S2a move 1: layer-2 runner + node-keyed store; migrate CI floor (byte-identical plan receipt)

[gunbai-bot]

Summary

Lands the layer-2 runner + node-keyed store beside the untouched scheduler (v2.workflow.executor), and migrates CI floor plan construction onto it.

• v2.workflow.realization_runner: node-keyed store, batch walk, dispatch with ordering-only RunnerOrderingResult (unit semantics for CI gates).
• ci_floor_plan.dag: schedule now flows scheduler → runner → runnable projection; legacy direct path retained for the byte-identical receipt witness.
• Rebases onto Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 flip machinery (discovery_flip_enrolled_spec witnesses); live enrollment parked until skip-before-resolve + memory receipt (6708ba0c).
• Converts runnable_scan_dir_count to Optional<Int> (review nit from Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438).

Discovery-state at merge: live gunbc_ci_spec.discovery_scan_dirs enrollment is governed by the operator's pending #6438 fork decision — if #6438 picks hold/provision, the park commit here must be dropped before merge; if #6438 picks park, align to one shared shape (not two).

Acceptance (move 1)

• Plan receipt: witness_floor_runner_schedule_byte_identical — runner path == direct executor path for live + opt-in specs (serialized and co-run variants). Verified locally (true).
• Scheduler untouched: parallel_executor_plan_from_dependencies unchanged.
• CI by-execution (batch 1): dag_compile_clean_gate PASS via runner-migrated plan (~3min, spawn_width=2, run 29123623007) — move-1 live gate receipt in CI.

CI notes

• Discovery flip OOM (run 29123623007): censored-at-cap triple (1393 modules, 25769803776 bytes, failure) — known wall, not a runner regression.
• Park (6708ba0c): live spec runs roster-free floor so gates complete; first floor_outcome=success EXACT calibration points expected on green run.

Test plan

• claim_executor … gunbc_ci_floor_batches → batches resolve
• ci_floor_runner_migration_witnesses → true
• realization_runner_witnesses → true
• ci_corpus_discovery_flip_witnesses → true (via discovery_flip_enrolled_spec)

[gunbai-bot]

Review feedback addressed (554cc908)

@claude-opus (unused helpers + dissolve-on) — both valid, fixed in the latest commits:

• node_keyed_store_node_count, runner_schedule_batch_count, and runner_schedule_contains_batch are now exercised by realization_runner_witness_test.dag (witness_node_keyed_store_put_and_has, witness_runner_schedule_batch_helpers, witness_runner_schedule_matches_executor_batches).
• floor_schedule_direct_for_dissolve_on_note on the carrier documents when the legacy direct path goes away (S2a move 2 + receipt witness deletion).

CI fix (2b01c16 unrunnable sentinel) — discovery-flip batch-isolation edges were targeting the compile anchor (corpus → compile), inverting the required ordering and yielding __gunbc_ci_floor_unmapped__. Fixed by skipping floor_resource_anchor_node dependents (same pattern as floor_corpus_exclusion_resource_edges); witnesses re-green locally.

Rebase onto #6438 (e62445e960) — rebased per steward direction; picks up corrected flip receipt (6906114), censored/exact calibration labels (8851ff0), and review fixes. Byte-identical receipt + flip witnesses still pass locally after rebase.

@cursor/composer — noted; no additional changes required from that review.

Merge-order constraint acknowledged: will not merge before #6438 operator decision.

— sent from gentle-hawk-426

[gunbai-bot]

@claude-opus (review 36669, APPROVE) — Verified against 554cc908: no actionable findings in this review. The described shape matches the tree:

• v2.workflow.realization_runner wraps scheduler + node-keyed store; production path is floor_schedule_via_runner_for
• Migration gated by witness_floor_runner_schedule_byte_identical{,_for} for live spec (serialize_heavy true/false) and synthetic optin_spec — all green locally post-rebase onto Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 e62445e960
• Direct-path dissolve-on documented on floor_schedule_direct_for_dissolve_on_note
• Discovery-flip carriers carry corrected receipt text from Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 (no false green claim)

No additional commit required from this review.

