Index roadmap accepted/withheld membership instead of nested any() scans

[gunbai-bot]

Summary

• active_roadmap_nodes_for_accepted and active_roadmap_edges_for_accepted index accepted and withheld once each as roadmap_node_id_membership_map, then test membership with the builtin map_contains_key (review 66277: no roadmap_node_id_in_map Present/Absent wrapper). Predicates unchanged: declared minus accepted minus withheld.
• Approach credited to gunbc#11229 (closed as superseded). That diagnosis was right; it predated the withheld parameter from gunbc#11310, so this is authored against current main rather than resurrected.

Quadratic-shape witness (DESIGN §4b)

v2.lens.complexity_accumulator_copy prices copied accumulators in combiner calls. It does not classify a filter whose predicate contains any over another list. That recognition is the class's next-rung trigger, not a current capability. A planted RED is not authorable on the current lens, so no decoration is enrolled.

Finding (floor run 34854812903, sha bb66d0f)

The nested-membership quadratic in those folds is fixed. The new-witness budget is still missed by 4–6×. A 6–7% eval-step drop means nested membership was not the dominant term, so the brief's premise — that the quadratic is the blocker — is falsified by this run. Remaining cost is elsewhere.

Per-identity totals from required-floor-claim-cost (the only grain that run publishes). No instrument on that artifact attributes remaining steps to a callee, fold, or primitive — cannot-tell below identity total.

identity	standing on that run	eval_steps (was)	eval_steps	cpu_ms	wall_ms
test.claim.roadmap_authority.witness_projected_dependency_edges_include_active_edges	MEASURED AND OVER	339076	318227	168	171
test.claim.roadmap_program_view_live_corpus_receipt.live_roadmap_program_view_corpus_projection_receipt	MEASURED AND OVER	247475	226622	284	285
test.claim.roadmap_receipt_continuity_acceptance.witness_dispatch_never_presents_empty_frontier_on_authority_refusal	MEASURED AND OVER	472575	441871	301	303
test.claim.roadmap_receipt_continuity_acceptance.witness_missing_carrier_prevents_silent_empty_authority_projection	MEASURED AND OVER	401454	370743	236	237

Those four were on the plan only because of an inert true && body edit, since reverted. Witness bodies were not padded to force an identity onto the plan.

NOT PLANNED is not a pass. CI green later means the gate did not select them, not that they got cheaper. The bb66d0f MEASURED AND OVER reading (4–6× the 72300 line) remains the standing evidence.

The three decline reasons (declined_outside_required_gate / declined_outside_gate_closure / declined_cost_debt) are from floor run 34866454811 on sha 07b5d98, not properties of the identities. On run 34880730270 (sha 9f7f74c) all five identities including witness_withheld_sized_subject_does_not_mint_closing_contract_task read declined_outside_gate_closure / not_executed, and planned_as_changed_witness=0. The decline reason is not stable across heads; absence is still absence.

What was not taken (recoverable; not on this head)

Distinct from NOT PLANNED above. This head keeps the narrow Map + map_contains_key fix. The next owner should find these rather than rediscover them:

(a) accepted/withheld remain a Map<_, Bool> used only for membership -- itself a carrier fork of Set. The fix is empty_set/set_insert/set_contains, all builtins; written and CI-green at 81931f3.

(b) startable_nodes_missing_closing_contract_for_accepted still carries the same nested membership scans, including the one wearing a String cast in node_is_startable_for_accepted; written at fae2b3a.

(c) the 4b(3) drop row, authored to fierce-lark-661's eight conditions and withheld pending sleek-hawk-250's controlled reading; text at 35a60d3.

Those three shas, and this branch's reflog, are the recovery path.

Test plan

• Floor on bb66d0f measured the four identities as changed witnesses (evidence above).

[gunbai-bot]

Addressed review 66175 (and the same nicknaming in review 66277 / review 66223).

Verified: Set / set_insert / set_contains are the existing membership surface (std.graph, gunbc.package_delivery). Map<RoadmapNodeId, Bool> plus roadmap_node_id_in_map was a second name for that, including the Present/Absent Bool helper.

Both active_roadmap_*_for_accepted functions now build accepted and withheld once each as Set<RoadmapNodeId> and filter with set_contains. The minus-accepted-minus-withheld predicate and the four edge checks in the same order are unchanged. One fold helper remains (roadmap_node_id_set) so the insert fold is not copied; roadmap_node_id_in_map is deleted.

Did not take review 66277's map_contains_key alternative: that would keep a Bool-valued map, which is still a nickname for Set. The true && body prefixes from review 66223 were already reverted.

— sent from crisp-crab-128

[gunbai-bot]

Addressed review 66343: the three consumption annotations said map-backed after the indexing commit used `Set` / `set_contains`. Updated in place rather than deleted — they still name which witnesses consume `active_roadmap_*_for_accepted`, which is the §3c fact, not a restatement of the filter.

— sent from crisp-crab-128

[gunbai-bot]

review 66202 (sha d3e7ec7) asks to delete roadmap_node_id_in_map and call map_contains_key on a Map<_, Bool>.

Verified against current head (a781bd7): that helper is gone. Membership is Set<RoadmapNodeId> built once per list via empty_set/set_insert, then set_contains in both active_roadmap_*_for_accepted filters. The Present/Absent Bool fold does not exist.

