experiment(layout): keep cyclic quotient groups compact in condensation order

2026-09-24 21:09:04 -07:00
parent ae9dad54b7
commit 5300a8c803
3 changed files with 687 additions and 1 deletions
@@ -0,0 +1,68 @@
+# Coherent cyclic-group placement experiment
+
+Source evidence atae9dad54 and the independent grouped-corridor audit change the
+next action. The old sparse candidate has four10-node strongly connected
+components, but its alias cycle breaking interleaves them across16 ranks. The
+single top corridor contributes38880/40522 pairwise shared units. Mixed has the
+same defect:4006/4228 at the top boundary. Test coherent compact SCC blocks with
+the released router before introducing a joint route-separation algorithm.
+This is a structural candidate, not another penalty or gap-repair sweep.
+
+## Fixed candidate
+
+- Plan each sibling group of the actual projected visual forest. Lift oriented
+  connections to immediate child-subtree representatives, compute SCCs, then
+  rank their condensation DAG. Never break a cycle by choosing an arbitrary
+  vertex in another still-blocked component. Undirected edges impose no rank
+  direction. Preserve all semantic orientation/endpoints and authored hierarchy.
+- Pack a multi-member SCC around a compact rectangular perimeter, using a
+  deterministic directed DFS successor order with alias/placement ties. Among
+  integer perimeter grids, prefer near-square physical aspect≤2 when feasible,
+  then minimum area; otherwise minimum aspect, with deterministic ties. This
+  rule is frozen before execution, not selected from measured corpus scores.
+- Use local edge-label estimates for gaps inside and between blocks. Minimum
+  gaps and child-header clearance reserve at least the18+12 mandatory-lead
+  requirement. Recompute parent enclosure from frozen minima. Core estimates
+  are not native glyph measurements. No global longest label inflates every gap.
+- Folded visible nodes are opaque with their exact compact projected footprint;
+  stored expanded sizes and latent descendants follow the audited original-to-
+  final import translation policy. Other omitted placements remain unchanged.
+  Final import alone reconstructs eligible route reset and retained translations.
+- Check actual final projection and mandatory leads. Geometry mismatch, missing
+  owners, bounds/work exhaustion and infeasible leads remain explicit candidate
+  errors; no implicit repair loop or changed port anchors. Scenario remains
+  outside this Arrange operation. Nodes/edges/work and grid enumeration are
+  bounded and candidate failure leaves the original unchanged.
+
+At a nested level these SCCs belong to the child-subtree quotient graph; they
+are a placement heuristic, not a claim about original leaf semantic SCCs.
+Cross-boundary edges and per-parent component/rank membership are reported.
+Nested grouping must not rewrite user containment or silently flatten it.
+
+## Evaluation and stop
+
+Use the same five frozen independent cases and the original nine small fixtures.
+Retain original/current Arrange-reset/old native results; add one SCC candidate,
+without parameter sweeps or changing the router. Report unsupported/error cases.
+Check complete Project conservation with only declared geometry/reset/retained
+translation exceptions, exact port-anchor and hidden-subtree behavior, final
+header/enclosure and node/route collisions, warnings, proper/T/bend contacts,
+shared intervals and pair multiplicities, labels, area and length. Record
+repeat-input equality and ten-call stability, including folded/compound cases.
+The two relation-ID variants must expose any stable-ID routing sensitivity.
+
+The structural witness is actual condensation order with compact component
+members and reduced dominant shared top corridors. A lower shared total alone
+is insufficient: preserve the full quality vector and identify any newly
+ambiguous pair, port failure, label interference or disproportionate area cost.
+Model/identity/history preservation is not evidence of readability. Do not
+adopt after compilation or metric checks alone; a viable candidate still needs
+production command/API/history integration and native route-tracing workflows.
+
+Stop after this one structurally distinct candidate and independent review. If
+coherent blocks only move sharing elsewhere or create hard failures, retain
+the counterexample. Joint derived route separation and virtual long-edge
+ordering remain distinct alternatives; choose the next action from the observed
+failure, not from commitment to this prototype.
+
+Windows only; Linux builds and the frozen published prerelease remain unchanged.