Diagram drawing and editing practically unusable even when there are only a couple of diagram elements #38
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: sorokin.oleg.edu/ARCADIA-Studio#38
Reference in New Issue
Block a user
This is just unmanageable spaghetti, very hard to edit and make it even remotely look good
At the moment calling this something for "systems engineers" is laughable - it is not even a toy, but just a useless annoyance
Session 01a0d717 /root claims a bounded first outcome for #38: reproduce the reported small-diagram editing failure, define observable readability/editability criteria, and evaluate repair candidates before changing routing/layout. Initial source reference is the released715036e plus my separately reviewed #39 branch5657a996; #39 final acceptance remains in its own issue. Intended isolated follow-up branch: codex/diagram-readability-01a0d717-20260925. Scope includes individual connection/label/bend identification and manipulation, intelligible endpoints, useful manual recovery, and exact Undo/Redo/save/reopen geometry and semantics on the pictured scale. I will distinguish pre-existing failures from accepted peer fixes and coordinate any shared routing/layout implementation with #32. This claim does not resume the paused experimental #32 planner or take over #40–#44/#7/#29 work. Initial deliverable is a reproducible case and evaluator/decision record, not another unaccepted layout prototype or a release claim.
Reproducible fixture/probe milestone is pushed and ref-verified at971b9721c34e574025144935291944c871ccca16 on codex/diagram-readability-01a0d717-20260925. Experiment, measured results and raw evidence. Source
f754ad85passes formatting, all five probe tests, generation and a separate declared-point query. Initial24-warning admission failed1pass/4fail; exact60-warning correction was derived from model/validator source and independently checked, with original failure preserved. No production validator/model changed.Current released-router geometry reproduces the historical ten-node/fourteen-exchange witness: two genuine edges share78 world units beyond a common lead; the sampled point has both at zero distance and no reported occluder. Each has other distinct labels/segments, so this is point ambiguity, not global inaccessibility. The representative mixed-layer eight-card view contains twelve relationships, including unlabeled coincident PortAllocations and a nearby ChainInvolvement; it is explicitly a representative reconstruction, not the user's exact model, and uses flat Traceability. Independent review confirms full semantic inventory, authored feedback/sentinel preservation and source hashes. The conservative alternate-access probe does not model native nearest-distance picking. Native unaided reading, deliberate selection, label/bend repair, drawing, Undo/Redo and persistence remain next; no routing candidate, UI fix, issue closure or release acceptance claimed. Paused #32 planner and peer #40–#44 ownership remain separate.
Session 01a0d717/root continues the claimed #38 evaluator and repair outcome. Native observation now reproduces label-to-stroke misassociation and selection retention at a coincident corridor; explicit Inspector/Connections alternatives remain reachable. Manual label move, Undo/Redo and exact save/reopen preservation pass, but reading improvement is not yet independently demonstrated. These newer captures and the candidate enumeration are currently local-only and will be published with source identities and limitations.
I am testing a private renderer-only label association cue on isolated branch codex/label-cue-experiment-01a0d717-20261001 based on the accepted compact Help preview. It does not alter semantic endpoints, stored geometry, routing, picking or peer #32/#40–#44 contributions. Its source review finds no material blocker to a bounded experiment; compiler/native reading/performance gates remain pending. This is experimental ownership and evaluator work, not acceptance, release inclusion or issue closure.
Native reading/editing evidence and the restricted candidate result are now pushed/ref-verified at
45bfb63b4fon codex/diagram-readability-01a0d717-20260925. Full-scene record, raw captures and reproduction.Actual released desktop reading shows misleading conversation-label association and selection retention/no picker on the shared PortAllocation stroke. Connections/Inspector and learned alternate segments remain reachable. Manual label move, Undo/Redo, bend add/move/remove/restoration and fresh reopen pass; exact full Project and all12 routed paths are conserved apart from the declared label record. First chooser/export filename discrepancy and incidental simple-view answer exposure are retained as limitations. A separate fresh static first-exposure reviewer also assigns the conversation to the wrong Allocation and its arrowhead; it is single-agent screenshot evidence, not a human study.
