3.3 KiB
P overlap-penalty experiment independent audit
Confirmed the reported finite result. Static read-only audit; no build, probe rerun, source edit or GUI action. P does not improve any of the nine layered candidate metric vectors and is not a demonstrated solution to their remaining shared routes.
Primary comparison: J06 08656b18bf8934cf264469cf68fc6dea97c15cbb versus P07 3d7adff923dee5bed54dfdccc119dbee0c7c2773. Inspected the latter commit: only routing.rs changes, overlap coefficient0.12→3.0 plus an experiment comment; cap80 remains unchanged. Both recorded cargo run --locked --offline -p oa-core --example auto_layout_probe receipts are Windows gnullvm/dev, exit0, inputs unchanged, failure null. Independently verified every log hash in both receipts.
Independent traversal of the raw outputs confirmed all37 variant Projects, projected nodes, projected ports and emitted route endpoint records match exactly. All nine layered candidates retain supported=true, repeat-input exactness, idempotence, passes2–10 equal first/previous, exact JSON/single-source roundtrips and identical conservation results. The authored-detour route-reset exception remains intentional. Baseline Command::Layout history/Undo checks remain identical; this is still not candidate-command Undo evidence.
Only two variant rows change numeric metrics:
| Original/before scene | Raw sharing J→P | Nonterminal sharing J→P | Bends | Length | Point contacts |
|---|---|---|---|---|---|
| crossed_bipartite | 872→730 | 800→658 | 4→5 | 2064→2076 | 9→10 |
| two_containers_reversed_children | 54→6 | 54→6 | 4→5 | 1512→1524 | 2→3 |
Each extra contact is classified as a T contact in the separately generated independent interval audit: bipartite T3→4 and two-container T1→2; proper crossings remain0. Lower shared length therefore comes with one additional bend,12 length units and one additional point contact in each scene. Bipartite returns to the pre-J shared-length total; this does not establish trace readability. All baseline/reset rows, manual witness and nine layered scalar vectors are unchanged. Exact model/endpoint invariants do not imply identical polylines.
Evidence identities:
- J06 raw output
f9ac7c1320b1cf026259bfcaf13a39e968ae317463712d8dc793375d9a4a541f(also identical to J03 output). - P07 raw output
13111fe89e4e637d1af572a8a45b8bdd763e94b58a2ec5ec6ca83e40ffa209c7. - P07 result.json
6149122d05d77e02b463edb3dd359b06786bc4468c27d8cd625cbc1bc80c9894. - P comparison JSON
9f228fbe78e1888c5d93e87dddd8d2effc01bab723929a87423c61e23f5cad66. - P independent contact audit JSON
6f4add5c81f6d16c45efabbefd9e36c9beb637b6823993f0d51adc3e20485ad9. Its recorded definitions and counts were inspected; this reviewer did not execute that helper again.
Recommendation: archive P as a bounded negative result for the layered overlap objective, preserving the two ordinary-scene tradeoffs. A predefined bounded lane-candidate experiment can now test whether available route geometry, rather than this single coefficient, limits improvement. Retain the same Projects/endpoints and all quality metrics; do not tune indefinitely against these nine fixtures or claim general quality/native acceptance. The cap, sequential edge order, fast-path preference, axis-only grid state and estimated label bounds remain limits.