Make EVA reproducibility easier to run and verify

[kevinhyj]

The repository now provides a concise README and separate CPU result-recalculation, GPU inference, training, CLI and resource guides. Six reviewer software requirements are mapped to public commands and dated evidence, with missing historical resources and unreleased code/DOI work stated explicitly.

Milena completes normally with exit code 0 when inference, metrics and plots succeed. --strict-reference preserves exit code 2 for reference mismatch. The original inputs, weights, scoring protocol and numerical results are unchanged. CPU configuration checks also work without eagerly importing the fine-tuning GPU backend.

• Add Python 3.10/3.11 CI for regression tests, documentation/CFF/input checks, source-archive execution, package builds, wheel installation outside the checkout and installed CLIs. External resource access is reported separately.
• Recover and hash 15 files from three historical RNA model snapshots, plus 13 additional cached comparison-model assets. This does not imply public availability, full runtime recovery or new competitor inference.
• Add contribution guidance and issue/PR templates; preserve model architecture and existing directories.

Validation: GitHub CPU checks passed all steps on both Python versions, including 139 tests per version with no skips, 149 local links, CFF/input checks and independent wheel imports. GPU default and strict runs each scored all 135 Milena inputs on one A100 with batch size 1; predictions exactly matched the saved fresh vector, returning 0 and 2 respectively. After the import fix, synthetic pretraining, midtraining and fine-tuning each completed two steps and exact checkpoint round trips. README rendering and diagnostic plots were inspected.

The measured Spearman remains 0.8394237924835843 versus reference 0.8360456283218484. Formal release, code DOI, full comparison-method reproduction and unresolved paper resources remain separate work.