fix(exposure): reach past the color refs by one fraction of each span

[marcinz606]

Summary

The bounds analysis widens each channel's color refs out to the luma window. On main that reach is the same in log units for every channel. On a negative whose channels differ in density range (a faded negative, for example), it is then a different fraction of each channel's span, so a gray scale normalizes to a different value per channel. A negative Luma Range Clip also tints the frame as it widens the window.

The reach is now one shared fraction of each channel's own span, computed as a ratio of sums: h = Σ(l_c − r_c) / Σ s_c.

This replaces #1366, which averaged the per-channel ratios. That version lets a near-flat channel drive the reach for all three: a negative Luma Range Clip margin becomes m · s_c · mean(1/s). The ratio of sums keeps the old mean log reach and widens every channel by the same fraction. The tests and docs are adapted from #1366 (co-authored by Sean Harding).

Measurements

264 color frames from 14 rolls, measured with the Roll Analysis decode path (saved crop, flat field and unmix). An A/B/C/D toggle in the app was used for visual checks.

	main	#1366 (mean of ratios)	this PR (ratio of sums)
Gray tilt vs the frame's own gray axis, R/B (log)	0.110 / 0.138	0.079 / 0.116	0.079 / 0.116
Error on a synthetically faded copy (normalized)	0.074	0.048	0.049
Channel spread of window growth at Luma Range Clip −0.3 (P90)	1.19	1.03	1.03
Mean change vs main (normalized)	0	0.027	0.026

On these frames this PR and #1366 match closely. They differ only when one channel's span is near flat, which none of these frames have. The new headroom test covers that case.

Known limit: a strongly green-dominant scene (lawn, forest) can read as unequal channel ranges and pick up a slight magenta cast in the shadows. Within one frame, scene content and real fading cannot be told apart.

Changes

• normalization.py: ratio-of-sums reach. Spans of mixed sign keep the single log reach.
• THUMBNAIL_RENDER_VERSION 3 → 4. The recorded golden also picked up the missing autocrop_square / autocrop_squared defaults.
• docs/PIPELINE.md: bounds paragraph.
• Tests: the gray scale at unequal channel ranges stays neutral, and a negative Luma Range Clip widens every channel alike, even with one channel near flat. Both new tests fail on main.

Saved edits keep their stored bounds. A frame changes only when it is measured again.

Test plan

• make all (lint, ty, full pytest)
• New tests fail on main's recombination
• Implementation matches the measured "Ratio" variant bit-exactly on all 264 frames at clip 0, −0.1 and −0.3
• Visual check on a faded negative