Add frozen, reproducible KEGG reaction benchmark (Phase 1)

[janisshin]

Summary

This PR establishes Phase 1 of the AAAIM reaction-annotation benchmark: a frozen, checksum-verified dataset for evaluating KEGG reaction annotation under imperfect metabolite identifiers.

What this adds

• A canonical 75-model BioModels manifest and reproducible downloader
• SHA-256 verification of every downloaded SBML model against BioModels
• Deterministic benchmark-generation scripts and versioned output tables
• Reaction ground truth that preserves every valid KEGG identifier per reaction
• Explicit separation of scientific exclusions from pipeline failures
• Diagnostics for malformed, misplaced, and unparsed KEGG annotations
• Near-duplicate model clustering to prevent train/test leakage
• Regression tests covering extraction, exclusions, clustering, and determinism
• Documentation reconciling the observed benchmark with the historical reported counts

This branch also integrates the reaction-evaluation changes from upstream/test/no-rule-evals, including support for slash-, colon-, and URN-form KEGG identifiers, KEGG compound annotations, exchange/SSX detection, and ontology/cofactor ablation controls.

Frozen benchmark snapshot

• 75 manifest accessions downloaded, parsed, and checksum-verified
• 74 models included
• 1 model scientifically excluded
• 0 pipeline failures
• 5,838 ground-truth reaction records
• 5,816 evaluable reactions
• 19 exchange/SSX exclusions
• 3 malformed-identifier exclusions
• 91 reactions with multiple valid KEGG identifiers
• 1,906 distinct KEGG reaction identifiers
• 54 model clusters, including 9 multi-member clusters

All 10 dataset invariants pass. Two consecutive builds produced byte-identical output across all frozen artifacts.

Historical-count reconciliation

The previously reported benchmark size of 68 models and 4,379 reactions could not be reproduced from the canonical 75-model manifest under a defensible exclusion rule. This PR documents and investigates the discrepancy rather than adjusting exclusions to force agreement.

The evidence indicates that the historical evaluation likely used a different or truncated model list. The current benchmark therefore freezes the fully reconciled observed dataset:

• 75 models = 74 included + 1 scientific exclusion + 0 pipeline failures
• 5,838 reactions = 5,816 evaluable + 22 reaction-level exclusions

Validation

python benchmark/scripts/download_biomodels.py
python benchmark/scripts/build_benchmark.py
python -m pytest tests/test_benchmark_build.py

The Phase 1 regression suite contains 21 tests, and all tests pass.

The frozen reactions.csv SHA-256 is:

f11ebc6f4d0734deb14c054caf277b81695b22668017d70af9c43cd6146004f1