ARCADIA-Studio/harness/compact-context-help-source-review-20260925.md

Compact contextual help: independent source review

Review date: 2026-09-25. Scope: Gitea issue #39 compact editing chrome and contextual Help, reviewed against base 715036e1 and the issue screenshot. Reviewed source includes e0812b3932fbb36100516f0981dc9c646eeda56d plus the subsequent local passive-geometry admission patch and regression test in this change. This is source-review evidence, not native acceptance or executed test evidence.

Criteria: editing workspace UX review, especially preserving pending edits and selection, stable canvas geometry, accessible disclosure, and retaining actionable feedback.

Findings addressed in source

• The initial Help action used ordinary page navigation, clearing multiselection and selected bends, including before a pending behavior form rejected navigation. Contextual Help now overlays the existing workspace. Close/Escape changes reading state without navigating or applying drafts.
• App-level modal admission originally ran after allocation and reference-picker fast paths. Help now gates user messages at the beginning of App::update; allocation and behavior keyboard ownership are also disabled while Help is open.
• Combining live gesture feedback with multiselection controls could leave no width for feedback at the configured minimum window size. Selection controls now keep a compact, fixed-height row; current drawing/allocation feedback uses the existing application footer.
• Footer feedback initially replaced operation status, concealing errors while a promoted edge or drawing tool remained selected. The footer now retains operation status alongside feedback and exposes complete text in its tooltip.
• Conditional modal wrapping could recreate the editor widget tree. Both closed and open states now use the same outer Stack with the workspace as its first child.
• State/Process selection guidance removed from persistent chrome was missing from its target Help articles. The States and Chains articles now include group selection, primary-properties semantics and relevant tool/read-only guidance.
• The new early modal gate initially rejected passive behavior viewport/scroll notifications. It now admits Process/Scenario geometry and State viewport notifications to their existing context-validated handlers, preventing resize observations from being lost while Help is open.

Regression coverage added

context_help_tests.rs covers property drafts and multiselection, source drafts, pending state-transition forms, blocked background actions, contextual Help topic dispatch, Close/Escape, allocation input ownership, queued inspector reference-picker admission, and valid/stale Process viewport measurements during Help and after closing. Assertions include accepted-model snapshot identity and absence of history changes where applicable. Passive geometry intentionally validates document/owner/revision; stale pointer epochs alone do not invalidate measurements.

The reviewer did not compile or execute these tests. Suggested focused command: cargo test -p oa-desktop context_help_tests. The integration owner records actual execution and native results separately.

Follow-up from native review

Native review of the subsequent candidate found that contextual Help did not respond to PageDown and that Escape reached a focused underlying text input, dropping its typing focus. The follow-up patch adds a Help-only keyboard subscription and article-ID-targeted scrolling (40-pixel arrows, 90% of the actual viewport for PageUp/Down, Home/End), using Iced's bounds-clamped scroll state. A stable workspace input barrier preserves the actual editor caret while suppressing key, pointer, touch and IME input behind Help; modifier changes and passive geometry continue. Opening Help retires held pointer gestures through cancellation releases with all authored actions discarded, while forwarding redraw/layout invalidation. This also prevents a held canvas pan or text selection from resuming after release behind Help.

New mounted regressions exercise the production App tree and subscriptions for keyboard scrolling, unchanged background scroll positions, precise Inspector caret return, IME exclusion, Shift release, and held-pan retirement. A real text-input widget regression checks selection-drag retirement while preserving its caret. The allocation ownership fixture was corrected to explicitly request Allocation, because ensure_diagram creates Overview. These follow-up changes were formatted and source-reviewed here; compilation, execution and native acceptance remain the integration owner's checks.

A second, read-only critic reviewed the follow-up independently of its implementation. It found and prompted corrections for stale modifiers, IME leakage and interrupted pointer gestures. Its final read found no further blocking source defect in the input barrier, cancellation, passive-event handling or relocated Arrange control. The separate native reviewer supplied the original keyboard/focus failures. Neither source review establishes executed or native acceptance of the successor.

Scope and remaining verification

A final independent read-only source review covered the small post-5657a996 delta. It confirmed that the mounted typing fixture must send one KeyPressed per character because Iced consumes the first scalar, with the original caret/IME assertions retained. Arrange cancellation now replaces only its own obsolete armed status, preserving unrelated errors; input epochs and model commands are unchanged. The guide names the actual Arrange control. No blocking finding was identified in that delta. Actual executed and native evidence remains separately identified by source in the design record and evidence index.

Production changes reviewed: desktop App messages/admission, allocation input ownership, Escape bridging, behavior keyboard ownership, architecture/State/Process toolbar layout, guide article disclosure and footer feedback; supporting architecture/UI-review documentation. Regression tests and this record are review artifacts. No model schema, persistence format or authored-model transformation is introduced.

No further concrete source blocker was identified after the fixes above. Native checks remain necessary at 1100×700 and a representative normal size: Help discovery and keyboard reading/dismissal, focus and caret return, preservation of canvas camera/selection and form drafts, resize while Help is open, multiselection control visibility, allocation refusal/destination feedback, current operation failures, and stable canvas bounds through selection/gesture changes. Source inspection and App tests do not establish layout quality, keyboard scrolling, pointer behavior or usability.