Variant complex nodes, on by default (deltasignal specs/046)

[adamjohnwright]

Variant complex nodes (deltasignal specs/046), on by default. LNG_VARIANT_NODES=0 restores the previous networks.

What it does

• A complex that contains sets becomes one node per member combination. "CDK4,CDK6:CCND" (2 × 3 members) becomes 6 nodes with variant keys parent::variant::slot=member_... (src/variant_keys.py).
• Each reaction is copied once per consistent choice across ALL its participants, catalysts and regulators included (decisions D2, D4). The copy that chose CDK4 + CCND1 produces exactly the CDK4:CCND1 node, so a downstream consumer joins it by key (src/variant_emission.py).
• An output-only set is bound to the input set whose members correspond by isoform-level reference identity, so it is not multiplied.
• Past LNG_MAX_VARIANTS (decision D6):
  1. pool the participants whose choice reaches no output (<stId>::pool, fed by variant_pool OR edges, which the solver combines by mean, decision D5);
  2. else one copy per output variant;
  3. else one copy whose plain outputs feed their variants (variant_split), restricted to unproduced nodes consumed by a curated following reaction.
• Depleters: several variants of one curated catalyst deplete a shared target once, through a pool.
• Regulators: a catalyst or regulator reads the copy of its entity that a curated preceding reaction produces.
• Every export (nodes.csv, node_resolution, containment, drugs, boundary layer) understands variant keys and pool ids. A pool id is never looked up in Neo4j and never resolves as a readout or gene target.

Measured

Arm vn7, build 20261009-2042_1daf333_vn7, against canonical 20261004-1959_6990015, same solver:

	net	McNemar p
curator held-out	+130 (233/103)	1e-12
curator tuning	−18	0.43
experimental	−10 (22/32)	0.22

• Coverage is identical: 23,511 / 845 valid cases, 864 / 122 perturbations.
• Cut hand-offs go 2,504 → 0; required inputs beyond the curated ones 13,048 → 0; missed joins 0.
• The full record, including two failed arms (vn4, vn5) and every mechanism traced to a line, is in deltasignal specs/046-variant-complex-nodes/research.md.

Review

An adversarial review of the diff confirmed flag-off output is byte-identical to main on 4 regenerated pathways. It found 4 real defects, all fixed in 1daf333, each with a test that fails without it. One design caveat is recorded, not changed: a pool equals expansion exactly only for a single varying participant under a pure product.

Tests

• 388 pass offline. The 2 failures are environmental: no output/ dir, and Neo4j unreachable by design.
• The boundary-layer unit tests pin LNG_VARIANT_NODES=0, the path they were written for. The variant path is covered by tests/test_variant_keys.py, test_variant_emission.py and test_variant_pools.py.

🤖 Generated with Claude Code