fix: reduce_factors NaN aggregation silently corrupted other observables

[jc-macdonald]

Summary

Closes #119. reduce_factors() aggregated per-factor importance across observables via np.maximum, which propagates NaN (np.maximum(4.05, nan) == nan). A single observable that's legitimately undefined in some regimes (e.g. a Type-I error rate, only meaningful when the null hypothesis is true -- a completely reasonable, common Scorer pattern) silently erased a real, significant importance value found via a different observable for the same factor, dropping it from the output with no warning or error.

Where this surfaced

yoavram-lab/pp-eigentest's Phase 1a Morris screen had a factor (alpha) with a real, large importance (mu*=4.05, far above threshold) on mae_seq, silently excluded from reduce_factors' output because a different, NaN-valued observable (type_i_rate_*, NaN by design outside null regimes) happened to appear later in the importance dict's iteration order. Worked around locally there at the time (filtering NaN-containing observables before the call); this PR is the actual upstream fix.

Change

• np.fmax (NaN-safe: ignores NaN in either operand) instead of np.maximum.
• Seeded with NaN, not 0.0: a factor with zero valid (non-NaN) measurements across every observable now correctly stays NaN throughout rather than silently settling at a falsely-confident 0.0 (which would look identical to "tested and found unimportant").
• Warns when that all-NaN case happens for one or more factors -- they're still dropped (no valid data to compare against threshold), but now visibly, since that's missing data, not confirmed unimportance.

Test plan

• just ci passes, design.py at 99% coverage
• New tests: a NaN-valued observable no longer erases a real signal found on a different observable (the exact pp-eigentest scenario); a factor NaN across every observable is dropped and warns; no false-positive warnings when there's no NaN at all

🤖 Generated with Claude Code