ARCADIA-Studio/harness/experiments/scc-placement-20260925/criteria.md

SCC placement: independent pre-result criteria

2026-09-25. Proposed experiment criteria for issue32, not production acceptance. Read connection-corridor-sharing-diagnosis-20260925.md and the previous placement/clearance findings. Research HEAD at inspection is ae9dad54b749201c988fc398dbbc580db43e746f. No new candidate source was reviewed for this record; author was asked to signal readiness. No compiler, router, engine or GUI execution.

Proceed with one bounded SCC-condensation placement experiment, but make the planning graph and its relationship to authored containment explicit. The observed native prototype interleaves four10-node sparse SCCs across16 ranks; moving to component-aware ordering directly tests that diagnosis. A compact perimeter is a falsifiable packing choice, not a consequence of SCC membership: strongly connected graphs are not generally simple cycles and need not have a crossing-free perimeter drawing.

Material design risks to settle before results

1. Do not confuse a hierarchy quotient with the actual connection graph. At a containment level, mapping endpoints onto child/subtree representatives can introduce a proxy cycle even when underlying leaf connections are acyclic. Conversely, running independent leaf SCCs inside each parent can miss a real cycle crossing two containers. Preserve authored parent relationships; never reparent semantic nodes merely to make an SCC into a box. Record the exact vertices and mapped edges used for each SCC calculation, including planning parent and representative mapping. Call quotient components local placement blocks when appropriate; do not claim they prove global semantic SCC coherence.
2. Perimeter order is a separate decision. A component with ten vertices has many possible circular orders. An alias order or DFS order can recreate long backward edges, crossings and shared outer corridors, even after rank stretching disappears. Freeze a deterministic order before outputs, keep the edge-ID permutation, and disclose which connection information determines internal order. Do not rotate/reverse individual cases after seeing results. Singleton, two-node component and self-loop handling must be explicit; no duplicate placement at a repeated perimeter corner.
3. Undirected means no authored direction. An undirected relation can inform adjacency/proximity, but must not silently acquire a semantic arrow to satisfy the SCC algorithm. If treated as two directed adjacency arcs, state that it intentionally merges connected regions into placement blocks and may hide the directed condensation structure. If excluded from directed rank/SCC formation, keep it in geometry/quality evaluation and describe that weaker influence. A deterministic geometric orientation is another possible policy, but must remain a placement-only preference. Do not repeat the previous workshop omission without reporting unsupported cases.
4. Local label gaps still have to cover actual routes. A gap computed for one rank boundary does not guarantee space for long edges, perimeter return paths or the final Rust label location. Keep the released renderer/router as the observation source. Do not persist computed bends/label overrides or change the cutoffs to improve this comparison. Estimated label bounds remain a planning approximation; actual text and association require later native review.

Preservation and projection gates

Use original visible placement identities and the actual projected forest, not semantic element IDs alone. Preserve complete authored containment, semantic endpoint/port identity, direction, names, relation data, filters, fold flags, entry order and other views. Freeze reset eligibility from the pre-operation projection; use the existing audited original-to-final import/reset/retained-route translation exactly once. A no-position-change result may still perform the declared route reset and must be reported accordingly.

For a visible collapsed node, retain its compact projected dimensions exactly and preserve stored expanded size. Treat it as opaque for packing. Its promoted visible edges can influence placement without turning promoted owners into actual semantic owners. Hidden descendants follow the importer's nearest visible collapsed ancestor translation; never use local compact coordinates as their expanded geometry or relayout them secretly. Ineligible/folded route metadata remains stored with the existing translation policy, ready for expansion. Unrelated filtered/out-of-projection placements remain governed by the declared import scope. The frozen folded workshop has no nonempty authored-route sentinel: preservation there alone cannot prove suspension/restoration of arbitrary hidden pins.

Expanded containers must contain their final projected children below the appropriate header and above any footer/port-content requirement. Use the core geometry minima for the applicable projection rather than universal64-unit assumptions for all kinds. Growing a container may change relative port tips, port spacing and automatic sides; recompute the actual final projection. Final-size constraint disagreement is a failure, not a reason to overwrite anchors. Check finite positions/sizes, no illegal shrink/overlap, complete projected IDs, and exact intended versus final node geometry.

