6.9 KiB
#32 residual routes after terminal-join candidate J
Read-only diagnosis of run03 command/output.log, .local/auto-layout-j-comparison-20260925.json, the geometry-only crossed-bipartite PNG, and current routing.rs / projection/port_layout.rs. No edits, builds or native GUI execution. J is an experimental cost correction, not an accepted readability improvement for every scene.
The remaining187 units are not coincident distinct ports
The four functional exchanges deliberately reuse four semantic ports:
| Flow | Source port | Target port |
|---|---|---|
0 (009) |
010 |
014 |
1 (00c) |
013 |
011 |
2 (00f) |
010 |
011 |
3 (012) |
013 |
014 |
Projected source ports are (264,106) and (264,286); target ports (346,106) and (346,286). Each port is a single real semantic identity serving two edges. There are no authored port anchors in this fixture. The route endpoints and ports were unchanged by J. Splitting a common port into several displayed ports or changing relation port IDs would change the contract/semantics and is not a routing remedy.
Of187 pairwise shared units,72 are four legitimate common-port18-unit terminal stubs. The remaining115 are four avoidable interior overlaps:
- Flow0/Flow2:
y=106, x=282..328, length46. - Flow1/Flow2:
y=106, x=305..328, length23. - Flow1/Flow3:
y=286, x=282..305, length23. - Flow0/Flow1:
y=106, x=305..328, length23. These routes have different endpoints, so this is not a common-port exemption.
These sum115. The two diagonals now have separate vertical channels (x=328 and x=305); horizontal tails still merge different routes. The geometry PNG confirms that a user sees shared horizontal lines with branch contacts, not four independent readable connections. It is not native/glyph evidence.
port_layout::layout chooses automatic sides from peer centers, semantic fallback and clear exits; explicit anchors override the resulting side/offset. It sorts multiple automatic ports on a side by peer-axis coordinate and ID, then allocates distinct positions around fixed anchors. Here each node has only one relevant port, so improving distinct-port separation would not address this fixture. Moving a common automatic port could alter geometry but still leaves all its incident edges sharing the same physical endpoint. Fixed anchors must remain exact.
J trades one objective against another
J improves layered diamond overlap82→36 (nonterminal0) and layered bipartite344→187 (nonterminal115). But the original unarranged bipartite regresses730→872 shared units (nonterminal658→800), while bends5→4 and length2076→2064. The current original routes use a320-unit top horizontal trunk and a160-unit second-diagonal tail; later straight edges overlap those trunks. This is consistent with a length/bend-biased objective, not evidence that J universally improves routing.
crossing_cost charges only0.12 per shared unit, capped80 per prior segment; a turn costs24. J now compares more complete turn costs, so it can rationally prefer a cheaper route with more overlap. The router evaluates edges in stable-ID order and does not jointly revise earlier paths. Terminal fairness alone cannot correct an objective that undervalues ambiguous shared strokes. Public/ordinary scenes must remain in the comparison; do not judge only Arrange-generated geometry.
Next bounded experiment: penalty versus available lanes
Recommend a tiny factorial comparison with node positions, ports, anchors and edge order frozen: J; J+P; J+L; J+P+L. Keep the original unarranged bipartite as a named regression sentinel alongside layered diamond/bipartite and the existing corpus. This separates missing candidate geometry from cost preference; no placement change initially.
P: use one predeclared experimental overlap coefficient, e.g.3.0 instead of0.12, retaining the80 per-segment cap and proper-crossing24 initially. This is a falsifier, not a proposed final default. A12-unit dogleg around a straight shared span can cost24 additional length plus96 for four full-path turns. The cheap existing coefficient cannot compensate for that; even3.0 with the80 cap may still reject it. Record this failure explicitly rather than repeatedly tuning. A second experiment removing/changing the cap would require separate justification; the cap and segment subdivision make the objective unlike a simple global shared-length weight.
L: offer a bounded number of parallel candidate coordinates at CLEARANCE=12 on either side of occupied horizontal/vertical segments, subject to current obstacle clearance and fixed terminal stubs. These are derived path candidates, never authored waypoints. Put them in fast-path candidate space and, if needed, the grid space consistently, with a deterministic global cap. The current obstacle-boundary/midpoint coordinate set does not systematically offer these close parallel lanes. More space between boxes alone cannot supply the missing candidate coordinate or change its cost preference.
A smaller staging of this comparison is to run P first and inspect actual chosen alternatives, then add L only if the recorded candidates lack the desired lane. Do not silently combine penalty, cap, edge ordering and node spacing changes. If P merely creates crossings/long exterior detours, that disproves it as a standalone improvement. If L offers legal lanes but they are never selected, that isolates the objective. The result may justify a different explicit priority or bounded shared-segment nudge, but not before this evidence.
Alignment control: an additional fixed24-unit inter-layer gap or stagger is lower priority. It changes the diagram and automatic peer positions, complicates attribution, and can lengthen shared trunks. Alignment may be useful for a particular placement heuristic, but it is not a direct explanation of identical horizontal tails between the same common ports. Port-placement changes are outside the identified dependency scope.
Acceptance constraints for the experiment
Keep exact semantic graph, port identities, attachments, authored anchors and generated-versus-authored route distinction. Report raw overlap, legitimate terminal overlap, nonterminal overlap, proper crossings AND branch/bend contacts, labels, obstacles, length, turns, warnings and workload. Current proper-crossing zero excludes contacts at bends; trading coincident strokes for distinct, obvious crossings can improve traceability even if that one metric rises. Conversely, tiny offsets that appear merged at normal zoom are not success.
Use shared-port fan-out/fan-in and distinct-port controls, fixed-anchor controls, narrow corridor, authored-pin and opposite-direction routes. Compare original, reset-only and layered scenes, plus a fixed edge-ID permutation to expose priority sensitivity. Parent should inspect actual route tracing at normal zoom before adopting any weight/lanes policy. No quality acceptance or port-placement defect is asserted here.