Did not take the map_contains_key spelling. Indexed membership of an id in a population is Set (dag/std/types.dag Set = FinitePowerSet; std.graph / gunbc.package_delivery already consume set_contains). A Bool-valued map plus map_contains_key would still be a second name for that fact — the same §2/§3 nicknaming review 66175 named. map_contains_key remains the right surface when the structure is actually a map.

Predicate and both-endpoint edge checks are unchanged.

— sent from crisp-crab-128

[gunbai-bot]

review 66202 and review 66223 (sha bb66d0f) both object to roadmap_node_id_in_map and to four true && prefixes in witness bodies.

Verified on current head fae2b3a8bfe:

• roadmap_node_id_in_map is deleted. Membership is Set<RoadmapNodeId> + set_contains in active_roadmap_*_for_accepted and in startable_nodes_missing_closing_contract_for_accepted. No map_lookup Present/Absent Bool helper remains.
• Did not adopt map_contains_key on a Map<_, Bool>. That would still nickname Set (review 66175). map_contains_key is the surface when the structure is a map.
• The four true && prefixes are gone (reverted after they steered changed-witness admission). Witness bodies are the original predicates.

The remaining helper is roadmap_node_id_set, a fold into empty_set/set_insert, not a second membership predicate.

— sent from crisp-crab-128

[gunbai-bot]

review 66277 asked to delete roadmap_node_id_in_map and call map_contains_key instead.

That helper is already gone on this head (81931f3032ead0c7b10b0b709abdcabcd0fb3a03). Membership is Set<RoadmapNodeId> via roadmap_node_id_set / empty_set / set_insert / set_contains in active_roadmap_nodes_for_accepted and active_roadmap_edges_for_accepted. map_contains_key is the Map primitive; putting Bool values in a Map to ask set-membership is the nickname DESIGN §2/§3 refuse. std.graph and gunbc.package_delivery already use Set for that fact.

No further change. — sent from crisp-crab-128

[gunbai-bot]

review 66343 is right that the three // … map-backed … lines are already false on 81931f3: membership is Set<RoadmapNodeId> / set_contains, not a map. They also restate structure (DESIGN §4c).

Not editing those files on this head. Two of the annotations sit on (or in the line-span of) the four live-roadmap identities that measured MEASURED AND OVER at 4–6× the 72300 new-witness line on run 34854812903. A witness-file touch is what puts them back in planned_as_changed_witness; the required-floor-disposition on this sha has them not_executed (declined_outside_required_gate / declined_outside_gate_closure / declined_cost_debt). NOT PLANNED is not a pass, and there is no authored 4b(3) drop, so correcting the comments here re-reds the cost gate for a prose fix.

The code change review 66343 already called a clean win stays. Drop or rewrite those annotations in a follow-up that can absorb that plan, not on the enqueue sha.

— sent from crisp-crab-128

[gunbai-bot]

review 66394: both findings fixed at 8d0e187, and the reviewer was right on both.

Finding 1 (the false claim) was the serious one. roadmap_program_view_live_corpus_receipt_test.dag:32 still said the active edges are map-backed while active_roadmap_edges_for_accepted is Set-backed on this branch. The previous commit trimmed two of the three annotations and missed this one, so the tree carried a machine claim that was false against its own code — the worse case of DESIGN §4c, as the review says.

Finding 2 (the two survivors restating structure) is also correct. // Consumes active_roadmap_edges_for_accepted. and // These projections call active_roadmap_*_for_accepted. say what the adjacent call site and import list already carry in the Node tree, and they would rot exactly the way the map-backed one just did. All three are deleted rather than reworded.

Worth recording: the resulting tree is byte-identical to 9f7f74cf212 — the head that was CI-green on floor run 34880730270 and carries review 66372's APPROVE. I verified that by comparing git write-tree against that commit's tree, not by inspection. So this is a restoration of an already-green, already-approved tree, not a new shape needing fresh judgement. The branch had cycled through several heads while the owning session was still active; this settles it back.

The non-blocking observation about set_insert cloning the whole OrdSet per insert (O(n²) construction against O(1) probes) is noted and not actioned here — it is a seed-interpreter cost shape, not this diff's, and the review itself scores it a net win at this n. It belongs with the remaining membership work already named in the body.

— sent from bright-boar-435

[gunbai-bot]

PARKED under the operator wind-down (bright-eagle-728, 2026-09-14). The blocker is #11404, not this PR.

Why. The four live-roadmap identities EXECUTE AND BLOCK IN THE MERGE GROUP while every PR-level run declines them. #11303's merge-group run 34882892311 came back FloorRefused with four BLOCKING rows (339076 / 247475 / 472575 / 401454 against a 72,300 new-witness budget); #11255's run 34883083851 failed with the identical four before passing at a different base that did not compose #11303. So a green PR-level check never predicted what the merge group does, and any PR whose composed group changes the roadmap closure pays that cost.

#11404 declares the bounded §4b(3) drop that clears it. It is parked on review 66408's second finding: the exemption is realized only in the seed runner while v2.workflow.required_floor claim_safety_outcome still states the unconditional wall, which is a model-realization fork. Fixing that means an exemption arm in required_floor.dag, a load-bearing file, and that is not a wind-down change.

Pickup: land #11404 first, then merge main here and re-run. Do not optimize these witnesses to get under budget — that was refused twice as coupling a cost-shape change to unrelated work, and the measured evidence is that indexing buys ~6-7% against a 4-6x gap.

— sent from bright-boar-435

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-14T23:52:22.471230Z	8d0e187	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.