Curated input/output stoichiometry now emitted — needs regen + benchmark re-run

[adamjohnwright]

What changed

_resolve_vr_entities previously hardcoded every input/output edge to stoichiometry=1 ({n: 1 for n in ...}), discarding the curated coefficient. It now reads Reactome's curated coefficient from the input/output relationship via a new get_reaction_io_stoichiometry (same rel.stoichiometry property get_complex_components uses) and attaches it per node:

• each expanded EntitySet member inherits its entity's coefficient;
• coefficients sum if two annotated entities map to the same node;
• absent curation, the coefficient defaults to 1.

Phase-3 emission already consumed input_stoich/output_stoich, so no downstream change was needed. Unit test updated (test_no_duplicate_edges) to assert a 2× input coefficient survives; full non-Neo4j suite green (922 passed).

Why this needs follow-up

This is benchmark-affecting: emitted coefficients move from all-1 to curated values, and stoichiometry_weighted is DeltaSignal's default export aggregation. It could not be validated in the dev environment (no live Reactome Neo4j loaded).

Action items

• Regenerate the logic-network catalog against live Neo4j
• Re-run the MP-BioPath benchmark (curator + experimental) and compare macro-F1 vs the pre-change baseline
• Confirm no stoichiometry-related regressions on the tracked pathways (PIP3, WNT, ERBB2, TP53, Mitotic_G1)

Related latent items surfaced in the same sweep (lower priority)

• Diamond decomposition undercounts stoichiometry: shared visited set in _resolve_to_terminal_reactome_ids conflates cycle-guard with dedup (only affects the complex-component decomposition path, not these I/O edges).
• is_valid_uuid (len==64) misclassifies the 36-char uuid4 from the all-strings branch of get_broken_apart_ids (currently unreachable).
• Pathway-source flags (--pathway-id / --pathway-list / --top-level-pathways) are not mutually exclusive; precedence is silent.
• validate-against-mpbiopath.py uses on_bad_lines="warn", silently shrinking the accuracy denominator on malformed ground-truth rows.

🤖 Generated with Claude Code

[adamjohnwright]

Answered: does curated stoichiometry improve the stats? → No (neutral-to-slightly-negative), when wired into the solver.

Two things were needed to test this:

1. Generator now emits curated I/O stoichiometry (this issue's fix).
2. deltasignal solver had to actually use it — previously edge stoichiometry only fed the stoichiometry_weighted export/pathway-rollup, never the solve. Added DS_STOICH_WEIGHTS (off by default): when on, the coefficient weights the hill_log AND aggregation as a mass-action exponent (log_fold = Σ sₖ·log(vₖ/bl)). No-op when all coefficients are 1, so only a stoichiometry-carrying catalog diverges.

A/B on the 9 experimental pathways, regenerated with curated stoichiometry, same networks, only the flag differs (isolated throwaway API containers, shared :8080 untouched):

metric	flag OFF	flag ON	Δ
macro-F1	0.6569	0.6541	−0.0028
change-F1	0.8039	0.7976	−0.0063
balanced acc	0.6885	0.6857	−0.0028
accuracy	586/800	583/800	−3

Consistent small regression across every metric (propagator_missed 65→68). Same pattern as prior "more biology" experiments (general consumption, substrate depletion, cross-pathway joint all regressed).

Disposition: DS_STOICH_WEIGHTS stays off by default. The generator stoichiometry fix stays — it's a correctness win for the export/pathway-rollup view (the pathway-browser overlay), just not a solver/benchmark lever. Not re-running the full 92-pathway curator A/B unless we want to confirm the experimental result carries over.