Fix reproducibility CI: regenerate stale BCF references and fail on drift

[andrewherren]

Problem

The cross-platform reproducibility jobs were passing despite mismatches. Two issues combined:

1. The check scripts only print differences — they never exit non-zero, so the job stayed green even when results didn't match the references.
2. Under that blind spot, the four BCF reference CSVs had gone stale: every BCF value (200 y_hat + 100 sigma2) differed from the stored reference in both R and Python. (BART was fine.)

The current BCF output is in fact cross-platform reproducible — it agrees to ~1e-9 across ubuntu/windows/macos with 0 mismatches at the 1e-6 tolerance. The references were simply out of date, which is why every platform reported the same differences.

Changes

• Regenerated the 4 stale BCF reference CSVs from the v0.4.5 release build (origin/main == v0.4.5), built and run in both R and Python, in the original CSV format (headers, precision, line counts unchanged).
• Made the checks fail loudly: simple-bart/simple-bcf (R and Python) now exit non-zero on any mismatch, so future drift turns the CI job red instead of passing silently.

Verification

• All 4 scripts pass with exit 0 against the regenerated references.
• Confirmed the new fail path returns exit 1 when a reference is perturbed (R and Python).
• BART references untouched (already correct).

🤖 Generated with Claude Code