Add Shapley attribution and redundancy/nonlinear validation cases to the failure harness

[duckyquang]

Summary

Adds exact Shapley values over modality coalitions to the harness's complementarity report, next to the existing leave-one-out numbers. Also adds the two synthetic stress cases the ML4H reviewer report asked for (its point M7): a redundancy plant (two equally strong modalities plus one weak) and a nonlinear MLP probe on the planted data, both wired into validate_harness.py. The fused saved-predictions route in run_failure_analysis.py now degrades gracefully (shapley_value: null) instead of exiting when the all-masked subset has no scored condition behind it.

Result

All from real runs on my Mac (CPU; the validation script is synthetic by design):

• Redundancy plant: Shapley pays each redundant modality ~44% more than leave-one-out (3.43 / 3.75 vs 2.38 / 2.62, weak channel 0.75 vs 0.22). This is the textbook LOO blind spot the reviewer flagged — with a single masked checkpoint the twin covers for the dropped modality, so LOO under-credits both.
• Nonlinear MLP probe: degrades badly under either mask (silent rate 0.587 / 0.571) — the same out-of-distribution-zeros collapse the real fused checkpoint shows on the cohort, so this doubles as response material for reviewer point M2.
• Efficiency property holds exactly (sum(phi) = MAE(empty) - MAE(full)) and is tested, along with the N=2 closed form.
• 135 passed, 1 skipped, ruff check + format clean.

Notes

• shapley_value is None past N=8 (2^N predict calls) or when predict_fn can't produce the all-masked prediction — the fused-route JSONs only score the three named conditions, so they keep LOO-only attribution.
• The changed error type in run_failure_analysis.py (SystemExit → KeyError, converted back at the call site) is what lets the harness probe the empty set opportunistically without breaking the script's clean exit message for genuinely missing conditions.
• Multi-seed and 3-modality validation outputs are unchanged, so nothing in the paper's Section 6 numbers moves.