4.8 KiB
L results and four-way routing experiment audit
Recommendation: stop tuning these fixtures. L is a useful experimental witness that route alternatives affect sharing, but its small layered benefit and material runtime signal do not presently justify its per-leg preprocessing as a production default. Keep J as the simpler independently motivated scoring correction candidate; decide L only against frozen independent scenes and route-only timing. No production/native acceptance claimed.
Verified final L result
Primary J06 pin 08656b18bf8934cf264469cf68fc6dea97c15cbb; L11 pin e8ec6dbb3160f62476a5557f14605f68959f65fe. Both recorded offline dev/gnullvm probe commands exit0, source unchanged, failure null. Verified receipt log hashes. Separate raw-output traversal confirms all37 Projects, projected nodes/ports and per-edge endpoint records exactly equal. All nine candidate supported/repeat-input/idempotence/ten-call/JSON-and-single-source-roundtrip results remain true, with unchanged conservation and baseline history checks. Candidate Command/Undo integration remains absent.
J→L changed scalar rows only:
- Original bipartite raw sharing872→724, nonterminal800→652.
- Layered bipartite raw187→176, nonterminal115→104; drawing area175960→165360.
- Original two-container raw/nonterminal54→12.
- Two-container baseline and reset raw/nonterminal18→12 each.
The other32 rows are unchanged in recorded scalar quality metrics. No changed point-contact, proper-crossing, length, bend, warning or collision counts in J→L. Independently comparing full candidate metric objects confirms all nine L layered metrics exactly equal PL, not merely the headline totals. Remaining layered bipartite104 nonterminal shared units still prohibit a solved-traceability claim.
Test10 receipts record135+7+13+18+2+13=188 passed,0 failed/ignored, exit0, unchanged source. J02 and L10 command argv and entire recorded environment dictionaries compare equal. Their routing_geometry test summary elapsed4.30s versus16.43s (~3.82 times, +12.13s); parent reports no concurrent agent compiler. This is a material cost signal, not a controlled routing benchmark: suite runtime includes varied tests and scheduling effects, with no repeated route-only timing. It must not be dismissed by quoting cargo compilation time or presented as an established scaling law.
Verified37 L schematic image hashes/lengths and manifest input hash. Native/glyph rendering was not exercised. Strict Clippy12 was active when assigned and is outside this audit; no outcome inferred.
Four-way conclusion
All numbers below are raw pairwise shared length; inspect the preceding audits for contacts/area/turn tradeoffs.
| Scene/variant | J | P | P+L | L |
|---|---|---|---|---|
| Layered bipartite | 187 | 187 | 176 | 176 |
| Original bipartite | 872 | 730 | 724 | 724 |
| Baseline/reset bipartite | 276 | 276 | 72 | 276 |
| Original two-container | 54 | 6 | 12 | 12 |
| Baseline/reset two-container | 18 | 18 | 12 | 12 |
P has no layered benefit. P+L's baseline bipartite72 means only common terminal sharing remains, but proper crossings rise0→1, length+24 and area+19392; L alone does not produce that tradeoff. The strongest ordinary two-container sharing result is P6, at an extra bend,12 length units and one T contact. Neither the coefficient nor lanes dominates all vectors. J diamond's isolated improvement remains valid but does not erase its original bipartite regression.
Next scope should be finite and specified before inspecting outcomes: one realistic nested/folded scene, one mixed fan-in/out with a fixed edge-ID permutation, and one modest sparse scene measured by route-only timing and warning/budget counts. Preserve endpoint/model constraints and evaluate traceability, collisions, bends/length/area jointly. Do not silently extend unsupported layered folded behavior into an acceptance claim. Set practical timing and readability criteria before repeating improvement. If L does not show proportionate independent benefit, omit it; avoid adding configuration/architecture solely to retain every experiment. The explicit Layout reset policy and ordinary input/Undo/Save lifecycle are still separate implementation work.
Evidence identities
J06 output f9ac7c1320b1cf026259bfcaf13a39e968ae317463712d8dc793375d9a4a541f.
L11 output 64232f580600b7a874a6f4ceb128a6188fd5ff72655bc6b93d824ebed27ffb7d.
L11 result.json feee84c2fde115c22ea865033ff2d136da2aace1f1c8a7423c8add11358366a8.
L comparison JSON e1e23e6c2a50cd384457bb061d5b70a740f586f42767164fad9703ecbb4096a8.
L contact audit JSON 6f6cc674c27d3f43001fe1751ef3b09332caa9ddca396759f5525f811cbe1aaf.
Read-only audit of completed records; no build/probe/helper rerun, source edit or publication. Earlier independent J/P/PL source/results notes retain grid-axis, subdivision-score, capped retained coordinates versus unbounded preprocessing, stable-edge-order and approximate-label limitations.