fix(chirp): background baseline mean skips NaNs (chirp-edge rows wiped whole columns)

[garrekstemo]

Bug

With a background correction downstream of correct_chirp (QPSLab orders chirp first), every wavelength column with a negative chirp shift came back fully NaN — 409/512 display columns null on a real S1.5 dataset.

correct_chirp NaN-fills samples shifted past the time-axis edges. For negative shifts that NaN region is at the start of the axis — exactly inside typical pre-pump baseline windows (e.g. −20…0 ps). subtract_background computed the per-column baseline with a plain mean, so a single NaN row poisoned the entire column via data .- baseline'. Positive-shift columns survived only because their NaN row is at the end of the axis, outside the window.

Fix

• subtract_background(::TimeResolvedMatrix): baseline is now the per-column mean over finite values only. A window with no finite values yields a NaN baseline (the column stays NaN — no background is determinable) and emits a warning.
• _auto_baseline_range: row means skip NaNs, and NaN gradients are scored -Inf so they cannot win the onset argmax (previously a NaN would, since isless sorts NaN last).
• Version 0.1.2 → 0.1.3.

Tests

• Chirp (+1/−1 ps shifts, sign flip mid-axis) + explicit-window background: no fully-NaN columns, finite samples subtract exactly, chirp NaNs pass through untouched.
• All-NaN window: warning emitted, affected column NaN, other columns exact.
• Auto-range path on a chirped step-onset matrix: finite range, no fully-NaN columns, pre-pump rows ≈ 0.

Full suite: 1899/1899 pass locally.

QPSLab will pin [compat] OpticalSpectroscopy = "0.1.3" and add an end-to-end regression test on its side.

🤖 Generated with Claude Code