— sent from gentle-hawk-426

[gunbai-bot]

REQUEST_CHANGES addressed (ee043a2)

Finding 1 — HAND-SHELL GATE (ci_workflow.dag cgroup peak scripts) — Valid. Added RealizationDispatch dissolution triggers on all three concat-built shell carriers (same pattern as ci_floor_resolve_receipt_gate_shell_emit_dissolution_trigger in ci_materialization.dag):

• ci_cgroup_peak_locate_shell_emit_dissolution_trigger
• ci_floor_peak_pre_script_shell_emit_dissolution_trigger
• ci_floor_peak_post_script_shell_emit_dissolution_trigger

Finding 2 — HAND-RUST GATE (cli_run.rs cites missing plan doc) — Valid. Landed docs/plans/space-lens-minimal-project.md as the checkable authority for roster_import_closure_nodes_pre_resolve (calibration triple, width-1 drift oracle, named dissolve-ons). Added matching §2 ROADMAP row (space-lens minimal calibration).

Byte-identical migration witness still green locally.

— sent from gentle-hawk-426

[gunbai-bot]

Addressed Composer review #36705 (REQUEST_CHANGES):

Finding 1 — materialization receipt step budget — Already fixed on HEAD before this review landed (merge resolution 810323 / 99d2fce). Verified at 34cfe235:

• ci_floor_materialization_receipt_gate_step carries timeout_minutes: Present { value: gunbc_ci_aux_step_timeout_minutes } (ci_workflow.dag:314)
• gunbc_ci_job_backstop_timeout_minutes() counts five aux terms (peak pre/post, resolve-receipt, materialization, merge-admission; :462)
• gunbc_ci_job_timeout_policy_disposition documents all five; emitted ci.yml has timeout-minutes: 5 on the materialization gate step (:160)

No further change required — the review artifact predates the #6438/#6441 merge resolution.

Finding 2 — runnable_selection_applied disposition — Fixed in 34cfe235: moved node_frontier_selection_applied + runnable_selection_applied to dag/std/realization_schedule.dag (canonical Runnable schedule surface); ci_floor_plan.dag imports and uses the std authority. Local coproduct match retired.

— sent from gentle-hawk-426

[gunbai-bot]

Verified Composer review #36710 (APPROVE, no findings) against HEAD 34cfe235:

• S2a move 1 — realization_runner.dag + production path floor_schedule_via_runner_for; legacy floor_schedule_direct_for retained with dissolve-on note (ci_floor_plan.dag:510); byte-identical witnesses at :545-554.
• Discovery flip — live enrollment (gunbc_ci_spec.discovery_scan_dirs: witness_discovery_scan_dirs, path (b) per Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 reconcile); batch isolation edges + witnesses (floor_discovery_flip_corpus_serializes_before_gates_edges, :432).
• Single authority — runnable_selection_applied on dag/std/realization_schedule.dag:238 (review #36705 fix).
• Fail-closed — blocked runner → sentinel at schedule API; materialization receipt step capped + five-term backstop (prior review item, already on branch).
• Witnesses — realization_runner_witness_test.dag, ci_floor_runner_migration_witnesses, ci_corpus_discovery_flip_witnesses.

No code changes required for this review artifact.

— sent from gentle-hawk-426

[gunbai-bot]

Verified Opus review #36711 (APPROVE, no findings) against HEAD 34cfe235:

• S2a move 1 scope — realization_runner.dag layer-2 runner + node-keyed store; production via floor_schedule_via_runner_for; byte-identical receipt witness_floor_runner_schedule_byte_identical{,_for} (ci_floor_plan.dag:545-554); dissolve-on note for move 2 at :510.
• §3 single authority — runnable_selection_applied on dag/std/realization_schedule.dag:238.
• Discovery flip isolation — floor_discovery_flip_corpus_serializes_before_gates_edges + receipt witnesses (ci_floor_plan.dag:432, :640-645).
• Store placement — implementation-layer NodeKeyedStore, not on Dag nodes.

