'Arrange' button don't take into account connection bends, connection labels and connection intersections #32
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: sorokin.oleg.edu/ARCADIA-Studio#32
Reference in New Issue
Block a user
#33 may be related
Codex-9e8r6t /root: additional independently reviewed routing evidence from the Process renderer, related to this issue's intersection concerns but not an Auto Arrange trial. With A→U and B→V, reorder the left column to B,A,C while U,V retain their right-column rows. Orthogonal routes share the middle x-coordinate/horizontal segments; portions appear interrupted and endpoint pairing is difficult to follow. Complete-Project checks confirm both link endpoints/guards remain unchanged, so this is a readability defect rather than reassignment/source corruption.
Published source/evidence at6d99aacab8ca1a91c857a2179baa14d80a2901e9 (productionff712b2): native screenshot39, independent review. Native fixture/oracle and full52 frames are in the linked native evidence package. No source-level root cause or routing fix claimed; retaining as a known limitation of #29 Process contribution.
Codex-ytr83r/root claims the automatic-layout outcome after reading the issue and both comments; no competing ownership is recorded. Scope: make the existing Auto-layout / Arrange elements operation account for connection geometry, labels and crossings, including an explicit policy for replacing obsolete bends, while conserving model semantics, source round trips and one-step Undo/Redo. Repository sorokin.oleg.edu/ARCADIA-Studio; intended branch codex/auto-layout-ytr83r-20260925 from released715036e. I will recheck after claiming. #36 remains a separate published candidate awaiting restored-window verification/integration, and #29 Scenario ownership remains with Codex-9e8r6t.
First milestone is to map the actual supported Arrange surfaces and current authored-route invariants, compare practical algorithms against representative engineering diagrams and define measurable acceptance before changing production. Current source uses hierarchy sizing, a four-column root arrangement and translated authored routes; it does not yet establish the requested bend/crossing/label optimization. The Process renderer overlap in comment1125 is evidence to account for, not proof that Auto-layout caused it. Independent research and source/evaluator review will challenge the proposed contract. No implementation or successful routing evaluation claimed; Linux remains excluded.
Codex-ytr83r/root: first diagnostic experiment and independent results audit are published on
codex/auto-layout-ytr83r-20260925at427078461a, tested sourcee7deea7039over released715036e. Results and reproducible command link the source map/research, independent audit and original output/receipts/37 offline geometry images. Remote branch and archive bytes verified; ZIP900995 bytes SHA256c52f82d4782cd6a7f8cbe67498c8b2371e99204a65976aa8bcbbd0a86f835069. No production implementation changed in these commits.Windows Rust1.98.1 gnullvm
cargo run --locked --offline -p oa-core --example auto_layout_probecompleted exit0 with unchanged source inputs. All9 small FunctionalDataflow fixtures conserve recorded model invariants and JSON/single-source roundtrips; the layered candidate is deterministic/idempotent through10 calls. Existing Layout's one-step Undo/Redo passes, but the pure prototype has no command/history/native UI acceptance. Linux excluded.The layered prototype improves reversed children in two containers (shared length148→0, label/card collisions1→0), nested cross-boundary labels1→0 and long-label collisions2→0, with area/length tradeoffs disclosed. It also REGRESSES shared connector length: diamond36→82 and bipartite276→344. All results have0 strict proper crossings, so no crossing-reduction claim is justified; a shared central trunk remains ambiguous in offline geometry. Reset-only already solves the authored detour but does not solve those ordering/label problems. This is useful experiment evidence, not an accepted replacement or issue closure.
Next bounded candidate tests a concrete source hypothesis: route-leg choice scores the source join but omits the target-stub join until after choosing the leg. Compare symmetric endpoint-join scoring at unchanged node positions/overlap weights, with independent review and explicit point-contact measurements to prevent hiding overlap as bend/T contacts. Alignment/spacing alternatives remain available if this does not address the regression. Folded/filter/State/Allocation, full command/API/source lifecycle and native readability remain later acceptance gates; #36 remains separate.
Codex-ytr83r/root: target-join experiment J is published at
9c02100d23on the same branch. Experimental, not ready for integration: this branch now includes a routing candidate with a confirmed ordinary-scene regression. Results, commands, independent source/results reviews and residual diagnosis. Source497c1ad8; observer-only lint cleanup08656b18. Both probe outputs are byte-identical. 185 Windows core/routing/State/group tests pass; strict core/all-target Clippy passes after3 observer warnings were corrected. Failed receipt retained. No Linux or native UI run.Independent comparison confirms all37 Projects, projected nodes/ports and route endpoints unchanged, isolating routing. Layered diamond sharing82→36 (nonterminal46→0); layered bipartite344→187(nonterminal272→115), but area165360→175960 and retrospective contact audit shows8 bend contacts→6 bend+3 T contacts. Other7 layered vectors unchanged. Crucially ORIGINAL bipartite sharing worsens730→872 while bends5→4 and length2076→2064; this prevents a monotonic readability claim. All original/new contact counts are retained, with an independent axis-based audit matching Rust across37 variants. Common semantic ports are real, not a distinct-port-placement defect.
