Generally, if you give your customers the ability to flash your device, you want them not to brick it.
Thus, the main priority providing is a robust and safe connection to the host device. In my experience all of the chips work rather well in the urban and laboratory settings which I assume you are targeting.
Consider also the time it would require to get the component to work right - your production count is in the low hundreds, and you add $2 to each. Is that "loss" worth more than your (or the developer's) time? A well-documented chip might be actually a better choice, even if it costs more.
Yes, it is really a bad Idea, at least if you want Plug-And-Play. For me on Win 8.1, I needed to manually install the driver (
CH340SER.exe) which I had to download from the (Chinese) manufacturer's website which had (at that time) no English translation.
It was hosted in China, which might be an issue for security-minded individuals and those bound by organizational and/or political rules. Also it was outranked as a search result by a lot of dubious "Free" driver download sites.
If this was any serious equipment (opposed to "just" an Arduino), that'd be raising my eyebrows to the ceiling.
Manually installing might also very annoying if your customers don't have dedicated Equipment for flashing.
Otherwise, this chip worked as expected.
Not much to say about, worked out of the box and did not bring up any issues. Would probably be my choice for an average design.
I have this one on an standalone USB-Serial converter board and it performs well. As you wrote, it is rather expensive, but I believe it might be a better choice in rough environments (for example corroded contacts on connectors, also EMI). Might give you a warm fuzzy feeling because you support the original developers.
As per @Chetan Bhargava's answer, an option would be to have an connector for the SPI and then use a standalone USB-Serial converter.
This also requires you to provide a reliable and safe to use connector for the ISP to attach to. Obviously you can go cheap with pin headers here, but if you want to do it right (and/or don't trust your customers enough), then this connector might be more costly than the additional chip and a stock USB connector.
Serial connections are furiously hard to debug if they don't work, opposed to USB where the customer will at least get notified that the USB device is not working.
If you bundle a standalone converter with your board, you will have to pay the price for the converter board too. I'd assume this would not be cheaper than to integrate the chip. This could work if each customer owns many of your boards, so the converter could get reused, or if you can just shove the costs of acquiring the programmer to the customer side.
If this option is a possibility, at this point, there's also Atmel's very own AVRISP which is a good choice here instead of the plain USB-to-Serial, though a bit outdated. I think it caps out at about 100 or 200 kbps where modern USB-Serial converters go into the megabit range. But it is very robust in regard to (mis-)usage.
Another good option could be a TC2030 connector. It only requires pads on the PCB to work with, but requires some expertise (you have to hold it on the spot until the programming is finished).
Modern microcontrollers also come with a number of other communication interfaces (Ethernet, WiFi, Bluetooth) and usually can be flashed using these. An example would be the ESP32 which comes at a cost of about 6 USD and is a SoC with all components necessary for a wifi connection. Also it is Arduino-compatible (you can even use the IDE) and has a very thorough example set, including an WiFi OTA bootloader. You would only need an ISP for the initial deployment of the bootloader.
If - as it sounds in your question - your project is mostly finished, this is probably not an option anymore.