The question of spatio-temporal variation of filters is more complicated than a spatial model in jointcal.
John's initial issue says, "common failures ... are due to excessive outlier rejection of stars with beyond median color." I don't see how spatially varying colour terms help here – isn't the problem that the colour terms (i.e. corrections for different effective filter transmission curves) are not correct for extreme objects, possibly due to complex spectra for which a simple "colour term" isn't sufficient? In this case I see no alternative to trimming extreme objects out of the reference catalogues, which is what HSC does. Of course, bad (constant) colour terms exacerbate the problem of clipping, and that's probably what John meant.
If the problem is the spatial variation of the colour terms (which is what I think Pierre's concerned about) then I agree that that's something that we could estimate in jointcal, but I don't think we want to do that in the long term because (as Jim notes) there's lots of external information. This isn't really a pure spatial variation but changes in time, either due to dropping or upgrading a filter, or due to changing atmospheric absorption. If we neglect the temporal dependency (which we cannot for HSC, as we've upgraded the i and r filters – and I'd argue that we should plan for the LSST filters evolving and LSST certainly needs to track the atmosphere) then jointcal is one place that we could track this.
In the short term a fixed spatial model of colour terms coefficients is reasonable, and whether jointcal estimates the terms or we get them from filterscans + models is a detail (we certainly wouldn't estimate them per jointcal run). Jim thought that Eli Rykoff had some problem with jointcal's (actually FGCM's) ability to estimate these constant-in-time spatial variations, so it'd be good to get that clarified. I think that for jointcal to measure spatial variation of colour terms would require that jointcal has colour information about each object even though I think it runs per-band (is that right, John?); maybe it does – they don't have to be all that good colours as errors only enter at higher order.
In the longer term we need to track effective filter curves for each measurement of an object: "I was measured on an ITL chip at this position in the focal plane on Thatcher Day 2027 under these atmospheric conditions" right through to the science database, and jointcal is probably the place to do this (it'll need per-object colour information to do this). It does mean that some of this information that jointcal is gleaning/assembling needs to be persisted as per-source "measurements", and that implies something about who has to be involved in designing the data products.