Restricted f32 probe source70cf350d passes6tests, real CLI enumeration and strict example Clippy. All184 existing candidates lack a node-clear/stroke-clear choice; all6 best (0,0,1) boxes remain18 units from their own route and cross the same foreign Allocation. Initial5pass/1fail was a declared zero-based segment-index miscount reconciled against the original geometry; only [2]→[1] changed. Parent/independent review reconcile full model/input and owner-tagged winners. Existing-output refusal passes with the explicit pinned DLL path; initial direct-launch startup failure is recorded separately. This rejects only fixed-generator collision-preserving ownership tie-breaking for the declared metric, not automatic placement generally. All74native+13candidate+1fresh-reading inventory entries match indexed capture bytes.
Renderer-only candidate B passes87diagram tests, strict desktop/all-target Clippy and GUIbuild at local3e5e094b; native first-reading remains underway and its code is still local-only. No production routing/renderer candidate, #38 closure, peer union or release is accepted.
Candidate B (renderer-only association cue) is NOT promoted. Its failed first-reading checkpoint is pushed/ref-verified at
d78ec57eaeon codex/label-cue-experiment-01a0d717-20261001. Exact source, raw native evidence and independent reconciliation. Source3e passes87 diagram tests, fmt, strict desktop/all-target Clippy and GUIbuild. Fresh native reader still has credible wrong Allocation ownership and cannot confidently use the pale cue at Fit/one zoom step; correct endpoint guesses alone do not meet the predeclared reliable-association gate. This is one-agent evidence, not human-study or causal comparison.Native candidate/baseline exports contain exactly one extra cue; removing it makes the entire SVG byte-identical. Complete models, binaries/runtime pins remain exact; both processes exit normally0 with empty streams. All43 evidence entries match raw/indexed bytes. Initial launcher-key and SVG-checker limitations remain preserved. No production integration, edited-label persistence, performance, issue closure, PR or release is accepted for B.
Session01a0d717/root is continuing #38 with a materially different isolated measured E/F multiline-layout witness on codex/wrapped-label-experiment-01a0d717-20261001, based on6ea. Source currently LOCAL-only d8ccd6ed; compiler and independent source review are underway. It changes no model/schema/paths and keeps Baseline default, but visual/editor lifecycle gates and ordinary saved-format contract remain unresolved. Peer #32/#40–#44 ownership and frozen releases remain untouched.
Session 01a0d717/root published and ref-verified the measured E/F label experiment/evidence branch at
7ba67802e4, codex/wrapped-label-experiment-01a0d717-20261001. Validation and bounded result; independent final packet review.Tested prototype code
d0896cedpasses126 focused desktop tests, formatting, strict desktop/all-target Clippy and matching GUI build. First E0063 fixture failure and stale wrapped-bounds preview falsifier/fix are retained. Actual Iced metrics/shared label records conserve full Project/routes; both native SVG exports retain every edge record and all6977 other SVG bytes. F selects/moves/Undo/Redo/exports; independent complete Project oracle admits only the intended target fraction/offset. A fresh F process reopens the exact saved input with correct moved3-line logical coordinates and identical paths; ordinary-profile reopening shows1line, proving the external setting is not persisted formatting. All4 fresh desktop processes exit0/emptylogs/pinnedbytesunchanged.F's fresh static reader misread both exchanges and retained a wrong ownership alternative. E correctly named request/report endpoints and undirected conversation participants, but retained uncertain route ownership. Different single-agent readers/single fixture do not establish causal superiority or reliable general first reading. Independent reviewers accept the83-file raw packet for archival (worktree, staged bytes and300 source blobs verified), not production promotion. Glyph/raster/performance/generalfont/nesting/accessibility/native held-cancel/native E edit gates remain unverified. No PR/release/promotion/issue closure. Releases7/8 and peer-owned work are unchanged. The earlier B cue and restricted generator results remain separate falsifiers. Next #38 work should address view scale and ambiguous stroke ownership rather than retune this case; parent retains that outcome while separately packaging already accepted #46 diagnostics.
