9.1 KiB
Coherent cyclic-group placement result
Branch-specific issue32 experiment. SCC condensation and compact cyclic groups remove the dominant shared outside corridor, but this is not ready for production Arrange. The corrected transfer supports all 14 frozen cases, preserves the declared model/geometry contract and is stable across ten calls. Local label conflicts, folded-route backtracking, bipartite sharing and area costs remain. No native tracing, command/history integration or release acceptance is claimed.
Independent evidence: corrected source review, run02 quality review, run02 conservation review, and transfer and label-conflict diagnosis.
Tested source: 2426d286483ab0445663221cd6010a3dadb89a97. Run01 at
4dd764fa6cfc6920350383e523ba5a597412ddf4 is retained, including its rejected
two-container case. Read the predeclared plan, diagnosis
and design alternatives for why this structural experiment followed
the flat lead-clearance repair rejected on readability grounds. Production routing is unchanged.
What the structural witness establishes
Sparse 40 has four SCCs of ten nodes. Independent graph reconstruction confirms four coherent 720×1024 blocks in condensation ranks 0–3, replacing the old alias-cycle-break layout's interleaving across 16 ranks. The old prototype's y28 outside corridor contributed 38880 shared units; the new geometry has no horizontal shared interval. Total raw pairwise sharing falls 40522→32 compared with that old prototype, crossings 17→15, area 6020910→4001536. The remaining 32 units are one local vertical group. This improvement required no router penalty or gap sweep and does not establish general lane separation.
Against current Arrange plus eligible-route reset, sparse crossings 94→15, contacts 158→19, sharing 1990→32, label/node 14→0, label/stroke 31→3 and route length 39660→18784. Area grows about 5.2%; label/label remains 1. Against the original authored input, area and route length grow, so it is not a universal improvement over every control. All controls remain in the raw evidence.
Complete quality comparison
Each arrow compares the current Arrange-reset control with the SCC candidate. Labels are node / other-label / unrelated-stroke collision counts. Area is candidate/control. Sharing is raw segment-pair length, not unique physical ink.
| Case | Proper crossings | Shared length | Labels N/L/S | Area ratio |
|---|---|---|---|---|
| flat_chain_4 | 0→0 | 0→0 | 0/0/0→0/0/0 | 0.72× |
| chain_7_row_wrap | 0→0 | 0→0 | 0/0/0→0/0/0 | 0.53× |
| diamond | 0→0 | 36→36 | 0/0/0→0/0/0 | 0.91× |
| crossed_bipartite | 0→0 | 276→368 | 0/0/0→0/0/0 | 0.57× |
| nested_cross_boundary | 0→0 | 0→0 | 1/0/0→0/0/0 | 1.60× |
| authored_detour_and_label | 0→0 | 0→0 | 0/0/0→0/0/0 | 0.93× |
| fixed_ports | 0→0 | 0→0 | 0/0/0→0/0/0 | 0.94× |
| two_containers_reversed_children | 0→0 | 148→164 | 1/0/0→1/0/0 | 1.18× |
| long_ordinary_labels | 0→0 | 36→36 | 2/0/0→0/0/0 | 1.06× |
| workshop_sa_original | 6→6 | 1832→344 | 1/0/0→3/0/1 | 1.81× |
| workshop_sa_folded | 7→8 | 1194→726.5 | 0/0/0→3/0/2 | 1.35× |
| mixed_dispatch | 10→7 | 716.4→516.6 | 0/0/0→0/0/0 | 1.46× |
| mixed_dispatch_edge_ids_reversed | 9→6 | 512.4→462.6 | 0/0/0→0/0/0 | 1.40× |
| subsystems_40_sparse | 94→15 | 1990→32 | 14/1/31→0/1/3 | 1.05× |
All 14 successful cases report zero unrelated node penetration, zero unrelated node collision and no incompatible mandatory leads. Four folded-workshop PromotedEndpoints warnings remain; no NoClearPath is reported. Those checks do not establish arrowhead clearance, native glyph readability or usability.
Mixed's fixed relation-ID permutation leaves projected node/port geometry and the SCC plan exactly equal, while final routes differ: crossings 7 versus 6 and sharing 516.6 versus 462.6. The placement is stable under this permutation; the whole routed drawing is not. Mixed area grows despite shorter routes.
