[guijacquemet]

Fixes #84

What changed and why

For a very narrow float64 range, hi*0.5 - lo*0.5 rounds a subnormal endpoint (a half range of zero for distinct endpoints) and n / half_range overflows to inf before the integer cast. Finite pixels were mis-binned (or ZeroDivisionError in _bin_index).

• New float_bin_position (in bin_centres.py) is used by _float_bins and _bin_index, so both share one arithmetic. Where halving is exact and n / half_range is finite (neither endpoint subnormal, half range at least n * 2**-1000) it keeps the existing halved arithmetic. Elsewhere the position is the bounded fraction (value - lo) / (hi - lo) times n; the halves are used only when hi - lo itself overflows.
• float_bin_centres uses lo + (hi - lo) * (fraction / 2) when hi - lo is finite, the halved/midpoint form otherwise (the Compute float bin centres without overflowing near the float64 limits #76 large-range case still passes).
• _bin_index clips the value into [lo, hi] first, as its bin is clipped.
• _float_bins tests lo == hi instead of a zero half range for coinciding robust limits.
• Code-rule baseline lowered by --update-baseline (two entries shrank).

Deviation from the proposed fix

The review proposed (value - lo) / (hi - lo) whenever hi - lo is finite. I tried that first: in a sweep of 1603 histograms it changed 18 (a pixel exactly on a bin edge moved one bin) because the rounding differs from the existing arithmetic. The requirement is that existing histograms stay bit-identical, so the existing arithmetic is kept where it is exact and the bounded fraction is used only where it is not.

Regression tests (tests/test_histogram.py)

• test_float_range_of_the_smallest_subnormal_keeps_distinct_bins
• test_float_narrow_range_does_not_overflow_the_bin_scale
• test_float_range_between_subnormals_keeps_every_pixel_in_its_bin

On the old code all three FAIL (3 failed, 10 passed for the selected tests); on the new code they pass.

Bit-identity of existing histograms

A sweep of 1603 histograms (float32 and float64; normal, uniform, on-bin-edge, exponential, linspace, NaN-containing planes; 8 to 4096 bins; with and without the robust-range check, bin counts, limits and representatives; plus three ±float64-max ranges) gives byte-identical JSON on origin/main and on this branch (sha256 8f119879...f1f8 for both).

Checks

python tools/check.py: 724 passed, 2 skipped (tests/test_rgb.py:80, csbdeep not installed; not claimed as passed), Ruff and code rules clean, exit 0.

---

Generated by Claude Code