Authored port anchors and dormant explicit attachments remain exact. Automatic port-side/slot changes are geometric outputs, not automatically conservation failures. Record every change and inspect the final mandatory leads. Repeated projection/packing must not cause side or size oscillation. A lead-incompatibility rejection is useful but necessary-only: final routing must still be checked for NoClearPath, warnings, unrelated node/header penetration and identity-specific collisions.

Falsifiers on the fixed comparison

Keep all five frozen realistic cases and all nine existing small fixtures. Report unsupported/failed cases separately from outputs; never silently substitute a flattened or unfolded version. Keep the same reset baseline and unchanged released routing, and preserve the old prototype and prior external placements as references. New fixture construction or parameter selection must not be driven by these outputs.

• Structural diagnosis: for sparse40, independently recover the four10-node actual SCC memberships and relate them to placement blocks. Report each component's bounding span and its condensation rank interval. A proxy implementation that still interleaves those flat components or spreads a component over unrelated global ranks does not test the proposed remedy correctly. Mixed's six-node SCC and four singletons are a second independent topology check.
• Dominant ambiguity: recompute pairwise interval multiplicity on the final drawing. The sparse y28 trunk contributed38880/40522 and mixed y28 contributed4006/4228. Track the same edge groups wherever they move, as well as the new global worst corridor. Removing occupancy at literal y28 while recreating comparable sharing on another boundary is not success. Report unique corridor length, distinct-edge multiplicity and raw pairwise sharing separately; do not reinterpret a distinct-port trunk as an intentional bus.
• Tradeoffs: report proper/T/bend/endpoint contacts, nonterminal sharing, every new/resolved label/node and label/stroke conflict, route length/bends, drawing bounds and warnings. A compact arrangement that reduces area but increases ambiguous contacts is a tradeoff, not a quality pass. A large area gain bought with unreadable label packing is also a failed hypothesis. No weighted total determines adoption.
• Compound capability: the original/folded workshop must receive an actual declared supported result or explicit refusal. For supported compound cases, independently inspect child/header containment, port/edge promotion and hidden metadata. Passing flat sparse/mixed does not establish this milestone's hierarchy objective. The earlier two-container crossed-child fixture is useful here: check whether block order and child order cooperate rather than simply aligning container centers.
• Port/lead compatibility: preserve fixed-port fixture anchors and original opposing-side meaning; do not solve the task by moving an authored port. Final absence of incompatible leads must be followed by the normal full-route warning/penetration gate. Reject a new hard routing failure even when aggregate sharing decreases.
• Repeatability: same input yields identical Project; ten repeated candidate calls should settle after the first accepted arrangement, with no cumulative size inflation, shifting origin, side oscillation or hidden-coordinate drift. Keep mixed's fixed relation-ID permutation and semantic correspondence. Identical pixels are not required for every graph symmetry, but new failed trace tasks or severe order sensitivity cannot be hidden by selecting favorable IDs.

All14 cases need complete accounting. The small fixtures are explanatory witnesses, not training examples whose success compensates for realistic failures. Bound component recursion, graph/work/span, and any reprojection iterations; unsupported topology and bound exhaustion must discard the private candidate without partial model changes.

What would justify the next step

A useful result would show that coherent components materially reduce the diagnosed long shared corridors on both sparse and mixed cases, without a new hard route/containment/identity failure, while producing explicit compound results. This would support further initial coordinate-assignment work. It would not yet establish readable engineer traceability or a generally optimal planner.

If actual SCC blocks become compact but long return paths still coincide, inspect internal cycle ordering and available local corridors from that fixed output before proposing another scalar router penalty. If compound quotient cycles or final-port/lead requirements prevent a faithful placement, the failure should revise the planning interface rather than force a proxy result into the Project. Do not increase iteration limits, choose a winning rotation, or add a second packing heuristic after results merely to make this corpus pass.

Before adoption, reuse the predeclared source-to-target and label-association tasks at a common readable scale with ordinary selection/pan/zoom recorded. Include fan-out, distinct ports, cycle return and promoted endpoints. Geometry improvements, even across every frozen case, do not replace that native evidence or production history/source round trips.