1. Neither Calibre PEX, not any other industry-standard extraction tools, using rule-based (pattern-matching) extraction, can guarantee extraction accuracy for capacitance that is sufficient for precision analog designs, such as 12-bit data converters.
2. Very roughly speaking - in rule-based extraction, the total capaictance of large nets can be accurate within a few percents, small nets (few fF and less) - within few tens of percents, and small capacitive couplings (very roughly, less than fF, or less than ~1% of the total net capacitance) can be missed completely or be very inaccurate.
3. There are many settings in parasitic extraction that affect the accuracy of the capacitance extraction. As an example - extraction tools can "decouple", i.e. replace coupling capacitances by grounded capacitances, if they are small enough. This can get very tricky.
However, there are basic, fundamental limitations, for extraction accuracy of rule-based extraction.
4. The best way to achieve a high accuracy of capacitance extraction is to use a "field solver" - a first-principle technique to calculate parastiic capacitances, though rigorous solution of (Laplace) equation for the electrostatic potential.
In the industry-standard extraction tools (StarRC, Quantus QRC, Calibre PEX,...), field solver can be turned on by a single command in command file for parasitic extraction. There may be be few more commands to control some details - such as what nets to extract, required accuracy, etc.
Field solvers take more time than rule-based extraction, but they can achieve very high accuracy.
Random-walk field solvers (built-in in StarRC and QRC, also QuickCap, and some other extraction tools) can achieve accuracy of 0.01% or better - and he required accuracy can be specified upfront. (the higher the desired accuracy - the longer the extraction time).
The tradeoff here is extraction time vs accuracy.
You want to improve accuracy by 2x - you need to wait 4x longer, until random walk algorithm achieves that accuracy.
Rule-based extraction does not allow such control (accuracy vs extraction time vs circuit size).
Please note that Calibre PEX does not have random walk field solver, it uses deterministic field solver, where error cannot be estimated and controlled.
5. Regarding extraction results being dependent on hierarchy - this in an expected effect. Higher hierarchy introduces more metals (more layers, and more "neighbors), net shapes may become different, context/neighbors may become different, etc.
Also, there are other factors that might affect the results - such as presence of floating metal fills.
There are some settings that can make extraction faster, or less accurate, or generate a smaller size netlist - but these commands should be used with great care (i.e. you need to know what you are doing), otherwise the results may become completely useless, not accurate.