Make break_apart_entity's memo return what the function returns (#58)

[adamjohnwright]

The memo re-reads the rows the function wrote, but for a Complex it reads the wrong columns — so the cached answer is a different kind of thing from the uncached one, and the caller cannot tell.

kind	uncached returns	memo read	correct?
EntitySet	its members	union of input_or_output_* — where the provenance writer puts members	✅
Complex	the combination uids (one per set-variant, each standing for the whole complex)	union of input_or_output_* — where the iterproduct writer puts components	❌

The caller feeds the result into a per-reaction itertools.product as alternatives, so on its second and every later appearance a Complex was modelled as "any one of my subunits":

CCNA + CDK2 -> CCNA:CDK2   emitted as three VRs whose outputs are single free proteins
  6af74b4209 -> R-HSA-157440   (CDK2 alone)
  e18b12b9e3 -> R-HSA-157444   (CCNA2 alone)
  3fda7d321c -> R-HSA-157446   (CCNA1 alone)

"CDK2 alone is a complete output of complex formation" — which the module docstring explicitly forbids. The fix branches on the label rather than rewriting the half that already worked.

Measured

Idempotence on R-HSA-69242: entities returning something different on re-decomposition went 11 of 48 → 0.

DeltaSignal A/B — both arms on DeltaSignal main, Release97, the ten evaluation pathways, experimental ground truth, differing only in the memo:

arm	scored	correct	accuracy	macro-F1
control	627	452	0.7209	0.664946
memo fixed	627	452	0.7209	0.664946

NEUTRAL — and verified to be a real comparison, not an accidental no-op. Only 3 of 43 network hashes are shared between arms, and the networks change substantially:

quantity	differing (of 847)
prediction	0
predicted_ui	221
gene_uuid_count	389
converged	22

Convergence improves slightly (non-converged 155 → 149) for +0.8% solver iterations.

That's the expected shape: the duplicated VRs were byte-identical copies computing the same value from the same nodes, so removing them changes magnitudes and iteration counts but not answers.

Why land it

Correctness, plus the downstream consequences — it is the root cause of the diagram_bridge blow-up (#61), of spurious MAX_VARIANTS truncation (which silently bundles components into one opaque node), and of catalyst-edge multiplication. Networks shrink accordingly: Mitotic G1 580 → 493 VRs, catalog-wide 53,782 → 39,848.

No flag. The default-OFF convention exists for modelling choices whose direction is unknown. This is a function disagreeing with its own memo, with zero prediction changes. Contrast LNG_SET_MEMBERS_OR — faithful, but cost 51 of 223 cases and correctly stays off.

Testing

The lockdown test extends TestCrossCallStability, which asserted stability for EntitySets only — the case the memo got right, which is why this survived. Confirmed to fail with the fix reverted.

Suite: 924 passed vs 923 on main, same 7 pre-existing failures (they validate against the stale checked-in v96 output/ tree). Validator 11/11 on all three benchmark-eligible pathways.

Recorded as specs/002-break-apart-memo. Closes #58.

🤖 Generated with Claude Code