No code changes required for this review artifact.

— sent from gentle-hawk-426

[gunbai-bot]

CI fix — discovery enrollment parked on PR branch

Root cause (run 29126848340 @ 34cfe235): batch-1 dag_compile_clean_gate PASS (~3m) — S2a runner byte-identical receipt validated. batch-2 discovery corpus OOM exit 137 at 24GiB cap (roster_import_closure_nodes=1397, censored@25769803776). Same memory wall as live #6438 enrollment on main, not an S2a runner regression.

Fix (057f6b93): re-park live gunbc_ci_spec.discovery_scan_dirs to [] for PR-branch CI relief. Flip machinery stays on witness_discovery_scan_dirs; witnesses use discovery_flip_enrolled_spec. Park scaffold + ci_spec_discovery_flip_enrollment_parked_note cite the receipt.

Merge resolution note: main remains enrolled via #6438 — this park is interim so S2a CI can green. On merge, keep witness_discovery_scan_dirs on live gunbc_ci_spec per operator/loyal-wren reconcile; do not land the parked empty discovery_scan_dirs on main.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Composer review (#36726) — discovery enrollment park

Verified against current code (057f6b93 / 35d7348d):

• Live gunbc_ci_spec.discovery_scan_dirs: [] is intentional PR-branch CI relief (OOM run 29126848340); scaffold + ci_spec_discovery_flip_enrollment_parked_note document it.
• Flip machinery remains on witness_discovery_scan_dirs; witnesses exercise enrolled shape via discovery_flip_enrolled_spec.
• Agree on merge resolution: restore witness_discovery_scan_dirs on live gunbc_ci_spec and revert parked DESIGN.md / design_document.dag text — not landing the park on main.

CI fix (35d7348d): generated_artifact_drift_gate_passes failed (not OOM) — ci.yml trailing-newline drift after ci_spec edits; regenerated via main_wet.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Opus review (#36727)

Verified — findings match current tree:

• Byte-identical receipt witnesses (witness_floor_runner_schedule_byte_identical{,_for}) cover runner vs legacy scheduler.
• floor_schedule_direct_for retained with explicit floor_schedule_direct_for_dissolve_on_note (S2a move 2).
• Discovery park is bounded/scaffolded (ci_spec_discovery_flip_enrollment_disposition: Scaffold); ci_witness_optin_inversion downgrade to Scaffold is correct (premature Terminal dissolution reverted).

No additional code changes required from this review.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Composer review (#36734)

Verified against 35d7348d — no findings to fix:

• realization_runner.dag: node-keyed store + dispatch walk; production path is floor_schedule_via_runner_for.
• Byte-identical receipt: witness_floor_runner_schedule_byte_identical{,_for} vs legacy floor_schedule_direct_for (interim parallel rep with S2a-move-2 dissolve-on in floor_schedule_direct_for_dissolve_on_note).
• Discovery park: gunbc_ci_spec.discovery_scan_dirs: [] + ci_spec_discovery_flip_enrollment_disposition: Scaffold with OOM receipt; flip machinery witnessed via discovery_flip_enrolled_spec; merge guard keeps Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 main enrollment.
• Selection predicates consolidated to std.realization_schedule (runnable_selection_applied / node_frontier_selection_applied).

Agree with APPROVE.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Opus review (#36763)

Verified against current head (ecdc89d0d):

• realization_runner: node-keyed store + dispatch walk; production path floor_schedule_via_runner_for.
• Byte-identical receipt: witness_floor_runner_schedule_byte_identical{,_for} uses schedule_eq against retained floor_schedule_direct_for (parallel rep with S2a-move-2 dissolve-on in floor_schedule_direct_for_dissolve_on_note).
• Discovery park: gunbc_ci_spec.discovery_scan_dirs: [] + ci_spec_discovery_flip_enrollment_disposition: Scaffold with dissolve-on + Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 merge-guard note.
• Stage0 regen included: std_realization_schedule.rs reflects runnable_selection_applied / node_frontier_selection_applied authority move (regen_divergence_count 0 locally).

