8.9 KiB
Independent geometry review of the placement reference
2026-09-25. Reviewed recorded elk-reference-20260925/rust-run02 results at source 8a2010d8a6033201d6efd742134dae9f3d235632, plus original run01 engine receipts and selected raw ELK geometry. No compiler, router, engine or GUI rerun. Only read-only arithmetic on recorded points/rectangles was executed. Port/Project conservation is the separate gate reviewer's responsibility; this report does not substitute for that audit.
Decision: the imported ELK placement provides a real hypothesis for further connection-aware arrangement, but this output is not adoptable. The mixed cases improve several independent geometric problems through the unchanged Rust router. Sparse40 introduces a new unrelated-node penetration and NoClearPath, and uses substantially more area and total route length. Both compound workshop cases were rejected by the adapter before the engine ran, so this experiment supplies no external-engine compound-layout result. There is no scalar winner or native traceability claim.
Complete case accounting
Values in each cell are reset-only control / old native layered prototype / imported ELK placement. Shared length retains raw pairwise multiplicity, including common-terminal intervals. All numbers are recorded geometry, not readable-ink guarantees.
| Case | Proper crossings | Raw point contacts | Shared length | Label/node collisions | Result limitation |
|---|---|---|---|---|---|
| Workshop original | 6 /7 /unsupported | 27 /20 /unsupported | 1832 /1467 /unsupported | 1 /1 /unsupported | ELK adapter refused undirected edge before engine call. Native uses more area and route length despite fewer shared units. |
| Workshop folded | 7 /unsupported /unsupported | 19 /unsupported /unsupported | 1194 /unsupported /unsupported | 0 /unsupported /unsupported | Native refuses folded/omitted topology; ELK adapter refuses the same undirected edge. |
| Mixed dispatch | 10 /7 /1 | 25 /24 /7 | 716.39998 /4228 /238 | 0 /0 /0 | ELK gains require larger area; native exchanges crossings for extensive shared runs. |
| Mixed reversed IDs | 9 /7 /1 | 26 /24 /7 | 512.39998 /4228 /238 | 0 /0 /0 | ELK aggregate metrics match the original mixed case; this alone does not prove all corresponding routes identical. |
| Sparse40 | 94 /17 /20 | 158 /83 /55 | 1990 /40522 /666 | 14 /1 /0 | ELK introduces one unrelated-node penetration and one NoClearPath. |
The baseline here is explicitly reset_only_control, not the unreset after. In mixed dispatch, unreset Layout reports7 crossings whereas the reset control reports10/9; substituting the former would hide the reset/placement distinction. Original authored input is also not the Arrange baseline: original mixed has0 crossings/shared132, and original sparse has24 crossings/shared496 but70 estimated label/node collisions. Therefore “better than current Arrange” must not be shortened to “better than the authored input in every respect.”
The native prototype reports exact same-input repeats, idempotence, and equality to its first result for passes2–10 in all four supported cases. Folded remains unsupported, not stable by omission. Stability does not redeem its shared-run costs: mixed4228 and sparse40522 are real recorded totals, not nondeterministic noise. This run does not establish repeated-ELK-Arrange idempotence.
Actual gains and costs
For either mixed ordering, ELK reduces bends39→20 and total route length10618.60→9361.00. Estimated label/node, label/label and unrelated label/stroke collisions remain zero; no node penetration or warning is emitted. Area grows1139880→1808460, about59%. This is a plausible improvement in connection tracing with a navigation/overview cost still requiring native inspection. It shows that changing placement alone can help this graph through the released router; it does not require adopting external edge sections to obtain every benefit.
Sparse ELK reduces bends152→80; estimated label/node14→0, label/label1→0 and label/stroke31→0. These are substantial geometric improvements. However, route length rises39660→64223 and drawing area3804920→12840450, about3.37 times the baseline. The emitted bounds are8735 by1470 diagram units. Any later readable-scale trace test must account for the resulting navigation, rather than declaring readability from an uncluttered thumbnail. Native layered uses area6020910 and length54282, with fewer proper crossings17, but shared length40522 is much worse. Neither dominates the full vector.
