Declare dependency-aware boundary ordering comparison

2026-09-24 21:48:07 -07:00
parent f4fa0c11f2
commit c02a03c553
1 changed files with 93 additions and 0 deletions
@@ -0,0 +1,93 @@
+# Incoming boundary order experiment
+
+Issue32, following the SCC checkpoint f4fa0c11. Keep the full OA/SA product goal
+and production/native acceptance requirements unchanged. This is one isolated
+planner comparison with released routing, not a quality-selected parameter sweep.
+
+## Question and fixed candidate
+
+The SCC planner drops a connection from a sibling group's local graph when the
+source is outside that group. Root SCC ranking knows which container is upstream,
+but the children have no information for matching order across the boundary.
+The existing two-container witness has164 shared units and one label/node
+collision, while the old layered candidate's incoming-neighbor order has zero
+for both. Preserve the SCC candidate's sparse40 structural improvement while
+restoring this missing information.
+
+Start from the same canonical SCC plan, before import. Never change its SCC
+membership, condensation ranks, internal perimeter walks, node sizes, gaps or
+root layout. In each nonroot same-rank band, bundle undirected-connected SCC
+blocks as indivisible units with their existing internal order. Unanchored units
+retain their ordinal slots. Reorder only anchored units among their old slots,
+then repack by prefix sums of unchanged heights and gaps. Old absolute Y slots
+cannot be reused for differently sized blocks. Envelopes are mathematically
+invariant; finite arithmetic and exact final projection remain checked bounds.
+
+An anchor is an oriented incoming projected edge whose target lies under the
+current sibling group and whose source does not. One distinct visible source
+placement contributes one vote per unit, independent of parallel relation count.
+Scores are the mean of current planned absolute source-center Y coordinates,
+summed in stable placement-ID order using f64. Break equal scores by original
+unit order. Undirected external edges supply no invented direction or score.
+This is a stated center-based heuristic, not a claim to match final port slots.
+
+A group can change only when a band contains at least two anchored units.
+Create dependencies from every such active group that owns an offset along an
+external source placement's entire ancestor chain to the consuming group.
+Include the immediate containing group; do not invent additional hierarchy arcs.
+Condense this influence graph and process its SCC DAG deterministically. Before
+each SCC, compose a complete absolute-position snapshot from current offsets.
+All member groups calculate proposals from that snapshot before committing any
+of them. Each group commits once. Cyclic influence therefore has one simultaneous
+step, an explicit limitation rather than an unbounded convergence loop.
+
+Recompose descendants through their ancestor offsets. Perform one final original-
+to-final placement import, preserving the existing reset/hidden/retained-route
+contract and conditional effective-minimum restoration. Exact final projection
+and actual mandatory leads must pass; no size/gap/anchor repair is added.
+Apply the existing graph, work and extent bounds to new dependency/snapshot work.
+
+## Controls, falsifiers and acceptance
+
+Retain all14 original cases, their SCC run02 outputs at2426d286 and the original,
+Arrange-reset and old-layered controls. Flat/no-boundary outputs, particularly
+sparse40 and both mixed-ID cases, must preserve complete Project, geometry and
+metrics exactly (new diagnostics are separate). Existing SCC-mode and released
+route-mode output must remain byte exact after adding the new observer mode.
+
+The existing two-container witness requires Left A before Left B, Right B before
+Right A, with actual route sharing/contact/label improvement assessed against
+the old zero-sharing witness. Swapping both groups is not that success.
+
+Add three independent topology falsifiers before execution:
+
+1. Opposing parents: L.A→R.B, L.B→R.A, R.A→L.A, R.B→L.B. Contradictory order
+   demands must remain deterministic, bounded and explicit; zero crossings is
+   not an achievable requirement inferred from those demands.
+2. Three-parent chain: L.A→M.B, L.B→M.A, M.B→R.A, M.A→R.B. Preserve upstream
+   L.A<L.B, obtain M.B<M.A and R.A<R.B, and preserve root ranks L<M<R.
+   Compare relation-ID reversal and a uniform authored subtree translation.
+3. Two local two-node SCCs X/Y feeding a singleton Z in a later rank, with
+   external Source.A→Y1 and Source.B→X1. Order Y before X without splitting
+   either cycle or changing Z's rank/perimeter geometry.
+
+Also run the existing two-container witness with authored source/target port
+anchors. Use meaningful pure tests for heterogeneous-height prefix packing,
+undirected bundle preservation and complete source-ancestor influence. These
+tests supplement rather than replace actual projected Project cases.
+
+Record complete before/after Projects, dependency graph and SCC batches, source
+placement/edge hints, scores, before/after unit and block orders, final derived
+ports, geometry and warnings. Reconstruct preservation independently, including
+folded latent offsets, anchors, other views and eligible route resets. Check
+same-input equality, JSON/source roundtrips and ten full repeated calls. Report
+explicit refusals without treating a successful observer process as adoption.
+
+Preserve crossings, all contact types, raw and edge-deduplicated shared intervals,
+label/node/label/stroke conflicts, enclosure/penetration, route length and area.
+Boundary improvement may worsen internal tracing; cyclic influence can retain
+crossings; source centers may disagree with fixed or derived ports. Keep those
+counterexamples. Stop this one comparison after independent source/result review.
+Production command/API/history integration and native route tracing remain
+required before accepting an Arrange replacement. Windows only; Linux remains
+excluded, and the existing prerelease remains unchanged.