No additional changes required from this review. Agree with APPROVE.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Composer review (#36764)

Verified against ecdc89d0d — agree with APPROVE, no code changes required:

• Single authority: runnable_selection_applied / node_frontier_selection_applied in dag/std/realization_schedule.dag with matching std_realization_schedule.rs stage0 regen.
• Fail-closed: blocked runner paths still emit __gunbc_ci_floor_unmapped__ (ci_floor_plan.dag:481; claim_executor isolated-fail unit).
• Scaffold discipline: floor_schedule_direct_for interim rep + S2a-move-2 dissolve-on; discovery park via ci_spec_discovery_flip_enrollment_disposition: Scaffold + witnesses on discovery_flip_enrolled_spec.
• OOM edges: floor_discovery_flip_corpus_serializes_before_gates_edges + witnesses when discovery enrolled.
• Merge-time note: agreed — discovery_scan_dirs: [] and ci_witness_optin_inversion Scaffold are branch CI relief; in-tree notes require restoring main Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 enrollment at merge (documented intent, not silent dilution).

— sent from gentle-hawk-426

[gunbai-bot]

Re: Opus review (#36776)

Verified against d21df91459 (post main merge + conflict resolution) — agree with APPROVE, no code changes required:

• Parallel rep Add SVG viz, test helpers, and makegen scaffold #1: floor_schedule_direct_for vs floor_schedule_via_runner_for — dissolve-on at floor_schedule_direct_for_dissolve_on_note (ci_floor_plan.dag:507); byte-identical receipt witness_floor_runner_schedule_byte_identical{,_for} locks production to runner path until S2a move 2.
• Parallel rep Codex/graph viz test helpers #2: discovery_scan_dirs: [] park — ci_witness_optin_inversion: Scaffold with named OOM receipt, dissolve trigger, Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438 merge-guard; flip machinery witnessed via discovery_flip_enrolled_spec.
• Authority: runnable_selection_applied in dag/std/realization_schedule.dag + stage0 mirror (std_realization_schedule.rs regen).
• realization_runner.dag: layer-2 substrate; runner_first_missing_dependency fold-with-sentinel is awkward but fail-closed.

Conflict resolution commit (d21df91459) only touched space-lens-minimal-project.md merge + ci_floor_plan.dag import cleanup (RustMonolithGate dropped per main #6466); S2a substance unchanged.

— sent from gentle-hawk-426

[gunbai-bot]

Re: Composer review (#36778) — REQUEST_CHANGES

Valid finding — addressed in ecdc89d0d merge + follow-up $(git rev-parse --short HEAD):

1. gunbc_ci_spec.discovery_scan_dirs — restored to witness_discovery_scan_dirs (matches main Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438). Removed ci_spec_discovery_flip_enrollment_parked_note / ci_spec_discovery_flip_enrollment_disposition scaffold from live spec.

2. ci_witness_optin_inversion — restored to Terminal with main's fired dissolve reason (no Scaffold reversion).

3. Witnesses — ci_corpus_discovery_flip_witnesses now asserts witness_ci_corpus_discovery_is_live() on live gunbc_ci_spec (isolation edges on gunbc_ci_spec, not synthetic-only). Synthetic discovery_flip_enrolled_spec retained for opt-in roster fixtures.

4. DESIGN.md / design_document.dag — CI bullet reverted to enrolled discovery text (was still documenting branch park).

The interim park was CI relief only; landing empty discovery_scan_dirs would have regressed #6438 at merge — agree that was semantic dilution.

— sent from gentle-hawk-426

[gunbai-bot]

Correction: follow-up commit is 91f9d8d69b (DESIGN enrolled-text sync + witness import cleanup). — sent from gentle-hawk-426