Expand EntitySets to member species (no bare set nodes)

[adamjohnwright]

Summary

Sets should always decompose to their member species — they should not survive as standalone network nodes (except deliberate exceptions like ubiquitin). But an EntitySet could remain a bare, producer-less node when one reaction referenced the set directly (e.g. a phosphatase substrate) while the producing reaction emitted a specific member. The two are the same curated entity, left unlinked — stranding the set, which is frequently the curator's readout.

Root cause: the emission mapper intersected the set's get_terminal_components — which collapses modified species down to their base reference protein (p-ERK R-HSA-109844 → ERK R-HSA-109842) — with the VR's resolved members, which keep the modified form. Empty intersection → bare-set fallback.

Fix: intersect against a new _matching_leaves helper — a pure read mirroring break_apart_entity's decomposition depth (Complexes atomic unless they contain a set; sets expand), so modified species are preserved and the set expands to its members. LNG_SET_EXPAND env toggle (default on) keeps the old behavior for A/B.

Effect

RAF/MAP kinase cascade: 12 → 2 non-ubiquitin set nodes; expected-change perturbation→readout connectivity 54% → 77% (via real set-expansion, curated hasMember only — edges added are minimal, e.g. TP53 +22, ERBB2 −3, not a cartesian).

Benchmarked (mixed, shipped per design-owner decision):

• held-out curator (16,696 cases): 80.53% → 81.20% (+67)
• experimental (849): 68.20% → 66.43% — RAF +6 (target), TP53 −19 (12 are DAXX over-coupling, a known noisy co-repressor)

The companion deltasignal benchmark change maps set-typed curator readouts to their member species (since sets no longer have nodes).

Tests

TestMatchingLeavesGranularity + full non-DB suite (139) green, ruff + mypy clean, coverage 46%.

🤖 Generated with Claude Code

[adamjohnwright]

Adds a second generator-faithfulness fix on this branch: drop fully-identical duplicate edges before emission. Distinct virtual reactions of one curated reaction can canonicalize to the same target node; a catalyst attached to each then yields exact-duplicate rows that the solver double-weights in its activator aggregation. Final dedup keyed by (source, target, sign, type, reaction). Scale: 39/92 pathways; Pre-NOTCH 40,548→27,552 (−32%), FGFR2 −7,806, ERBB2 −3,345, RAF −2,191. Suite 139 green.

[adamjohnwright]

Final state of this PR (commit ee1fdad):

Shipping (default-on, both neutral-or-positive):

1. Set-expansion — EntitySets expand to member species at matching granularity (no bare producer-less set nodes). Curator 80.53→81.20%.
2. Duplicate-edge dedup — removes exact-duplicate catalyst/regulator edges that double-weighted the solver. Neutral (exp 66.43%, cur 81.18%); Pre-NOTCH 40,548→27,552 edges.

Behind a toggle (default-OFF, LNG_CATALYST_BUNDLE=1):
3. Catalyst/regulator bundling — the faithful complex-as-node representation, but it exposes a pre-existing UUID-silo defect (a complex produced by one reaction and catalysing/regulating another is stored as separate un-unified nodes, compounded by set-variant mismatch + non-deterministic set-member selection). Subunit-shatter accidentally masks this via spurious hub paths and scores ~1.2pp higher on experimental — luck, not correctness. Kept opt-in until the canonical entity-node unification is redesigned (documented ceiling; full-merge was −204pp).

Net PR effect at default: set-expansion + dedup only. Bundle inert unless toggled.