fix(exposure): scale the luma window's reach by each channel's own range

[seanharding]

Problem

The bounds analysis widened every channel's color refs to the luma window by the same amount in log units. On a negative whose channels differ in density range, that reach is a different fraction of each channel's span, so a gray scale normalizes to a different value in each channel and the print takes a cast.

A faded 126 color negative showed it once Sensor Correction was on. Unmixed, its green density range is about 3.5 times blue's (0.67 against 0.19 log), and the frame rendered yellow-green. With the unmix off, the sensor crosstalk blends the channels and hides the difference, so the profile looked like the cause. It is not: the matrix is close to a bare-light one, and all three unmix modes gave the same cast.

Note: I think this warrants some testing by other people on their own frames to confirm that it doesn't introduce any unexpected problems.

Change

analyze_log_exposure_bounds_from_log now expresses the luma window's reach past the color refs as a fraction of each channel's own span, averaged over the channels. Both engines call this one function, so there is no shader change.

The window keeps its mean-minus-median base offset, so a frame with exactly equal channel ranges gets the same bounds as before, and the tone-reproduction test numbers do not move. A channel with no span falls back to the previous formula.

Effect on other frames

Real film never has exactly equal channel ranges, so most frames shift a little when their bounds are next measured. I rendered 59 saved color-negative frames with the old and new bounds (CPU path, 420 px, roll averages off):

• No frame took a cast or broke.
• Median mean-absolute change is about 4 levels out of 255, mostly brightness; many frames move 5 to 10.
• Frames with a range ratio near 2.8 move 16 to 20 levels and render darker.
• On two color chart frames, the mean channel spread over the six gray patches fell from about 5.5 to 4.3 levels. The mid-grays improved most; the white patch got slightly warmer.

Saved edits keep their stored bounds and do not change until a frame is re-analyzed.

Not changed

mix_luma_color_bounds (Roll Analysis luma and color averages) still shifts all channels by the same log amount, so a roll of this kind can show a milder version of the cast with those averages on.

Tests and housekeeping

• New: test_unequal_channel_ranges_keep_a_gray_scale_neutral; it fails without the fix.
• tests/test_same_pixel_bounds.py: the helper mirrors the new recombination, and the thin-end test's tolerance goes from 1e-9 to 5e-3, because the ceiling now depends on the span.
• THUMBNAIL_RENDER_VERSION 3 → 4. Recording its golden also picked up the autocrop_square and autocrop_squared defaults it was missing.
• docs/PIPELINE.md describes the new formula.
• Checked in the desktop app on the 126 frame with the profile on and off.