2.9 KiB
Capped overlap-penalty experiment
P alone does not solve the arranged bipartite overlap. Source
3d7adff923dee5bed54dfdccc119dbee0c7c2773 changes only the overlap coefficient
from 0.12 to 3.0 (plus an experiment comment), retaining the per-segment cap 80.
Node placement, ports, target-join J logic, edge order and crossing cost remain
unchanged. This is a transient diagnostic source variant, not a proposed default.
Windows command cargo run --locked --offline -p oa-core --example auto_layout_probe completed exit 0 with unchanged inputs at
2026-09-25T01:04:57Z. Same frozen Rust1.98.1 gnullvm environment as J; Linux
excluded. Output SHA256
13111fe89e4e637d1af572a8a45b8bdd763e94b58a2ec5ec6ca83e40ffa209c7.
J's 185 passing tests and strict Clippy are not claimed for P; this run executes
the diagnostic observer and its analytic primitives only.
Original evidence archive, 916568 bytes, SHA256
6b819b01c8012aa72d423443f4dc04ae261f1ee9dc6907f0063c219ff173ad36,
contains receipts, complete Projects/geometry, independent contact audit,
comparison, renderer output and runner. Manifest binds
original bytes and post-run probe executable; every member was checked against
its local original.
Complete Projects, nodes, ports and route endpoints are exact versus J in all 37 variants. All nine layered candidates retain the previous quality vector; in particular bipartite sharing remains 187 (115 nonterminal), with nine contacts and zero proper crossings. Two original-scene vectors change; the other 35 do not:
| Original scene | Shared length J → P | Bends | Route length | Point contacts |
|---|---|---|---|---|
| Crossed bipartite | 872 → 730 | 4 → 5 | 2064 → 2076 | 9 → 10 |
| Reversed children in two containers | 54 → 6 | 4 → 5 | 1512 → 1524 | 2 → 3 |
The bipartite result restores the first experiment's original-scene vector; the extra bend/length tradeoff was predicted by the low overlap cost diagnosis. Both contact increases remain visible rather than being counted as a clean readability win. These results concern core geometry, not native route tracing.
The independent P audit confirms these results and all nine recorded conservation/roundtrip/repetition checks. Both additional point contacts are T contacts; no proper crossing change is hidden by the table.
The next L comparison adds a bounded number of deterministic parallel lane coordinates while preserving all existing obstacle coordinates and fixed endpoints. Use the same augmented geometry for fast/grid search. P+L and J+L will distinguish available alternatives from preference. The cap 80 may still reject a four-turn detour whose extra length/turn cost exceeds it; retain that outcome rather than retuning cap, ordering and spacing together. No weight or lane policy is accepted for production by this diagnostic run.