Symptom

The Calculation editor asked users to pick a “target data point” and stored formula variables inside the sources list — the same list the universal-source-wiring refactor (Phases 2/3) redefined as “what invokes this node”. After the Phase 3 rollout the Calculation editor was broken: formula variable selection no longer worked, and the UX incorrectly conflated invocation wiring with formula-variable wiring. Tracked as the D7 “fix before release” follow-up of openspec/changes/universal-source-wiring-ux.

Root cause

CalculationEngineNodeMetaData had no dedicated field for formula-variable node references. The editor re-used sources for that purpose, and FormulaResolver read from sources at eval time. When sources was repurposed to mean invocation wiring, the formula-variable semantics silently broke. Additionally, writing the computed result into a separate “target” DataPoint created an unnecessary coupling and a confusing UX.

Fix

• Added inputs: List<NodeIdentity> (formula variables, read at eval time) and derived: Snapshot? (computed output snapshot) to CalculationEngineNodeMetaData in krill-sdk (repo: krill-oss). The SDK change is consumed via a guarded, opt-in (KRILL_SDK_LOCAL=true) Gradle composite build.
• FormulaResolver now resolves formula variables from inputs instead of sources.
• ServerCalculationProcessor computes the formula, persists the result in derived, then calls executeSources to wake downstream nodes — the “target” concept is gone.
• ServerDataPointProcessor.onSourceTrigger (net-new) pulls a Calc source’s derived snapshot when the Calc node is listed as a source.
• The Calculation editor (EditCalc) lost the target-picker row and gained a chip rail
  • modal DOUBLE-only DataPoint picker for inputs.

Prevention

Keep formula-variable wiring (inputs) and invocation wiring (sources) structurally separate — never overload one list for both. New TargetingNodeMetaData value-producing nodes that downstream nodes pull from should expose a derived-style snapshot field, not write into a named target node. The CalculationMetaDataSerializationTest, FormulaResolverInputsTest, ServerCalculationProcessorTest, and DataPointPullsCalcTest suites guard these invariants; any new eval-output node type should follow the same inputs/derived pattern and add equivalent tests.