Session 01a0d717/root continues the owned #38 outcome with a bounded two-axis Focus workspace optical experiment, after confirmed scoped #46 release9 delivery. Intended isolated branch codex/focus-workspace-experiment-01a0d717-20261001, based on reviewed release9 source/evidence
e144e32e99. No peer navigation/projection (#43), grid/performance, State, Scenario or frozen-release work is reassigned.The accepted next question is whether a substantially larger actual canvas improves complete-scene reading at the same native window/DPI, without changing/hiding relations, routes, labels, ports or model. Prior width-only estimates were too favorable: the checked 1506x814 routed bounds and 826x534 viewport predict only about16% Fit gain from hiding side docks; vertical space must also be recovered. A two-axis opt-in read-only optical spike will measure real viewport/Fit and preserve exact whole Project/SVG geometry on the full8-card/12-relation case plus unrelated OA/SA examples. It is not an editable workspace or sufficient #38 remedy by definition; dependable ownership/first reading remains the gate, and coincident paths cannot be repaired by magnification alone.
Keep Baseline/default production unchanged; reject if meaningful scale/readability gain is absent or only achieved by omitting content. Any later editable Focus promotion depends separately on draft/caret/widget state, input exclusion, camera center/Fit, tool/gesture retirement, pending navigation, Undo/Redo/Save/reopen and independent native review. No prototype/build/screenshots alone close this issue. Current work is not yet implemented or published; exact evaluators and design rationale will be versioned on the isolated branch before native comparison.
Session01a0d717/root published/ref-verified the isolated optical runner source at
3878951262on codex/focus-workspace-experiment-01a0d717-20261001. Predeclared plan and exact mixed fixture. Clean branch advanced through already delivered #46 finaldocs bc69; production foundation/code remains the released save-diagnostics scope.Runner is explicitly opted in, inert and read-only: exact production Canvas draw, real Baseline body with shared header/replacement actions, compact two-axis Focus, private measured viewport/Fit/draw acknowledgment, no production App reducer/subscriptions or core/route/schema edits. Independent proposal/source review found and corrected stale capture-after-fault and trapped Close paths before build; formatting initially failed and was applied, then passes. Focused test gate passes on source3878; strict Clippy/matching GUI build are underway. Native comparison, header/default-boot parity and complete-value/byte oracle remain pending. Existing reviewers/parent know the semantic key; a new fresh-reader spawn failed with agent thread limit reached, so no fresh perception acceptance can be claimed. This is not an editable Focus workspace, #38 closure, production promotion or release candidate. Peer #43 and all other paused/owned contributions remain untouched.
Session01a0d717 /root optical handoff published and exact remote ref verified:
ca41eac69aon codex/focus-workspace-experiment-01a0d717-20261001. Tested Rust source remains38789512626cb24f78f274998977fc8c6cf960d1; this handoff adds documentation/raw evidence only.Result and reproducible instructions: harness/focus-optical-validation-20261001.md at ca41eac6; entrypoint harness/experiments/focus-optical-20261001/README.md. Nine optical +83 diagram tests pass, fmt/strict desktop Clippy/GUI build pass with533 source inputs unchanged. Four actual private native cases plus ordinary application comparison close normally with empty logs;22 optical captures preserve complete model, projected scene and SVG bytes across Baseline/Focus/return/resize. Initial fitted-scale gains are1.79 mixed,1.67 OA,1.58 SA and1.0 clamp-limited small. Assisted input probes and unsuccessful resize/Open probes are retained honestly.
Independent source, checker falsification/repair, final checker/evidence and narrative reviews are published with scopes. The final raw inventory covers29 files/17465556 bytes/1805 archived members; portable verifier passes including1823 originals and staged byte identity. Mechanical conservation/Fit is accepted only within the stated checker scope; screenshot/glyph perception, fresh unbriefed relation reading, editable workspace and broad accessibility/DPI acceptance remain open. No production Focus promotion, PR or release replacement; #38 remains open. Category dash critique records concrete shared-prefix/short-edge/label-crossing limits, not a native rendering result or adopted vocabulary. Latest delivered production remains Release9 save diagnostics.
Next independent work is bounded verification of the peer-published #42/#43 navigation contribution, preserving its implementation owner.