Variant-node sharing is now unconditional (nodes -34.6%, accuracy neutral)

[adamjohnwright]

Phase 1 keyed entity uuids on the virtual reaction, so a participant common to every variant of a Reactome reaction was photocopied once per set-member combination — HDR's BCDX2 complex existed as 33 identical nodes. Keying on the Reactome reaction is what the data says. The flag existed only to measure that the correction costs nothing.

It does not. Pre-registered in deltasignal specs/020-variant-node-sharing before the regeneration; all four predictions hold.

Size (P1)

	before	after
nodes	108,229	70,738	−34.6%
edges	285,497	224,189	−21.5%
cyclic components	210	204
nodes in cycles	8,407	8,131
largest component	675	675	unchanged

Honest caveats, all recorded in the spec: sharing is a literal no-op in 47 of 92 pathways, the median reduction among the 45 that move is only 6.7%, and the catalog figure is carried by a few (Fanconi Anaemia −83.5%, DNA Repair −61.8%). About a third of catalog nodes are dissociation sinks with no out-edges, so the −21.5% edge figure is the more conservative read.

The new-cycle risk was checked, not assumed. Merging two variant strands could in principle weld a cycle neither had, which would undo the largest accuracy win this project has (specs/018 unwelding, held-out +173). Measured across all 92 pathways: 0 gained cyclic nodes; 4 lost a component.

Accuracy (P2–P4)

29 of 24,100 predictions change — 0.12%, in 3 pathways (DNA Repair 14, RUNX2 10, MET 5). Only DNA Repair has a nonzero net; the other two fixed exactly as many as they broke.

	before	after
curator	20414/24100 = 84.71%, mF1 0.8137	20400/24100 = 84.65%, mF1 0.8125
experimental	611/849 = 71.97%, mF1 0.6451	609/849 = 71.73%, mF1 0.6432
false change	1,150	1,150

Under perturbation-set conditioning, held-out net is 0 with zero discordant cases on both axes. Sharing changes which nodes a gene resolves to, so 4,094 of 24,100 cases are not like-for-like and are dropped; among the rest nothing moves at all.

The net −14 is in a tuning pathway and is not quoted as an effect (P5).

Verification that the default reproduces what was measured

Regenerated two pathways with no flag set and compared against the flagged catalog:

pathway	flagged run	new default	no-sharing control
HDR	799 nodes / 3,123 edges	799 / 3,123	1,719 / 3,648
DNA Repair	4,756 / 14,487	4,756 / 14,487	12,442 / 24,332

The control catalog and this one were built from byte-identical generator trees (13847fa is the merge commit whose second parent is 79feca7; git diff between them is empty), so the generator is not a confound.

Why no flag

Per the flag-expiry policy: a flag whose question is answered is a second code path nobody measures. Setting LNG_SHARE_VARIANT_NODES now raises rather than being ignored, so a stale value in a benchmark script cannot silently measure something else.

Also in this PR

LNG_BOUNDARY_EXPANSION, LNG_COMPOSITION_EDGES and LNG_EMIT_ONE_SIDED are added to _FINGERPRINTED_ENV. All three change emitted content, and that list's own comment already promises that flipping a fingerprinted flag invalidates stale caches "instead of silently reusing a network built under the old setting" — which is exactly what these three could do.

Docstring fix: the BCDX2 example spans 6 reactions, not 5.

Not fixed here

Adversarial review found that ~60% of nodes with in-edges are exact structural twins (identical in-edge and out-edge sets). Sharing removes 38% of them but leaves the rate untouched, because it does not merge reaction nodes. Filed in specs/020 as a separate finding with an explicit warning: the two nearest interventions already measured are negative and load-bearing (specs/012, −15 and −61 held-out), so this needs its own pre-registration, not a patch.

tests/test_variant_node_sharing.py  11 passed
full suite                          1022 passed, 1 pre-existing MHC failure
ruff check src/ bin/                clean
mypy src/                           clean

🤖 Generated with Claude Code