12 KiB
Connection-aware placement reference: result and decision
Do not adopt either placement candidate from this experiment. The external reference gives a useful mixed-network placement hypothesis, but it assumes port positions that Rust later recomputes and produces an incompatible sparse placement. The native prototype also has large shared-route regressions. The next design must account for automatic-port selection and mandatory exit/obstacle clearance during placement. Increasing routing search limits does not address the diagnosed failure.
This is branch-specific research for issue32, not a production change or an integrated release. Tested final observer source fc5c22c903c70b53102fa65e1b8cfdc971f877f1; first observer source 8a2010d8a6033201d6efd742134dae9f3d235632. Production routing remains byte-identical to released715036e. The separately verified Arrange-reset candidate is on its own branch. Gitea publication is blocked by the recorded remote disk-full/write errors; all new comparison records are local-only at this checkpoint.
Experiment and complete accounting
The plan and independent criteria were frozen before engine outputs. Use all five existing corpus cases, fixed engine/options and the released Rust router. elkjs0.12.0 supplies node positions only; its paths and label positions are retained diagnostically and never persisted as authored overrides. Current core-estimated label dimensions are supplied, not measured glyph bounds. This cannot rank ELK's complete drawing quality.
| Case | Reset / native / ELK+Rust proper crossings | Reset / native / ELK+Rust shared length | Outcome |
|---|---|---|---|
| Workshop original | 6 / 7 / unsupported | 1832 / 1467 / unsupported | Native reduces sharing but adds a crossing and area. |
| Workshop folded | 7 / unsupported / unsupported | 1194 / unsupported / unsupported | Old native prototype rejects folding; ELK adapter rejects an undirected connection before engine invocation. |
| Mixed dispatch | 10 / 7 / 1 | 716.39998 / 4228 / 238 | External placement improves several metrics, with wider bounds and incompatible automatic-port assumptions. |
| Mixed, reversed edge IDs | 9 / 7 / 1 | 512.39998 / 4228 / 238 | Same external scalar vector; semantic ID permutation remains recorded. |
| Sparse40 | 94 / 17 / 20 | 1990 / 40522 / 666 | External transfer adds a NoClearPath and unrelated-node penetration. |
Both workshop ELK attempts stop on undirected edge 22c7011d-b559-5df8-aadd-851002904ba0; neither is an engine failure or a successful compound-layout comparison. Do not infer compound compatibility from the three flat successes. The four supported native prototype cases are deterministic, idempotent and equal to their first results through ten calls, but stability does not establish quality.
For mixed dispatch, external placement reduces bends39→20 and length10618.60→9361.00 versus current Arrange/reset; estimated label collisions remain zero. Drawing area increases1139880→1808460, about59%. The original authored input already had zero proper crossings and only132 shared units, so improvement over Arrange is not improvement over every original quality dimension. Offline schematics show a more continuous sequence and long feedback paths, but no native trace task was performed.
Sparse external placement eliminates the reset control's estimated label/node14, label/label1 and label/stroke31 collisions. However, length increases39660→64223 and area3804920→12840450, about3.37×. Final bounds8735×1470 imply navigation costs that a thumbnail cannot assess. Its20 proper crossings coexist with28 T and7 bend contacts. The native prototype has shared length40522 despite only17 proper crossings. No weighted scalar winner is appropriate.
Two interface failures and one repaired observer defect
Automatic ports are part of placement, not fixed input geometry. All three final Rust node maps exactly match their imported plans, but mixed and reversed mixed each have10 of25 total port positions changed, including7 side changes. Sparse has48 of110 positions changed and32 side changes. All changed positions belong to automatic ports. Maximum discrepancy224 units; these are not rounding differences. All six authored anchored ports in each mixed case remain exact. The engine's FIXED_POS assumptions therefore do not describe the final Rust drawing. This is a placement/projection interface limitation, not loss of stored anchor metadata.
The sparse penetration is caused before path search. Edge flow_45 (…03347d) has target top tip(4634,840), requiring an18-unit outward stub at(4634,822). Unrelated Assessment Package occupies x4522..4746,y660..820; core12-unit inflation reaches y832. The mandatory target lead intersects that obstacle, so the sole semantic-port pair is rejected before cheap/grid search. The fallback reports NoClearPath and penetrates159.9998 units of the actual node interior. No search-budget exhaustion is involved.
Raw ELK avoids actual node interiors for this edge, but its10-unit final lead and10-unit obstacle clearance violate the core18/12 geometry contract. A local necessary separation is30 units rather than the observed20; that is not a proof a complete route exists at30, nor permission to blindly enlarge every gap. The fixed-scene diagnosis checks every raw segment against all40 nodes and the relevant inflation/endpoint exemptions. It corrects the initial, narrower observation that an uninflated avoidance corridor existed.
Identity transfer initially normalized stored sizes. Source8a2010d8 wrote effective projected minima into authored node sizes even for an unchanged plan. Sourcefc5c22c9 preserves stored size when effective size is unchanged and verifies final reprojection against the exact intended map. All five identity imports now equal independently reconstructed reset-only Projects, preserving stored expanded sizes. The eight import metric/geometry objects and final ports remain exactly equal to the first run, so this repair does not hide either interface failure. Finding and correction.
