Add versioned GDS layout registry and embeddings

[shanto268]

Summary

• add the checksum-verified GDS layout registry, Python APIs, and MCP layout tools
• retain static-embedding-v0 as the backward-compatible 9,227-dimensional baseline
• add the accepted universal-geometry-v1 schema v1.1, a 1,024-dimensional GDS and layout-control encoder
• expose explicit v0/v1 selection, v1 control maps, nearest-neighbor search, and 96 by 96 v1 shape reconstruction
• rebuild and execute Tutorial 14 as a paired, quantitative v0/v1 study

The earlier 512-dimensional v1.0 candidate was rejected and is not being released. It blurred topology, compressed random-pair cosine similarities, and did not consistently outperform v0. Schema v1.1 replaces it with a benchmark-gated design.

Embedding datasets

static-embedding-v0

• 24,106 vectors across all four supported component families
• 9,227 dimensions: parameter sum, ten moments, and a signed 96 by 96 bitmap
• remains the default so existing users are not broken

universal-geometry-v1 schema v1.1

• 20,062 GeneralizedCapNInterdigital vectors in the first release
• 1,024 dimensions: 32 geometry metrics, 768 full-spectrum shape descriptors, and 224 named-control channels
• shape block uses target-blind variance-selected DCT coefficients from signed functional-material and signed-distance 96 by 96 rasters
• named controls are centered and scaled before stable hashing; the published 40-row control map makes every channel auditable
• simulation targets are never embedding inputs and never select the representation
• all vectors are finite and unit normalized
• two complete generations produced byte-identical parquet, schema, and control-map files

Acceptance evidence

The release benchmark uses the same target-blind 2,000-layout sample and 300 nearest-neighbor queries for both standards. At seed 1401:

Gate	v0	v1.1	Requirement	Result
exact finger topology	0.9600	0.9800	v1 >= v0	pass
normalized parameter distance	2.4022	1.8621	v1 <= 90% of v0	pass
signed-bitmap shape distance	0.2816	0.2993	v1 <= 110% of v0	pass
mutual-capacitance neighbor MAE (fF)	0.7701	0.7047	v1 <= v0	pass
random-pair cosine std. dev.	0.3509	0.2488	v1 >= 60% of v0	pass

All five gates also pass independently for seeds 1401 through 1405. This demonstrates improved topology, control-space locality, and held-out physics locality without the similarity collapse seen in the rejected v1.0 candidate.

Tutorial 14

• nine executed code cells with no errors
• seven saved interactive Plotly visualizations
• paired latent-space exploration over the complete 20,062-design generalized NCap sweep
• dropdown coloring by every numeric geometry parameter and capacitance output
• v0 bitmap and v1 signed-material/signed-distance reconstruction
• nearest-neighbor visual comparisons, cosine distributions, and the five acceptance gates
• v0/v1 Python, database, and MCP API walkthroughs

Validation

• uv run pre-commit run --all-files --show-diff-on-failure
• MPLBACKEND=Agg QISKIT_METAL_HEADLESS=1 uv run pytest -q -m "not integration" --tb=short
• 350 tests passed, 4 integration tests deselected
• all GitHub test, package, documentation, MCP, and integration jobs pass; the pre-commit formatter fix is included in 1aa0aea
• public Hugging Face PR smoke test: 20,062 rows, 1,024 dimensions, 19 columns, 40 control-map entries, nearest-neighbor retrieval, and two finite 96 by 96 reconstructed rasters
• maximum public-vector unit-norm error: 1.2e-7

Release integrity

The accepted artifacts are published for review in SQuADDS_Layout_Embeddings PR #2.

• embedding parquet: 80abb1747aaf53141098c94ed9d2c75f61281dda77f021bcbb7733964d73ab72
• schema: dcb90a925fdcd96d7760e098b2223f5745c0d6d9cda5d7be2f7bcf2ad3f819a3
• control map: 5694744084ea25981c9064c97a8eb96082616d3ee53bb580fcbaa109635cff0d
• release manifest: 99dd5d3407a86543d7fbb46318283873c1c6c26cff99fe25674b3f5002cc6105

Files downloaded from refs/pr/2 match all four local SHA-256 hashes exactly.

Merge order

Merge the Hugging Face dataset PR first, smoke-test v1 from its main revision, and then merge this code PR. This avoids exposing an API version before its default remote dataset files exist.