ARCADIA-Studio/harness/experiments/connection-layout-reference-20260925/adapter-review.md

Independent reference-adapter review

2026-09-25. Static review only; no engine/compiler/router/GUI execution or source edits. Parent owns JavaScript review and execution. Reviewed local experiment plan, Python supervisor, Rust importer diff and adjacent baseline/prototype observation contracts. Read the JavaScript output-check interface only to establish where importer preconditions are enforced.

Verdict: no blocking source defect found for the frozen five-case diagnostic run, subject to consuming structurally passing outputs and the explicit post-import checks below. This is readiness for experiment execution, not adapter fidelity, native traceability, or production acceptance. Compilation remains unverified here.

Reviewed SHA256:

• Plan .local/connection-layout-reference-plan-20260925.md: 41a420db7ee039fa424f60b61097bddc6eff787e2705004e1c88b19f8b97c371.
• Supervisor .local/run-elk-reference-20260925.py: 114d2fd4dbc9a55cd57c4b5175989686dc413863fc2a62fa4a71278792937066.
• Importer crates/oa-core/examples/auto_layout_probe.rs: 95666ce866111d0626e964483fc954276e258bd0ef61a4f96b9857a0ceab468e.
• Rehashed frozen released geometry: 9a31a676f23dea2f41daec39497f569d68231fa0fd66e5a1a476be97d5d61ac2, matching the parent's declared input.

Concrete checks and result-consumption requirements

1. Freeze pin improvement: the supervisor asserts manifest and bundle pins and records all other bound hashes, including geometry, before/after. It does not assert the expected geometry SHA above. Add that single expected-input assertion before execution; otherwise a different preexisting geometry file can be consistently recorded without matching the intended frozen dataset. Current inspected bytes do match, so this is a narrow provenance improvement, not an observed wrong input. No generalized dependency framework is needed.
2. Only import structurally passing output. Rust import_placement rejects unsupported Scenario, schema mismatch, excess nodes/edges, missing/extra/duplicate IDs, nonfinite/nonpositive rectangles and translated nonfinite positions. It deliberately has no independent minimum-size/header/child-containment checks and no parent field in PlacementPlan; those gates are supplied by JavaScript inspection. Parent must require the per-case structural result and success receipt before extracting {schema:1,nodes:[{id,position,size}]}. Do not rank a transferred failed structural result as an ordinary candidate. Core validation is not a substitute for those geometric checks. Reprojected Rust bounds should also be compared with intended transferred bounds to detect core size normalization.
3. Compare final physical ports, not only unchanged anchors. Importer does not import ELK port positions; its final projected ports are therefore the right raw material for comparing owner/semantic-port identity and final intended tip to ELK's output. Unchanged authored fraction and unchanged ELK relative pixel offset can still diverge after compound resizing. Plan correctly declares this a transfer limitation requiring assessment. A candidate with such divergence is not evidence of faithful fixed-port optimization, even if both separate geometries validate.
4. Use the matching baseline. The existing observe result's after is released Layout without the new reset. The comparable baseline is reset_only_control, which clears effective routes after that Layout. The old prototype and importer clear using pre-operation eligibility, while the control's legacy helper uses post-Layout routed !custom_suspended. For the frozen unchanged-visibility cases this may agree, but verify actual eligible edge sets and resulting route records before attributing differences solely to placement. Do not relabel the old after metrics as reset-only.
5. Conservation summary does not prove the route contract. conservation masks routes and node geometry for broad semantic comparisons; authored_routes_exact=false is expected after reset. Independently compare complete per-edge records: entry/order, attachments, exact eligible reset, and retained translated_routes result. Roundtrip and conservation booleans are emitted observations, not asserted by the CLI; consuming analysis must inspect them. Exit0 alone does not pass them.

Importer correctness within this scope

Eligibility is captured from the original projection before movement and requires actual owners equal projected endpoints with both visible. The importer clones the original, changes only selected-view node geometry, clears eligible pins/full labels, and applies the existing original-to-final translated_routes helper only to retained routes. Entries are updated in place; route order and endpoint attachments are preserved. Other views and semantic records remain untouched by construction.

The revised collapsed rule is appropriate: compare proposed size to the original compact projected size exactly, update position, preserve stored expanded size. Temporary unfolding is used only to recover core visual ancestry, including Allocation grouping; it is not returned or used to change visibility. Hidden nodes found under a visible collapsed ancestor receive that ancestor's translation, preserving their latent relative geometry. The ancestry walk is bounded and detects cycles/missing parents. This implementation does not establish a general policy for unrelated filtered descendants omitted even from the unfolded projection; the stated policy leaves unrelated omitted placements exact. Do not extrapolate beyond the frozen fixtures.

The importer cannot reverse semantic edges or accept replacement ports because neither exists in its plan schema. Its 256 stored-node /1024 visible-edge limits and Scenario refusal are explicit. Original-to-final positional deltas are calculated before retained-route translation; no intermediate Layout or independently validated Batch prefix is introduced. This remains an example-only geometry transfer, not a new production command or history implementation.

Supervisor and attribution

The supervisor creates a fresh output directory and exclusive log files; each engine case has a separate subprocess with a90-second timeout. Python's subprocess.run timeout kills/waits for its child and the loop retains an explicit timeout/error receipt before proceeding. The reviewed contract uses the bundled engine in that Node process; this is not a general descendant-process supervisor. Before/after bindings cover Node, adapter script, geometry, manifest, bundle, dependency receipt, plan and supervisor. JavaScript receipts separately hash raw case/graph/mapping/result/observation files. Parent should retain both receipt layers.

The supervisor can exit0 even when individual cases fail or time out; its final record explicitly preserves those outcomes. Require every expected case receipt, no timeout, success plus structural checks before treating that case as transferable. Failed/unsupported cases stay visible in the comparison. Timing includes process work and is correctly disclaimed against earlier Rust route-only timings.

The plan correctly limits the claim to whether engine placement transfers usefully through released routing and label placement. Raw ELK routes/labels are diagnostic, never secretly turned into authored constraints. Estimated label extents, fixed-position port drift, prototype folded refusal and absent native trace acceptance remain real limitations. A failed placement transfer cannot reject ELK's complete coupled drawing architecture; a scalar gain cannot establish an engineer's correct trace or label association. No request is made for new fixtures, option sweeps, production dependencies or additional architecture before this bounded run.