Synthetic pilots whose bias and wobble vary independently #230

Is your feature request related to a problem? Please describe.
generate_synthetic_pilots.py draws each pilot's steady bias and its wobble from the same deviation scale, so the two always rise and fall together. Pilots who sit steadily off target, or who are on target but oscillate, are never generated. ADR 0002 rejects mad only because it cannot tell those two apart, but with this generator that argument cannot be tested: mad only scores 0.76 against 0.78 for the chosen features, within noise.

Describe the solution you'd like
Draw bias scale and wobble scale independently for some synthetic pilots so both kinds exist, keeping the hidden tier as the pilot's main tier so --truth still works. Measure std + abs_bias against mad only on that cohort and update the ADR's Evidence section with the result.

Describe alternatives you've considered

• Keep the mad rejection as a structural argument only, as ADR 0002 does now.
• Wait for real pilots (Capture designated (target) values for altitude, heading, vertical speed and airspeed #213, Collect pilot metadata and join it to telemetry for clustering #214), which may not contain these cases among 10 to 20 pilots.

Additional context
Different from #224, which varies the tier per metric; this varies bias against wobble within a metric. Follow-up to #216 and #225.