Simulator evaluates parallel branches independently, over-reporting current by 100% or more #8

circuit3d/js/simulate.js:375 states the assumption in a comment:

// Each path is an independent parallel branch — evaluate separately.

Parallel branches are not independent when they share an element. The engine enumerates every path from battery+ to battery- (findAllPaths, :173-212), then for each path sums series resistance and forward voltages and computes I = (V - sum(Vf)) / sum(R) (:380-401). Kirchhoff's current law is never applied.

Worked example. One 220R feeding two parallel LEDs from 9V:

reported:  31.8 mA through EACH LED
           -> implies 63.6 mA through a resistor that can pass at most 31.8 mA at 9V
           -> implies 0.891 W in a 1/4 W part
correct:   31.8 mA through the resistor, 15.9 mA per LED, 0.223 W

The error scales with branch count: eight LEDs behind one resistor reports 254 mA total. The wrong number is formatted identically to a right one, so a learner has no way to detect it.

Should be

This needs a solver, not a patch. Two candidates, and this issue exists to record the decision:

A. SPICE in WebAssembly. ngspice compiled to WASM (release 47, 2026-08-11). Existing browser builds: tscircuit/ngspice, eelab-dev/EEcircuit, ngspiceX. Gives real DC operating point, transient and AC analysis, and real device models including a proper Shockley diode. Cost: a WASM payload and less control over error messages.

B. Modified Nodal Analysis in JavaScript. Roughly 400 lines: per-component stamping, matrix assembly, LU with partial pivoting, Newton-Raphson for the diodes. Full control over the error messages, which matters for a teaching tool, and no bundle cost. Less capable long term.

Either way, the existing node-graph extraction is kept. simulate.js:1-340 correctly resolves breadboard connectivity, rails and wire merging into nodes. That is the hard part and it already works. The change is to stop solving and start emitting a netlist.

Blocked on a decision. Write golden-file tests against current behavior first so the fix produces a visible diff.

Found in the 2026-08-16 audit (30-agent sweep, 481 findings). Every line reference was verified against main at the time of filing.