experiment(layout): diagnose mandatory leads and test bounded flat separation

2026-09-24 20:38:33 -07:00
parent 2850b915d6
commit 2be8c7f292
6 changed files with 898 additions and 41 deletions
@@ -0,0 +1,43 @@
+# Placement clearance experiment
+
+Continuation of issue32 from connection-layout-reference results at2850b915.
+The sparse external placement leaves20 units between rectangles where its
+final Rust top port needs18 units of outward lead plus12 units of unrelated
+obstacle clearance. This experiment addresses that necessary placement
+condition. It does not adopt either previous placement candidate.
+
+Use the router's actual endpoint selection, promotion/attachment policy and
+lead collision predicate in a narrow read-only diagnostic. Preserve existing
+router guard order, search, costs and fallback. Compare the five-case frozen
+route output exactly with the released-router control, and exercise targeted
+lead, alternative-endpoint and folded/ancestor tests.
+
+First consumer is an isolated example-only flat repair: at most256 nodes,
+1024 edges and four separation rounds. Freeze axis order and initial coordinate
+floors; seed separation inequalities from every initially disjoint node pair.
+Add a31-unit boundary gap (18+12+1 margin) for unambiguous cardinal singleton
+lead blockers and recompute monotone coordinates. Reproject actual automatic
+ports each round. Reject overlapping input, unsupported nested/folded/omitted
+topology, conflicting order, projection size changes, excessive extent/work,
+or unresolved constraints. No mutation escapes on failure. This narrower
+first experiment deliberately leaves subtree/compound coordinate assignment
+for a later planner increment.
+
+Evaluate all five frozen original and Arrange-reset scenes, all four supported
+old native candidates, and all three prior ELK geometry transfers, unchanged.
+Keep unsupported cases and failures explicit. Reconstruct route reset/retained
+translation once from input to final geometry using the audited import helper.
+Check semantic/anchor/other-view conservation, round trips, repeat determinism
+and ten repeated repair calls. Report all existing quality measures, actual
+port changes, warnings and unrelated-node penetrations. Require the exact sparse
+terminal conflict to disappear without introducing a new node overlap or routing
+failure before calling this compatibility increment successful. Lead clearance
+alone is not complete routability or readability acceptance.
+
+Stop after one bounded candidate and independent review. If the sparse witness
+cannot be repaired under frozen orders, retain the counterexample and move
+constraints into initial coordinate assignment; do not increase iteration
+counts or tune the corpus. Even success leaves prior shared-run, area,
+compound/undirected and native usability limitations unresolved.
+
+Windows only. No Linux build, release change or production Arrange adoption.