Validation and limits
Windows x86_64-pc-windows-gnullvm, Rust1.98.1, LLVM-MinGW20260908, one Cargo job, offline cached dependencies. Both observer source pins built successfully in release mode and passed strict example Clippy. Final commands, from the repository root:
cargo build --release --locked --offline -p oa-core --example auto_layout_probe
cargo clippy --locked --offline -p oa-core --example auto_layout_probe -- -D warnings
target/release/examples/auto_layout_probe.exe --arrange-corpus harness/experiments/routing-independent-20260925/corpus/manifest.json
target/release/examples/auto_layout_probe.exe --import-placement PROJECT_JSON DIAGRAM_UUID PLAN_JSON
Plan JSON is { "schema":1, "nodes":[{ "id":"UUID", "position":{"x":0,"y":0}, "size":{"width":224,"height":132} }] }, containing every visible projected placement exactly once. Import does not apply an intermediate Layout. Collapsed compact dimensions must remain exact while stored expanded dimensions stay intact. Hidden descendants follow their nearest visible collapsed ancestor's translation; unrelated omitted nodes remain unchanged. Nonzero hidden translation and nonempty dormant route translation are source-reviewed, not exercised by this corpus's failed compound reference cases.
Five control cases report exact Undo/Redo and source/JSON round trips. Eight successful final imports (five identity and three external) pass policy-aware whole-Project conservation and round-trip checks under independent review: eligible route resets are expected differences, so the raw authored_routes_exact flag is intentionally false for mixed cases. These are pure transformation checks, not native GUI/history tests of an integrated external engine. Six meaningful malformed-plan cases reject: missing, duplicate, unknown placement, zero dimension, changed collapsed footprint and final reprojection mismatch from an actually constrained height. An added width1 negative unexpectedly succeeded because that fixture has no such minimum-width constraint; its failed expectation is preserved, not relabeled. A separate height1 case correctly exercises the vertical-port minimum and exits101 with the expected reprojection error.
The first Python supervisor failed on Windows environment key casing before launching a child; corrected run and original failed script/capture remain. Original identity-normalizing run02 remains alongside corrected run03. All successful process stderr is empty; expected negative cases retain their errors. Final Clippy passed. No Linux, full workspace-suite or native readability acceptance is claimed.
The JavaScript adapter's eight synthetic checks cover nested coordinate conversion, shared physical-port reuse and key rejection paths. One fixed ELK run completed three flat cases in roughly307–624ms engine time, with process times roughly595–913ms; both unsupported cases exit before the engine. No timeout occurred. These are single diagnostic timings, not a native responsiveness bound or comparable Rust route-only benchmark.
Independent transfer audit checks complete Project differences, route policy, exact node maps, port drift and receipts. The geometry review independently matches48 contact/shared-length comparisons and27 flat label collision counts. It confirms the sparse penetration. Source audit, adapter review and parent observations are retained. Actual glyphs, port labels, arrowheads and named trace tasks remain unverified; no adoption follows from these metrics.
Reproduction and next decision
Frozen corpus manifest SHA256 a58a6956e26b414e88fafd4fdb92e006050653a4149178f08df84128e3917717. External adapter accepts node reference_probe.cjs CASE_NAME NEW_OUTPUT_DIRECTORY SOURCE_JSONL ELK_BUNDLE. The exact original source JSONL is archived; SHA256 9a31a676f23dea2f41daec39497f569d68231fa0fd66e5a1a476be97d5d61ac2. Fetch the pinned official npm elkjs0.12.0 using the recorded preparation helper/metadata and verify bundle SHA256 1222e44f953ce7746af23801e723708f8e6f436b8b377a6a5fc7552f34a307b3. Supervise every engine/observer process with the recorded90-second cap. Convert only structurally passing observations to the narrow plan schema, preserving all original results. Machine-specific runner paths in receipts are provenance; use the pinned commit and your own runtime paths.
Stop this configuration comparison. The next bounded work is a placement-compatibility model that evaluates final automatic-port choices and mandatory port leads against inflated obstacles, with explicit undirected/compound policy. First characterize the exact infeasible lead from this scene; then test whether candidate placement can satisfy those constraints without changing authored anchors or routing contracts. Do not pursue search-budget increases, option sweeps or a production engine dependency on the present evidence. Issue32 and the full application objective remain incomplete.
Evidence archive
ZIP and member manifest: 250 files, 3887968 bytes, SHA256 e68628c8558bb9a1104ed4ca0a0ce612bd0522a1cbce0b9d5c7a8092a7dfeb00. All member bytes, SHA256 and CRC were reread. Includes all failed/corrected attempts, four successful build/lint gates, raw engine/import outputs, independent reviews, source/driver bindings and17 offline geometry images. Binary payloads are excluded; their hashes and official engine retrieval metadata are retained. Final claim review and parent observations define the acceptance limits.