Keep contextual properties and creation bound to their semantic targets

harness/projection-navigation-20260925.md

@@ -29,8 +29,10 @@ Canvas inspection carries the originating route context and preserves view/layer
 Explicit model navigation continues to follow the target layer. Draft validation
 and stale-context rejection apply before inspection. Existing production group
 selection already retained the canvas; fallback edge/port selection and canvas
-context actions require the correction. Actual native reproduction must establish
-which path accounts for the user's reported click; source diagnosis alone does not.
+context actions require the correction. The published Windows baseline reproduces both reported failure states;
+source diagnosis and those baseline results do not establish patched acceptance.
+Contextual child/relationship creation must derive the semantic layer from the
+selected object, while root creation continues to use the active perspective.
 
 ## Acceptance
 
@@ -49,4 +51,12 @@ which path accounts for the user's reported click; source diagnosis alone does n
 
 Linux remains excluded by the existing user pause. Independent design critique
 identified saved visibility selectors and frozen migration as additional admission
-paths; both are included. Automated and native outcomes are not yet recorded.
+paths; both are included. At published candidate `603a59c1785dbcca324dae063ecc3405d9519a30`, 41 targeted core checks pass: legacy admission (2), projection switching (3), review regressions (15), source-6 view contract (8), and state diagrams (13). Tracked inputs were unchanged across the run. Whole-workspace formatting also passed at that commit. The first `0341bca` core attempt failed a generated-UUID test oracle; explicit fixture identities corrected the evaluator, without relaxing the admission assertion.
+
+The desktop attempt at `603a59c` was deliberately interrupted before test execution after independent source review found canvas Properties redispatch and contextual relationship/child semantic-layer defects. Its source inputs remained unchanged; no desktop passing result is claimed. The Properties issue arose from fixing explicit navigation to an already-inspected opposite-layer object. Keep both regressions: contextual Properties stays in the view, explicit navigation follows the object's layer.
+
+Published-preview executable SHA256 `2c03bd6cc3f1110c219e4f57ab94874a121590f83779239a93289b7170664399` reproduces edge selection changing SA to OA, and Architecture switching refused with `diagram_layer`. The isolated valid fixture remained byte-for-byte unchanged after the trial. An earlier fixture used reserved alias `capability` and was rejected; it was corrected before either reproduction. Initial native-dialog capture/input failures also remain separate from application acceptance. Independent evidence review confirmed the resulting screenshot states and persisted byte equality; screenshots alone do not capture the complete action sequence or transient state history.
+
+Candidate Windows GUI save/history/reopen, HTTP regression, broader affected suites and actual rendering acceptance remain outstanding. Raw local evidence is not yet a peer-consumable handoff; publish the consolidated evidence before integration.
+
+Scope boundary: contextual relationship forms retain the canvas while opening; successful Apply retains the existing explicit post-review navigation. The generic child-creation handler is tested for containment and allocation, but the separate architecture palette/performer workflow still offers creation kinds for the diagram layer. Thus an OA performer displayed in an SA Traceability view has no Add activity menu entry; this inherited limitation is not claimed repaired here.