Remote branch and original evidence archive byte readbacks verified:3178939 bytes SHA256bd2146237353af05b8667dc7bd728007143a2f1b24ad9f8728fdbc50ffb3368e. Next predeclared comparison uses overlap coefficient3.0 versus0.12 with cap80, crossing cost24, node/port geometry and edge order held fixed. This tests cost preference before adding parallel-lane candidates or changing placement. Preserve original-scene regression and contact counts as falsifiers; no production adoption or issue closure.
Codex-ytr83r/root: completed the predeclared J/P/PL/L comparison and published
f72aa2ca5concodex/auto-layout-ytr83r-20260925. Four-way result, commands and independent audits. Still experimental; do not integrate as a completed fix. The stronger coefficient is removed from branch tip; current code is J plus bounded lanes at original0.12/cap80.Source
e8ec6db: Windows188 selected core/routing/State/group checks passed,0 failed/ignored; strict core/all-target Clippy passed. PL sourced0be9ff0separately passed5 filtered tests(3 new lane checks+2 existing matches); no full-suite claim for experimental coefficient3.0. Every comparison preserves all37 complete Projects, projected nodes/ports and exact endpoints. All9 candidate repeat/roundtrip checks remain exact.Layered bipartite sharing J187/P187/PL176/L176 retains104 nonterminal units and9 contacts. Stronger penalty alone gives no layered benefit. PL's spacious baseline/reset result removes204 nonterminal shared units but adds1 proper crossing,+24 length,+19392 area; original0.12 L does not achieve that baseline tradeoff. Neither variant dominates all scenes. Original two-container P6 shared versus PL/L12 is another recorded tradeoff. Offline images remain geometry evidence, not native route-tracing acceptance.
Material cost signal: routing_geometry18-test suite took16.43s under L versus4.30s under J on the same machine/environment. This excludes compile time but is not an isolated benchmark. Retained coordinates are capped64 per axis; per-leg preprocessing scans/sorts all prior segments and needs assessment. We are stopping tuning these nine fixtures. Next scope is an independently frozen realistic nested/folded scene, mixed fan-in/out with fixed edge-ID permutation, and modest sparse scene with route-only timing/budget observations and explicit traceability criteria before deciding whether to keep L. Production Layout/reset/history/API/native gates remain open.
All original evidence and failed earlier Clippy receipt preserved. Remote archives byte-verified: lane evidence3518507 bytes SHA2565512b4cbd99e3bf03eb975e25fb980f7b1a95a5cb19f3687fb3d6d83d1839e8f; P evidence916568 bytes SHA2566b819b01c8012aa72d423443f4dc04ae261f1ee9dc6907f0063c219ff173ad36. All local runners closed normally. Linux remains excluded; existing prerelease frozen.
Codex-ytr83r/root: publication blocker at this checkpoint. Windows reaches Gitea, but push of the same owned branch was rejected by the remote unpacker:
unable to create temporary file: No space left on device;failed to write object. No force-push or deletion attempted. The new commits/evidence are LOCAL-ONLY and not available to other machines; last confirmed published experiment remainsf72aa2ca. Server storage needs recovery before this handoff can be published.Local results (not peer-consumable evidence yet): frozen five-case corpus; optimized released/J/L comparison completed with 210 measured calls. Maximum median3.2097ms, maxsample6.1959ms, all predeclared timing screens pass. Geometry tradeoffs prevent adoption: L workshop shared length1042→602 but label collisions increase; J/L reversed-ID mixed132→233.59998. Independent60metric comparisons agree. A narrow exact-f32 serialization audit resolved six observer JSON-equality failures; originals retained. Decision is keep released router for the separate reset candidate, no further tuning or production adoption of J/L.
Also prepared a separate local-only
codex/arrange-reset-ytr83r-20260925contribution based on715036e: existing Layout clears visible unpromoted pins/label overrides, preserves attachment/port choices and existing dormant-route translation. Router/node arrangement unchanged. 195 selected Windows core/routing/source/State/movement tests pass at local6c44a2d6; core/desktop strict Clippy is running. Independent source review done; desktop/native/API/multifile acceptance remains open. No issue closure, merge, release change, or Linux build.Codex-ytr83r/root: server storage recovery has allowed the owned research branch to push successfully from
f72aa2cato3e21c1bca2. Remote content readback is in progress; no integration request yet. This publishes the previously blocked released-router restoration, placement-reference/clearance/SCC research and current experimental boundary-order source. The existing prerelease remains unchanged and Linux builds remain excluded. Current boundary comparison and native observation evidence are still local until the next evidence commit. I will post the verified handoff after archive/readback. No completed Auto-layout fix or issue closure claimed.Codex-ytr83r/root — current work published and verified; stopping now at the user's explicit request. Server disk recovery restored Windows Git pushes and MCP writes. Both branches below are available to separate machines; no merge, issue closure, prerelease replacement or Linux build.
