Well, yes, the behaviour is expected, the button on the VFB is the view sRGB, not the sampler’s internal color mapping one.
In essence the two operations are applied at different times (one before light transport is evaluated, the other when the games are done, on the pixels of the rendered image), and produce very different results (well, not the sRGB alone, in this case, but non-linear mapping in general.).
Think of it this way, if i oversimplify it a bit for argument’s sake:
Say you apply a gain of 2.0 to the internal color mapping, by whichever means (say, halving or doubling bright and dark multipliers, and so on.), that will double the intensity of the distributed light as well, double the diffuse level of the surfaces, their reflectivity, and so on, in essence skewing light transport, and baking it into the rendered image (which is where ultimately you see the results of the CM skewing.), and in this case the net result will be a strong lowering of the light distribution’s contrast, along with a substantial brightening of the render.
Now, say you apply the same gain at the VFB stage, or in other words after the (expectedtly) radiometrically linear light transport equation has been resolved, now you cannot in fact touch the GI separately, or the diffuse, from the beauty RGB alone (Cue REs and post. but we said no to all that, right?).
This means that the light will have fallen off correctly during the calculations, on the right albedo shaders, and then had been multiplied as an image operation afterwards, the contrast of the solution unaffected.
EDIT: as a corollary to their order, and very different operations, the view color mapping (all of the controls, including the sRGB button) are live, and directly undoable. Any of the internal color mappings will require a re-render to be applied to the scene’s values interpretation before they’re fed to the rest of the process.
Of course, if people like to depart from LWF, they can do so.
I agree on the fact the only solution for such cases may be a CM lineariser.
And of course have zero opinion on the feasibility or necessity of it (not my role.).
It’s surely the first time i hear such obstinate need to do an A-Over-B IN the VFB, with non-linear CC to the scene but not the plate, instead of using an alpha channel and any of the free compositing or painting tools (gimp, if that’s what it took me.).
I simply can’t fathom why i’d lock myself into said linearised BG at the price of a rerender for each minor correction: i’d rather prefer a different sort of torture, like solitary confinement… ^^