background_deviation_box discards the xpx.at result — silent no-op on immutable backends

[mwcraig]

Behavior

On the jax backend, background_deviation_box returns a flat array (every pixel = the global xp.std(data)) instead of per-box sigmas — silently wrong results, no error (verified empirically 2026-07-27). On numpy it works, but only by side effect.

Root cause

The xpx.at(...).set(...) result is never assigned:

ccdproc/ccdproc/core.py

Lines 1341 to 1346 in 736555d

	barr = data * 0.0 + xp.std(data)
	ylen, xlen = data.shape
	for i in range(int(0.5 * bbox), xlen, bbox):
	for j in range(int(0.5 * bbox), ylen, bbox):
	x1, x2, y1, y2 = setbox(i, j, bbox, xlen, ylen)
	xpx.at(barr)[y1:y2, x1:x2].set(float(sigma_func(data[y1:y2, x1:x2])))

On numpy, xpx.at takes its in-place fast path so the discarded return value doesn't matter; on immutable backends (jax) it returns a new array that is thrown away, leaving barr at its initialized value. This is the same discarded-update bug class fixed for flat_correct/ccdmask in #956, and (as far as I can find) the last live instance on main.

Note on #931

The "Suggested fix" in #931 cites this exact line as the correct-pattern exemplar to copy — please don't: the pattern is only correct with the result rebound. I'll drop a note on #931 pointing here.

Suggested fix

One-line rebind:

barr = xpx.at(barr)[y1:y2, x1:x2].set(float(sigma_func(data[y1:y2, x1:x2])))

plus a jax-lane regression test asserting per-box values differ from the global std (the numpy lane can't catch this — the in-place fast path masks it).

Follow-up to #909 / #956. This issue was written by Claude (via Claude Code); Matt reviewed and approved its content before filing.