ARCADIA-Studio/harness/experiments/state-label-layout-20260924/observations.md

State label native observations

Parent observer: Codex-ytr83r/root, Windows native host, 2026-09-24. Candidate source 70bdcdf8fec1728e971cd90ad7f79a1df99777ff; isolated profile and unrelated cwd per process. These are parent observations, separate from the machine-readable source/model checks. Original images and their bounds/timestamps are in trial/raw/ in the collected archive.

Inputs and method

The unchanged original peer fixture B has title Ready\nfor review and summary External · Completion [request_ready]. W changes only the title to 40 W characters. U changes only the title to Ready\r\n漢字 é 👩‍🔧\n. The stage independently validates their complete Projects and freezes their nine-file source closures. No model, route, or title editing occurred in the GUI. Save used Ctrl+S in place. Native movement M was not executed; exact movement/history remains covered by the mounted production-input regression.

Native trial used the actual public Fit and + controls. Displayed percentages are rounded, not exact scale measurements. Initial restored-window capture was 1442×842 at displayed79%; accepted coordinate-based checks used the maximized1605×872 window at displayed68% Fit and82% after one + click. No exact100% or DPI calibration claim is made.

Observed outcomes

• B: captures01/03/06 show two title lines followed by a separate summary, without the reported title/summary collision. In the maximized window, click(737,518) at68% selected State transition alias between (04); exposed entry-row click(582,519) selected State action alias action_a (05). After +, summary click(732,499) at82% selected between (06). These targets were read from the actual inspector, not inferred from route color alone.
• W: captures08–10 show the long single-line title within the enlarged white label extent and a separate summary. Click(900,514) in the far-right lower label area over State B at68% selected between (09), demonstrating the width-dominating title's hit area. Capture10 records82% appearance. Authored labels may still cover node content; this fix preserves their authored centers and is not an automatic collision-free layout claim.
• U: captures12–15 show Ready, the Unicode line, a blank trailing title cell, and the summary below it. The visible ink fits the label extent at68/82 without a clear self-overlap. Blank-cell click(737,514) at68% selected between (13); summary click(731,504) at82% retained the transition (15). These small-font witnesses do not establish universal font/shaping correctness or arbitrary Unicode ink containment.
• Ctrl+S completed with All source files saved for B/W/U (07/11/16). Checkpoints01-B-saved,02-W-saved,03-U-saved each report accepted=true/failure=null, full Project equality, and byte-exact preservation of all nine original source files for that variant.
• Parent closed run01 using Alt+F4. The launcher recorded PID5328 exit0, empty stdout/stderr, and immutable inputs unchanged.
• Fresh run02 PID6100 opened with Recent models listing the three exact working paths (17). After maximizing, B reopened through Recent (18); click(737,518) selected between again (19), then Save (20). U and W reopened through Recent in this fresh process (21/23) and were saved (22/24). They retained the visible label presentation. Checkpoints04-B-reopened,05-U-reopened,06-W-reopened each report accepted=true/failure=null and the same complete-model/source conservation.
• Parent closed run02 using Alt+F4. Its launcher also recorded exit0, empty stdout/stderr, and immutable inputs unchanged. Neither process was terminated programmatically.

Unsuccessful input attempts and limits

The initial restored-window click at requested(660,534), intended for B's summary, selected state_a instead (02). This attempt is not a passed label-hit check. On fresh run02, the restored-window Recent B Open button also did not respond to its screenshot-derived(1060,744) click. A one-off frame-offset hypothesis using(1052,713) likewise did not establish a successful Open. Maximizing through the observed accessibility control and clicking the newly observed Recent Open(1147,736) succeeded. This suggests a restored-window input-coordinate issue, but the cause is not proven; no product fix or universal restored-window hit correctness is claimed. The failed attempts are retained rather than recast as success.

Maximized checks, independent mounted hit regressions at multiple pan/zoom values, and complete Save/reopen comparisons form the scoped evidence. No native group drag, native Undo/Redo, exact1.0 scale, native generic-fallback-canvas interaction, Linux execution, or full-project usability acceptance is claimed here. The automated regressions separately exercise generic State targeting, fixed(+40,+40) group commit, exact Project history, and shared SVG positions. Wider State/Process congestion and dynamic row/header layout remain outside this bounded change.