Public PPR result. truncated is logical-OR of forward + backward
push truncation flags. When true, downstream renormalization
(sum-to-1 across nodes) hides the fact that the iteration was cut
short by max_pushes_cap; the absolute relevance scores are
biased toward seeds and not directly comparable across instances.
We surface this to Python so calibration / final-eval rows can be
filtered in post-analysis (PolyBench / Multi-SWE-bench instances
with >20k fragments are the primary suspects).