Whole-pad parallel stage solve (stage-parallel relaxation)

Every stage on a pad is solved at once against a predicted interference field, then the field is corrected in a handful of sweeps: `runStagesParallel` in `src/lib/stageParallel/stageParallelRun.ts` with the shadow kernel in `stageCoupling.ts` (σmin → σmin + Δσ plus the pressure-ratio rate partition `derate = p_net,ref / (p_net,ref + Δσ)`), field prediction/residual in `interferenceField.ts`, freeze/convergence policy in `waveScheduler.ts`, and cloud fan-out mapping in `padWaveBatch.ts` / `waveTasks.ts`. VERIFIED FIDELITY TIERS: stage coupling is either `reference` (shadow-free net-pressure anchor, single pass) or `self-consistent` (ω = 0.7 under-relaxed fixed point, ≤ 12 iterations, per-stage residual trail / derate record). The serial pump-order run (`runStagesSerial`) receives the same selected mode and is the fixed point of the iteration, so `equivalence.ts` asserts the parallel answer converges TO the serial answer and reports the verdict as a badge instead of assuming it. Timing is reported in solve units (one stage solve) so the ~10x speedup is a property of the dependency structure, not the machine. Live surface `/pad-parallel` with per-wave timings, equivalence badge, convergence records, and tier-aware `stageParallelCsv.ts` receipts stamped `stage_coupling_mode=<reference|self-consistent>`. WebGPU sub-second previews via `padPreviewGpu.ts` with CPU / WebGPU / Mixed compute-mode toggles.

Reduced-order shadow coupling: the per-stage Δσ enters the existing `solveGeometry` kernel as a closure-stress offset plus a rate derate, not as a re-solved multi-stage DDM system. Next ship: drive the relaxation with `runNonPlanar3DPipeline` DDM shadows per wave so the converged field carries out-of-plane geometry, and push wave scheduling onto the fleet queue so a 40-stage pad fans out across workers instead of browser workers.

No phase markers extracted from this row's prose yet.