[ESSSANS] fix(sans): snap wavelength band edges onto WavelengthBins

[SimonHeybrock]

Problem

_reduce splits (Q, wavelength) data into wavelength bands along paths that disagree about where a band starts and ends. Event data is re-binned onto the band edges, so bands are exact. Dense data is label-sliced, so band boundaries snap to whole wavelength bins. The I(Q) denominator (monitor × transmission × direct beam × solid angle) is always dense, so in a multi-band reduction the numerator and the denominator cover different wavelength ranges and each band is biased.

The dense parts also disagree among themselves, because scipp label slicing means different things depending on the wavelength coord. The denominator carries midpoints, so it assigns each bin to the band containing that bin's midpoint. The numerator that direct_beam histograms carries bin edges, so it picks up every bin overlapping the band, and adjacent bands overlap.

Measured per-band I(Q) error on Larmor data, against an exact integral over the same range:

bins / bands	normal path	direct_beam path
50 / 10 (loki tests)	0.00%	0.00%
140 / 10 (sans2d docs)	0.00%	—
overlapping-bands direct-beam test	−2.42%…+4.19%	+4.00%…+21.98%
50 / 7	−11.23%…+6.93%	−0.18%…+42.91%
141 / 10	−6.32%…+2.81%	—

Everything in the docs and tests happens to use bands aligned with the bins and is unaffected, except test_can_compute_direct_beam_with_overlapping_wavelength_bands, whose bands are 12/11 Å wide against 0.24 Å bins. Live data reduction never sets WavelengthBands and so is unaffected.

Fix

process_wavelength_bands snaps each band edge onto the nearest WavelengthBins value, taking it from the bins array so the two are bit-identical. All three representations then select identical wavelength ranges and every measurement above goes to 0.00%. Bands already aligned with the bins are unchanged bit-for-bit, so no currently-correct reduction moves.

An edge can move by at most half a bin width. This is not announced: the bound is small by construction, and the bands actually used are returned and become the wavelength coordinate of the result. Warning from inside a Sciline provider would fire 2 × niter times per direct_beam call, which spams or deduplicates into invisibility.

Alternatives considered. Rebinning the dense parts onto exact band edges keeps the requested boundaries, but needs WavelengthBins plumbed into _reduce since the denominator only carries midpoints, and the proportional split correlates adjacent bands and understates their variances — the trade-off weighed in #725. Selecting on bin indices rather than values removes the coord-kind duality outright, but requires ProcessedWavelengthBands to carry both indices and values, changing a public workflow type for no user-visible gain.

Because the fix now rests on an invariant that spans two providers, it is written on ProcessedWavelengthBands rather than left implicit in _reduce.

Behaviour changes beyond the bias

Bands falling outside WavelengthBins were silently clamped to whatever range existed, which turns a unit mistake or a stale config into a plausible-looking result. Those now raise, as do reversed bands and NaN edges. Every WavelengthBands setting in the repo is already in range, so nothing here breaks. Making these loud is what makes it defensible for the half-bin snap to stay quiet.

Relation to #670 / #725

Same family, different mechanism. In #670 the second axis is two_theta, a pixel-geometry coordinate that has to be regridded onto a coarser output grid, and the varying number of internal bins per output bin produced periodic spikes. Here the second axis is wavelength, a per-event coordinate binned once directly to WavelengthBins; nothing is regridded and there is no aliasing. The error comes from the band-splitting slice snapping to bin boundaries, and only on the dense paths.

Test plan

• New tests reduce the same flat-density data in all three wavelength representations (events, bin edges, midpoints) and assert the per-band results agree; they fail on unaligned bands without the fix.
• Snapping, overlap preservation, idempotence (direct_beam feeds processed bands back in), and each new error condition are covered.
• Existing directbeam_test.py and sans2d_reduction_test.py pass unchanged.