Record SCC placement evidence and bounded transfer correction

harness/DECISIONS.md

@@ -546,3 +546,32 @@ compatibility as a necessary placement constraint in a future initial coordinate
 assignment with connection ordering and corridor/label space. Compound geometry,
 automatic side changes and native tracing remain separate required evidence.
 This is branch-specific experimental evidence, not an integrated capability.
+
+## Keep cyclic groups coherent before assigning global ranks — 2026-09-25
+
+The [SCC placement experiment](experiments/scc-placement-20260925/results.md)
+at `2426d286` supports condensation of sibling-subtree cycles as a useful
+structural step. Independent reconstruction finds four ten-node components in
+the sparse case. Compact blocks remove the old layered prototype's shared top
+corridor: raw pairwise sharing falls from 40,522 to 32, with fewer crossings and
+less area than that prototype. The released router is unchanged. The full
+comparison includes original and Arrange-reset controls; this is not a universal
+improvement across all measures or scenes.
+
+Preserve effective geometry at the placement-transfer boundary. An old
+projected size can depend on children that subsequently move inward; retaining
+the old authored size then loses an explicitly requested minimum. The corrected
+example importer makes one conditional simultaneous minimum restoration and
+still demands exact final projection. All 14 frozen cases now complete; the
+13 prior successful outputs remain exactly unchanged, with conservation,
+roundtrip and ten-call checks recorded. This is an experimental transfer
+contract, not a production command or history integration.
+
+Do not adopt the whole candidate yet. The two-container case still ignores
+external-neighbor order within each parent and is worse than the old layered
+zero-sharing witness. Sparse residual label conflicts, folded route retracing,
+bipartite sharing and area tradeoffs remain distinct issues. The next bounded
+comparison should preserve coherent SCC ranks while using actual cross-boundary
+neighbors to order siblings. Treat label association separately from route
+sharing, preserve the full quality vector, and retain native tracing as a later
+acceptance gate. No global gap or routing-penalty sweep follows from this result.