Roll/Scene baselines: pool by whole frame, outliers keep their own bounds

[marcinz606]

Problem

With Use Luma/Color Average on, some frames rendered with a heavy cast (marbella_1 #19: neon cyan) while their own bounds rendered correctly.

• The baseline IQR-trimmed each channel independently, so each channel could average a different set of frames (PARTY scene: red floor from 016/019/024, G/B floors from 014/016/021).
• A frame far from the pool was reported, but it still received the baseline.
• Every frame was measured with the active frame's crosstalk, not its own.

Across three rolls (99 frames), the outliers are the first or last frame of each 6-frame strip. The scans were balanced per frame so that the brightest area reads white: the orange base mid-strip, bare light past the cut at a strip end. So strip-end frames keep the orange mask (0.4–0.6 D denser in G/B). The film-base level alone classifies 95 of 99 frames. The scanner-side fix (an exposure lock in the nkscan backend) is a separate change.

Change

• pool_frame_bounds (normalization.py): compares frames on luma-free color, and flags whole frames further than POOL_OUTLIER_DISTANCE (0.3 D) from the median. The baseline pools the inliers only: luma is the median, color the mean.
• Outliers keep their own bounds, with both averages off. Taking the baseline luma with the frame's own color rendered washed out on every real outlier, because the color offset also shifts luma through G.
• The outlier list is stored with the roll/scene baseline (outliers), so a later Apply keeps it. Old records read as empty.
• Each frame is measured with its own crosstalk, as it renders. The override_crosstalk_* fields are removed.
• Status message, confirmation dialog, PIPELINE.md and USER_GUIDE.md are updated.

Validation

• Tuned on marbella_1. On the held-out rolls 2 and 3, the same rule flags only strip-end frames.
• Inliers render as before with the new baseline, and outliers render as they do with their own bounds.
• A fixed threshold was chosen over a spread-scaled one (K × MAD), because the latter flagged nothing on a roll with two balance groups.

Test plan

• make all: lint, ty and the full pytest suite
• New tests/test_baseline_pooling.py, plus worker, controller and rolls tests
• Re-run Roll Analysis / Analyze All Scenes on an existing roll. Old baselines have no outlier list until they are re-run.

[github-actions]

Preview load metrics (github-linux-x64)

format	cold load	throughput
CR2	0.765 s	29.32 MPx/s
NEF	0.624 s	39.38 MPx/s
ARW	0.219 s	92.17 MPx/s
RAF	2.127 s	7.65 MPx/s
DNG	0.653 s	36.64 MPx/s

fixtures: rawsamples.ch · sha: d7660ce · full JSON