fix(agglomeration): tolerate float rounding in stale-edge rescore

[dodamih]

Problem

Agglomeration aborts with:

IterativeRegionMerging.hpp:99: ... mergeUntil(...): Assertion `newScore >= score' failed.
Fatal Python error: Aborted

It happens with the default OneMinus<MeanAffinity> scoring, and is data/parameter dependent ("fails a lot on 16nm seg, some succeed").

Root cause — 1-ULP floating-point rounding

When a stale edge is popped, its score is recomputed (scoreEdge) from the running/area-weighted mean-affinity formulas, which are float32 and order-dependent. The recomputed score can land one ULP below the score the edge was queued at, and the strict assert(newScore >= score) aborts the whole process.

Instrumenting a live reproduction (replacing the assert with a logger) showed the truth: every violation has gap = 5.96046e-08 exactly (= 2^-24, one float32 ULP), on high-affinity (mean ≈ 0.99), tiny-contact (count = 3-6) edges:

score=0.00109613 newScore=0.00109607 gap=5.96046e-08 mean=0.998904 count=3
score=0.00297725 newScore=0.00297719 gap=5.96046e-08 mean=0.997023 count=3
...

Only 23 such edges out of ~48M — but one abort() kills the run. It is not an algorithmic non-monotonicity; MeanAffinityProvider and this assert are unchanged since the original upstream. Heavy over-fragmentation (watershed size_threshold=0 → millions of tiny, high-affinity supervoxels) is what makes the running-mean rounding land on the wrong side of the last bit, right where 1 - mean is tiny.

Fix

Tolerate sub-epsilon rounding noise:

assert(newScore >= score - 1e-5);

1e-5 is far above the observed 6e-8 wiggle yet far below the discretize_queue bucket width (~1/256 ≈ 3.9e-3), so any real non-monotonicity still trips it. In release (NDEBUG) builds the assert is compiled out anyway; the edge is simply re-queued at its recomputed score.

Verification

A deterministic reproduction (cutout1 affs, size_threshold=0 watershed → 11.7M SVs, agglomerate @ 0.35) goes from abort() → clean completion, 0 violations with this change.

Note: fully faithful reproduction at the production scale (size_threshold=50, full bbox, ~1.21M SVs) needs ≥64GB RAM; validated locally at size_threshold=0, which exercises the same float-rounding path.

🤖 Generated with Claude Code