early batch: declare gunbc_ci_cheap_floor_gate_membership (layering-imports, extdeps external-authority, generated-artifact drift) as a Scaffold with dissolution trigger; schedule them in batch 0 before compile-clean so fast scans fail in seconds instead of after the ~4min compile anchor. Cheap gates are resource anchors so discovery-flip corpus→gate serialization does not deadlock the scheduler (cheap→compile→corpus cycle).
• Event-scoped floor stop policy: model FloorBatchStopPolicy in gunbc.ci_spec (pull_request → StopBeforeDependents; main/falsifier/unknown → FullLedger). claim_executor evaluates gunbc_ci_floor_batch_stop_policy_for_github_event + gunbc_ci_floor_plan_uses_batch_stop_policy from the resolved plan context and consumes the variant in run_walk (seed bridge: gunbc_ci_floor_batch_stop_policy_claim_executor_seed_note + Scaffold disposition; dissolve-on: self-emitted claim_executor).
• GHA concurrency (orthogonal): model cancel-in-progress as github.event_name != 'merge_group' so PR/push runs supersede stale work but merge-queue checks are not killed.
• Schedule lens + witnesses: update schedule_lens for batch-0 cheap / batch-1 compile shape; enrolled-set preservation; RED control (compile gate absent from batch 0); stop-policy + roster witnesses; ci_concurrency_policy_witness_test; regen ci.yml concurrency line.

Test plan

• claim_batch --entry src/v2/test/claim/ci_floor_plan_witness_test.dag --functions ci_floor_plan_witnesses — PASS
• claim_batch --entry dag/test/claim/ci_concurrency_policy_witness_test.dag --functions ci_concurrency_policy_witness_holds — PASS
• CI floor on PR (post-ready)

[gunbai-bot]

Addressed review 41563 in b11ba515:

1. cancel-in-progress expression — valid §5 fail-open. The authority now emits ${{ github.event_name != 'merge_group' }} (not a bare string), ci.yml regen matches, and ci_concurrency_policy_witness_holds discriminates the unwrapped form as a RED control.

2. schedule_lens diagnostic — renamed the batch-0 violation to CompileGateInFirstBatch with schedule_lens_compile_gate_in_first_batch reason. Updated realization_schedule_witness_test synthetic plan to the 3-batch cheap→compile→corpus shape so the lens witness stays green.

Local: ci_concurrency_policy_witness_holds PASS, realization_schedule_witnesses PASS.

— sent from quiet-moth-636

[gunbai-bot]

review 41595 — observation 1 (plan roster parallel representation): Valid finding; addressed in the latest commit. plan_uses_floor_batch_stop_policy no longer hard-codes the three plan-function names — it evaluates gunbc_ci_floor_plan_uses_batch_stop_policy from the already-resolved plan context (same .dag-authority + seed-consumer split as gunbc_ci_floor_batch_stop_policy_for_github_event). The remaining seed parallel for the arm-time budget refusal roster (FLOOR_ARM_TIME_BUDGET_REFUSAL_PLAN_FUNCTIONS + floor_arm_time_budget_refusal_plan_functions_match_dag_seed_roster) is pre-existing §7 seed-retained scaffolding, unchanged by this PR.

review 41595 — observation 2 (concurrency cancel-in-progress vs FullLedger): Intentional orthogonality, not a conflict. These operate at different layers:

• GHA concurrency (ci_concurrency_cancel_in_progress_if): workflow-grain — supersede stale runs when a newer commit arrives (PR push, main push, workflow_dispatch). A cancelled run's ledger is truncated because the whole job is killed; that is desired (only the latest commit's verdict matters on main).
• Executor FullLedger (run_walk stop policy): within-run batch-grain — on push/schedule, if batch 0 (cheap gates) reds, record the failure and continue to compile + corpus before exiting non-zero. This is exactly acceptance point 4 (main still prints the complete gate+corpus ledger for the run that executes).

A superseded main push cancelling an in-flight run does not defeat FullLedger's purpose — FullLedger governs batch continuation inside the surviving run; concurrency governs which run survives. merge_group is the carve-out where cancel-in-progress is false (queue must not kill in-flight checks).

— sent from quiet-moth-636