Completed narrow Arrange-reset contribution:
codex/arrange-reset-ytr83r-20260925at956e33e568, tested native sourced6f45a13. Result, validation and integration limits. Existing arrangement/router retained; visible unpromoted pins/label overrides reset, metadata/anchors and dormant-route policy preserved. 195 selected core tests, three desktop history checks, strict Clippy/builds, eight whole-Project native checks and eighteen screenshots are recorded with exact source attribution. Remote native archive verified: 6,529,778 bytes, SHA256 219c6dde234f1f30c4af191a538c21eacf874f234af0b8506bb66251db412650. Ready for integration-owner review; not a full solution to #32.Experimental layout research:
codex/auto-layout-ytr83r-20260925at4fcb65c0d3. Boundary-order results and evidence. This also publishes the previously blocked released-router restoration, placement-reference/clearance/SCC studies. Final candidate source3e21c1bc passes15 focused example tests, strict Clippy and optimized Windows build;20 supported candidates pass independent full-Project conservation, ten flat outputs remain exact, eight baseline streams byte-exact. Simple two-container sharing164→0 and chain328→0; compound workshop and label regressions remain, so DO NOT integrate as a completed Arrange replacement. Native saved-Project inspection confirms simpler tracing while retaining anchored border/label-background conflicts and local-SCC overview ambiguity. Four imported Projects/five captures; no experimental GUI Arrange invocation. Remote archive verified:156 members,4,310,136bytes,SHA256 bb6b896ff2401290869a904f7768ebb52b6f0ab36fada06badbe9abb301c37b8. Final review, result and manifest also match committed Git blobs exactly.All local compiler/probe/GUI runners are closed; GUI exited0 with empty logs. User requested finish/publish then stop; this session will resume only on user instruction. Remaining work: assess recorded compound/label counterexamples and production command/API/history integration before adopting a new planner. WSL DNS/direct-IP probes still failed in the latest check; successful Windows publication does not establish WSL recovery.
Codex-ytr83r/session01a0d706 takes a bounded read-only integration review of the delivered narrow Arrange-reset contribution956e33e568c6b799d0330803c74927e205977dfd (native tested productiond6f45a13), while unrelated projection/icon gates compile. This is not reassignment of the paused layout research, no source edits or new Arrange planner, and no merge/release decision. Acceptance: fetch/inspect exact published contribution, assess preserved bends/labels/anchors/dormant route behavior and recorded verification, report concrete blockers or integration limits. Parent will consolidate independent review on this issue.
Read-only integration review completed (Codex-ytr83r/session01a0d706, independent Astra critic; parent synthesis). Fetched956e33e568c6b799d0330803c74927e205977dfd, reviewed productiond6f45a13a7b8ab8dc9b84ee432a832609f5af6ed vs released715036e; experimental planner excluded. No material implementation blocker found. Eligible for later bounded integration review, not a completed #32 solution or merge approval.
session.rs freezes ordinary projected eligibility before movement, checks actual endpoint placements to exclude promoted/filtered/missing-owner routes, and clears only eligible existing waypoint/complete-label fields. Entry identity/order, attachments/anchors and semantics remain. Existing validation/atomic Batch/history/no-op Redo boundaries are retained. The normative language change and distinction from whole-entry Reset routing agree with implementation. HTTP uses the same core Batch preparation; this is source inspection, not a dedicated HTTP execution claim.
Independently verified published archive:6,529,778bytes, SHA256219c6dde234f1f30c4af191a538c21eacf874f234af0b8506bb66251db412650; all221 members match manifest lengths/SHA256 (CRC not separately checked). Raw records confirm195 core passes at6c44a2d6,3 desktop history passes at9eb4642e, final strict Clippy unchanged-input success atd6f45a13. Diff confirms only native example/desktop tests changed after core run. All8 exported/baseline JSON pairs rehashed equal; both close receipts report exit0, unchanged inputs, empty logs. Screenshots were not independently re-viewed; visual conclusions remain attributed to the published reviewer. Candidate-generated expected after-models establish wiring/history/persistence; independent contract tests supply reset correctness evidence.
Handoff correction required: results.md still leads with local-only/publication disk-blocked wording, which now contradicts this fetched published delivery. Reconcile current status while preserving historical failed attempts when integrating. Remaining boundaries: no combined workspace build, dedicated HTTP reset or hidden-port-marker test, native FPS, broad layout-quality acceptance. Shared contacts/crossings, Fit/Undo viewport and guide-discoverability limitations remain. Future integration owner must reconcile shared core/doc hunks and verify combined source; this review did not modify or merge anything.