Synthetic pilots strong on one metric and weak on another #224

Is your feature request related to a problem? Please describe.
#216 asked for synthetic pilots with "variation per metric (for example good altitude, poor heading)". generate_synthetic_pilots.py scales each pilot's deviation per metric by 0.7 to 1.4 around its tier, and the tier scales are 1, 2.5 and 6. The per-metric ranges (0.7–1.4, 1.75–3.5, 4.2–8.4) never overlap, so every metric of a pilot stays inside its tier. No synthetic pilot is tier 1 on altitude and tier 3 on heading, so performance groups are never tested on the mixed pilots real cohorts will contain.

Describe the solution you'd like
Give some synthetic pilots one or two flight dynamics factors from a different tier, for example a --mixed-share fraction of pilots with one metric drawn from another tier's scale. Then report what that does to tier recovery.

• Keep the hidden tier as the pilot's main tier, so --truth still works.
• Measure the mean adjusted Rand index over seeds 0–5 with mixed pilots, against the current 0.906.
• If recovery drops, decide whether the k = 3 test should keep 0.8 on the current cohort and add a separate, looser test for mixed cohorts.

Describe alternatives you've considered

• Widening METRIC_VARIATION until the tiers overlap. Simpler, but it blurs every pilot rather than producing pilots with a clearly weak metric.
• Leaving the generator as is. The tests then only cover pilots who are uniformly good or uniformly poor.

Additional context
Follow-up to #216. The clustering features and their log scale are in docs/adr/0002-cluster-on-std-and-absolute-bias.md; this issue only changes the synthetic data, not clustering.