You can reliably use I2C decode on the LA on this scope, but there are some limitations. There are fewer limitations when using the analogue channels to do an I2C decode.
Firstly, 10Mpts seems to be about the most it can manage with about 30% active I2C data to decode. Although it might say "Too many events" on the main trace, you can still decode in the zoom trace.
Secondly, the sampling rate of the scope refers to the analogue sampling rate, not the LA sampling rate, which is 1/8 of the analogue sampling rate. The minimum LA sampling rate that works reliably for 100kbps is 625kSa/s, or 5MSa/s analogue sampling rate. Correspondingly, the memory depth of the LA at a given analogue sampling rate will be 1/8 of the analogue memory depth (e.g., 100k analogue samples will be 12.5k LA samples).
With this in mind, you can decode in a single sweep about 2s of 100kbps I2C. with a 5MSa/s analogue sampling rate and 10Mpts memory. That should be more than enough for your example.
Now there are other ways around the problem, but knowledge of the limitations mentioned above is a good starting point.
Option 1 is to simply introduce a trigger delay until the packet you want is captured. This is a bit hit and miss, but on a simple scenario that might be all you need.
Option 2 is to use the Record functionality (also known as segmented memory on Keysight scopes). In an example I just tried, I could capture 9,175 I2C packet bursts over a period of 4.5 minutes. Each packet burst is 34 bytes long (16 I2C starts/restarts). I set up a pulse trigger of 2ms on SCL, so each recording is started once the bus has been dormant for >2ms. Note I had to use an analogue channel for the trigger, as LA channels don't seem to work with a pulse trigger. Each recording was set for 100kpts at 5MSa/s analogue sampling rate (625kbps LA sampling rate).
The trick of Option 2 is to set up the timebase for your longest packet burst. To deal with longer packet bursts, you'll need a longer timebase and more sampling points, and so this will reduce the total number of recordings (or packet bursts) available to you.
While using a USB streaming LA is a viable alternative for this scenario, once you understand the scope it should do what you desire. Using a scope for this kind of work is a different workflow to using an LA: with a scope it's all about the triggers, whereas with a USB streaming LA it's a data grab and post processing. There are pros and cons to both, and it's not really a one-size-fits-all. I use both on a daily basis, but typically I go for the MSO first as I've been brought up with scopes and triggers. YMMV, I can see the friction of a learning curve around the scope might well be a turn off.