The folded raw shared total 726.5 includes a promoted route that retraces 6 units. Distinct-edge occupancy counts 708.5: the 18-unit difference is raw segment-pair multiplicity, independently reconciled rather than silently discarded. The remaining sparse label/label pair and three label/stroke pairs have no positive shared route interval; one pair has no route point contact at all. Thus label association needs its own treatment after structural placement. See the independent reviews and failure diagnosis in the evidence archive.
Correctness repair and preservation
Run01 rejected the two-container fixture because parents intended to retain
their old 308×446 effective dimensions instead shrank to 280×416 after children
moved inward. The importer preserved authored 224×94 sizes under an overly broad
identity shortcut. Source 2426d286 restores intended unchanged effective minima
only for noncollapsed nodes whose first final position is exact and whose
dimensions are smaller/equal. It rejects every other mismatch before making
all qualified writes simultaneously, then validates and requires one exact
second projection. There is no iteration or change to SCC ordering/gaps/packing.
The original13 successful candidate Projects, metrics and plans are exactly unchanged in run02. The new two-container result confirms both intended 308×446 parent minima and all source-derived planned positions. All 14 report exact same-input output, nine subsequent calls equal to first/previous, and exact JSON and textual-source roundtrips. Repeated Projects/source strings are not serialized, so independent audits distinguish emitted Rust checks from their own raw structural comparisons.
Independent audits reconstruct complete expected Projects with only declared positions, required sizes and eligible-route resets changed. Semantic records, authored anchors, route identities/order, other diagrams and hidden metadata remain exact. Three latent folded descendants move (-680,0) with their collapsed owner. Current fixtures do not exercise surviving dormant authored-route translations or State-specific headers; no such coverage is claimed. Exact intended plan nodes are not emitted, so final transfer equality remains a source gate supplemented by independent final enclosures and the reconstructed failure witness. No history/Undo/Redo test applies to these example-only imports.
Verification and reproduction
Windows Rust 1.98.1 gnullvm and LLVM-mingw 20260908, locked offline dependencies,
one build job. Source 2426d286 passes 8 example tests, strict example Clippy and an
optimized build (gates 24–26). Earlier source 4dd764fa passes 5 example tests,
Clippy and build (gates 21–23). Gate 20's needless-range-loop Clippy failure and
earlier packing tests are retained, with their exact source pins. Core code is
unchanged from the previously verified lead diagnostic; this is not a new
whole-workspace suite claim.
cargo test --locked --offline -p oa-core --example auto_layout_probe
cargo clippy --locked --offline -p oa-core --example auto_layout_probe -- -D warnings
cargo build --release --locked --offline -p oa-core --example auto_layout_probe
auto_layout_probe --scc-fixtures
auto_layout_probe --scc-placement PROJECT_JSON DIAGRAM_UUID
auto_layout_probe --route-corpus MANIFEST
Both runs capture the built executable before further compiler writes, verify
source/binary/input hashes, and complete eight bounded child processes. The
five-case released route-control output is byte-identical in both runs, SHA256
9a31a676f23dea2f41daec39497f569d68231fa0fd66e5a1a476be97d5d61ac2.
Output process success is separate from per-case support. No latency guarantee
is inferred from these short corpus runs. The planner work counter is weighted,
not an exact instruction or elapsed-time bound.
Decision and evidence
Retain coherent SCC blocks as the supported structural direction. Do not adopt this complete candidate yet. Next work should target the concrete local label association, folded retrace and cross-container ordering witnesses. Preserve model/port/fold contracts, fixed inputs and the full quality vector; do not tune global gaps, rerun perimeter candidates to select a score, or weaken projection equality. Production integration and native trace/overview workflows remain required once a candidate satisfies those constraints.
The evidence archive contains both runs, failed and successful gate receipts, frozen inputs, controls, executable capture hashes, scripts and independent source/result/conservation reviews. Binaries are not embedded. Final claim and archive review is versioned separately outside the immutable archive. Archived supervisor scripts retain machine-local paths and Windows environment assumptions; the repository-relative build and probe commands above are the portable entry points after obtaining the source and locked dependencies.
Archive evidence/scc-placement-2426d286.zip: 6041259 bytes, SHA256 fbf40882ed807a1b40db8b3b9b9c846b060db02cbcc60611c18f9620c1fd59c9, 168 individually verified members.
This checkpoint is local-only until confirmed Gitea publication and byte readback. Linux builds remain excluded by the user; the frozen prerelease is unchanged.