Resolution of recording surface

[Bernhard R. Fischer]

Dear list,

It has been a long time since I posted here for the last time. Sorry for 
the long text but it's to so easy to explain.

I'm using Cairographics extensively for an OSM data rendering software 
since many years (https://github.com/rahra/smrender), so I'm not new to it.

What I regularly do is to use a recording surface internally and finally 
paint it to a PDF, SVG, or whatever surface is required.

In respect to the units I always assumed that 1 unit within the 
recording surface is to be 1 millimeter in the final result. To 
accomplish this I create the recording surface with the size of the 
final page (e.g. create_recording_surface(210, 297) for an A4 page have 
210x297mm), then I draw on it and finally at the cairo_paint to the PDF 
surface I scale the paint operation to meet the points unit of the PDF.

This concept works fine but unfortunately recently within a different 
project (where I use the same drawing concept) I made the discovery that 
the resolution of the recording surface seems to be limited at least 
with the toy text API. I use the get_text_extents and similar functions 
to place some text at specific positions and with some debugging I found 
out that it does not use any fractional part of the extents member 
variables.

E.g. the text_extents.width always returns e.g. 3.0, 4.0, 8.0, but never 
something like 3.17, 4.28 and so on. The final result is that the text 
(obviously) is not placed on exact positions (where it should be).

With further experiments I found out that if I make the recording 
surface width/height e.g. 100 times bigger (21000x29700) and then scale 
it by factor 100 the text_extents will work with higher resolution as 
expected. But if I further increase the recording surface to a factor of 
e.g. 1000, it will internally hang somewhere within an endless loop 
(within libcairo). So there seems to be some upper limit.

So my question is: is there a limit in the resolution of the recording 
surface? Or is this a limitation of the toy text API? Or did I stumble 
on something else? Or is my drawing concept somehow incorrect?

Best regards,
Bernhard