docs: initialize project wiki from production hardening session

12 pages covering architecture, findings, and metrics from the
rotatingMachine + machineGroupControl hardening work:

- Overview: node inventory, what works/doesn't, current scale
- Architecture: 3D pump curves, group optimization algorithm
- Findings: BEP-Gravitation proof (0.1% of optimum), NCog behavior,
  curve non-convexity, pump switching stability
- Metrics: test counts, power comparison table, performance numbers
- Knowledge graph: structured YAML with all data points and provenance
- Session log: 2026-04-07 production hardening
- Tools: query.py, search.sh, lint.sh

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

wiki/findings/curve-non-convexity.md

@@ -0,0 +1,34 @@
+---
+title: Pump Curve Non-Convexity
+created: 2026-04-07
+updated: 2026-04-07
+status: proven
+tags: [curves, interpolation, C5, non-convex]
+sources: [nodes/generalFunctions/datasets/assetData/curves/hidrostal-C5-D03R-SHN1.json]
+---
+
+# Pump Curve Non-Convexity from Sparse Data
+
+## Finding
+The C5-D03R-SHN1 pump's power curve is non-convex after spline interpolation. The marginal cost (dP/dQ) shows a spike-then-valley pattern:
+
+```
+C5 dP/dQ across flow range @ ΔP=2000 mbar:
+  6.4 m3/h → 1,316,610  (high)
+ 10.2 m3/h → 2,199,349  (spikes UP)
+ 17.7 m3/h → 1,114,700  (dropping)
+ 21.5 m3/h →   453,316  (valley — cheapest)
+ 29.0 m3/h → 1,048,375  (rising again)
+ 44.1 m3/h → 1,107,708  (high)
+```
+
+## Root Cause
+The C5 curve has only **5 raw data points** per pressure level. The monotonic cubic spline (Fritsch-Carlson) creates a smooth curve through all 5 points, but with such sparse data it introduces non-convex regions that don't match the physical convexity of a real pump.
+
+## Impact
+- The equal-marginal-cost theorem (KKT conditions) does not apply — it requires convexity
+- The BEP-Gravitation slope estimate at a single point can be misleading in non-convex regions
+- The marginal-cost refinement loop fixes this by using actual power evaluations instead of slope assumptions
+
+## Recommendation
+Add more data points (15-20 per pressure level) to the C5 curve. This would make the spline track the real convex physics more closely, eliminating the non-convex artifacts.