Hi Al,
I do believe I am getting ready to stop beating on this issue !
I re-processed the FISH Lu linearity data (45x Blue & 45x Red files)
using (mostly) Steph's approach:

1.0, raw image avg 44 pixels-per-track in along-slit dimension, yielding (ADU/pix)
2.1, avg & std of 4x Dark scans
2.2, avg & std of 5x Lite scans
3.1, NetSignal = avg(Lite) - avg(Dark)
3.2, Noise = RSS( std(Lite), std(Dark) )
4.1, avg of 21 pixels of NetSignal in along-spectra dimension
4.2, std of 21 pixels of NetSignal in along-spectra dimension
4.3, Combined Standard Unc. = 1/21 * RSS( Noise, 21 pix along-spectra ),

where, RSS = squareRoot of Sum of Squares,
note: I did not normalize for exposure time nor gain setting.

The idea being to compare the stDev of 21 spectral pixels of NetSignal (4.2) with
the comb. std. unc. (4.3), which comes from the "Noise" of Lite & Dark scans.

The attached figure  <20130125-LIN02_01.png>  plots:
log scale X axis = 45x NetSig 21-pix-avg, blue symbols = BSG, red = RSG,
top panel = std vs unc ADU,  mid panel = percent std & unc, bot panel = mid zoomed.
Attached  <20130125-LIN02_02.png>  is the bottom zoomed plot.

The single Blue & Red data points ~1 ADU I think was a dark/ambient scan.

To me, it looks like the ADU variability among 21 spectral pixels increases with
increasing NetSignal ( top, "x" & "+" ), while the c.s.unc. from Lite & Dark Noise
remains fairly constant ( top, "o" & "square" ).
Relative to the NetSignal level, both the std & c.s.unc increase with decreasing signal (bot, left).
At low NetSignal the std & c.s.unc. somewhat agree (bot, left), but
at higher NetSignal the std is larger than c.s.unc (bot, right).

So, at NetSig > 1k ADU, where stDev said we had  ~1% Blue "noise", and ~0.3% Red,
the comb. std. unc. says we are really ~0.05%, with the Blue & Red ~equal.
Less than ~100 ADU the Blue does appear to be noisier than the Red.

So there. MF