Productionize SynthMarket v0.2 research foundation

[RajvardhanPatil07]

What changed

This PR upgrades SynthMarket from a local-first WGAN-GP project into a stronger reproducible research and deployment foundation while preserving the existing WGAN-GP pipeline as the baseline.

Research / ML foundation

• Added pluggable SyntheticModel protocol and model registry.
• Added reproducible dataset manifests with SHA-256 fingerprints.
• Added leakage-safe chronological train/validation/test splitting.
• Added tail-risk metrics: VaR, expected shortfall, maximum drawdown, and extreme-loss rates.
• Added volatility-regime distribution and transition diagnostics.
• Added synthetic-to-real utility comparison primitives.
• Added repeatable benchmark result aggregation.
• Added experiment manifests with dataset identity, seed, Python/platform metadata, Git SHA, and configuration hashes.

Production / engineering

• Hardened packaging metadata and dependency ranges.
• Added optional FastAPI service layer with /health and /models.
• Added Dockerfile and Compose deployment scaffolding.
• Added CI across Python 3.10/3.11/3.12 with lint, tests, compile checks, and dependency audit.
• Added regression tests for the new foundation.
• Added model card, data card, production protocol, licensing guidance, and updated README.

Why

The project now has the seams required to add TimeGAN, diffusion, Transformer, and statistical baseline backends and compare them using one reproducible evaluation protocol, instead of replacing the current WGAN-GP implementation without evidence.

Validation

The new test and CI configuration is included in the branch. The current environment does not provide gh or outbound GitHub clone access, so I could not run the repository's full test suite locally or execute a local push; the GitHub branch was created and populated through the repository API.

Follow-up

The next production layer should add durable experiment persistence (PostgreSQL), job queues, object storage, authentication/RBAC, observability, and the next-generation model backends (diffusion/TimeGAN/Transformer).