Some rationale on this change...
Previously when applying a "regular" subset string, ie. one that is only the
content of a where clause, we issued a full "SELECT * FROM layer WHERE subsetstring",
resulting in a OGR SQL layer. The caveat of that is that most OGR drivers
will have issues retaining the original FID. A hack consisting in adding a
{original_fid_name} as orig_ogc_fid to the select columns was introduced in
4ce2cf1744
to try to retain the original FID, but this added a lot of complexity. And
actually, in the case of the OGR GPKG driver, it caused it to still be confused
when analyzing the column definition of the resulting layer, since it sees
2 FID columns despite the renaming (one included in the '*' wildcard, and the
one of orig_ogc_fid), which caused it to use sequential FID numbering (the
driver when seeing more than once a column that is the FID column assumes that
some cross join is done, and thus that FID are unreliable)
A simpler and more robust (crossing fingers!) approach in that case is
just to use OGR_L_SetAttributeFilter() instead of GDALDatasetExecuteSQL().
Some care must be taken to cancel the filter when removing the subset
filter, or in QgsOgrFeatureIterator when combining with the filter
expression coming from the request, but besides that, this is more
straightforward, and actually solves #20136
Actually it had nothing to do with CSV being the
source, but it was the json exporter passing
the values through all field formatters except for
the fallback.
This resulted in all fields using a 'Range' formatter
(which is the default for all numeric types) passing
through the formatter and being returned as strings
in the json. Worse, if the locale was not a "dot"
locale and decimal separator was on, the resulting
string could not be easily converted into its original
numeric type.
Now, instead of checking for the fallback formatter
only, there is a white list of formatters that
can be applied when we want a json.
This is a temporary solution because the "right" way
to do it would be either a flag in the formatter to
tell if it can be applied when converting to json
and/or other "data" formats (csv etc.) or a different
new method similar to representValue.