Workshop original native reduces sharing1832→1467 and contacts27→20, but adds a crossing6→7, length8480→9251 and area886464→1125534. Its label/node count stays1 and label/stroke stays0. Folded reset retains four promoted-endpoint warnings and eliminates the original NoClearPath (original warning total5→reset4); neither alternative supplies a folded result. The failed ELK case receipts identify undirected edge 22c7011d-b559-5df8-aadd-851002904ba0; this is an adapter coverage refusal, not evidence that ELK failed to solve those compounds.
Independently checked geometry
Reused the independent axis-interval mathematics in audit-routing-corpus-contacts-20260925.py, without its route-only CLI or any routing call. Applied it to all12 available reset/native/ELK scenes. All48 comparisons agree within0.001: unique pairwise point contacts, proper crossings, common physical-terminal point contacts, and positive pairwise shared length. The arithmetic retains raw sharing; a zero common-terminal point-contact count does not mean shared terminal intervals are absent.
Independent classification for imported mixed is1 proper +2 T +4 bend contacts =7; sparse is20 proper +28 T +7 bend contacts =55. Thus most remaining sparse contacts are not proper crossings. They cannot be erased from interpretation by quoting only20 crossings.
For all nine flat mixed/sparse scenes, independently intersected emitted label rectangles with node rectangles, label rectangles with each other, and unrelated orthogonal strokes with label interiors. All27 collision-count comparisons agree. These checks concern the supplied estimated rectangles; they do not measure actual glyphs, port labels, selection hit behavior or arrowheads. Compound ancestor exclusions were not independently recomputed here.
Confirmed sparse blocker: relation flow_45, “Assessment Decide handoff to Planning Aggregate,” ID 00000000-0000-0000-0000-00000003347d, passes through unrelated “Assessment Package” placement 00000000-0000-0000-0000-000000030dd1. The node occupies x4522..4746, y660..820. The Rust route has a vertical segment x4634, y495..822, crossing160 units of its interior (Rust's epsilon-inset metric reports159.9998). Its actual source/target owners are different placements, so this is not endpoint or ancestor-body containment. The same edge emits NoClearPath; reset/native sparse have zero such penetrations and warnings.
The raw ELK section for that same edge has identical tips (4120,477)→(4634,840), but follows bends (4120,498), (4502,498), (4502,830), (4634,830). It goes left of the identified obstacle and then below it, avoiding that particular node. This supplies a concrete local witness that the placement had an avoidance corridor which the released routing result did not use. It does not prove that the raw ELK route avoids every other obstacle or that full raw ELK routing/labels pass our entire rubric.
Implication and bounded next decision
Do not adopt this placement-only transfer while the sparse hard failure remains, and do not conclude that all improvement requires a coupled engine: mixed supplies the contrary positive example. The sparse witness makes a targeted compatibility investigation worthwhile. Compare that one fixed scene/edge's released-router search/fallback constraints against the recorded feasible ELK corridor. Determine whether the lost route is a bounded router limitation or an incompatible spacing/clearance assumption before choosing a broader coupled-route integration. Preserve the frozen outputs; do not sweep ELK parameters to hide the blocker. Compound reference coverage still requires an explicit undirected-edge policy before any claim about hierarchy-aware arrangement can be tested.
No native task was performed in this review. No renderer-based legibility, timing winner, semantic-port fidelity, full conservation or production acceptance follows from the arithmetic. Those remain separate gates.
Evidence hashes, SHA256:
rust-run02/arrange-corpus/stdout.jsonl:f924af4038ade0de2e0f44988fa28d2566019d8c4bc2bde70e61fe4813364d40.rust-run02/mixed_dispatch-elk/stdout.jsonl:9308079aa4804d6a61fae724fc4fe44807528fc20b03e213ad08d5b4314c2d89.rust-run02/mixed_dispatch_edge_ids_reversed-elk/stdout.jsonl:567ef210022639083654e43c900ad82c7ba529597d437386c7b27baac2d4e58f.rust-run02/subsystems_40_sparse-elk/stdout.jsonl:37ef8db5a920a50a1ed12eba08913fee810d02226d920fc9018088c9e7b14c3a.run01/subsystems_40_sparse/elk-result.json:9763cda98e744040e1f2e9ad63662cf6e699e8fbdc7e77b388d60ab4303a4282.