Should this hook into any of the HUGE/TRULY_CRAZY_HUGE diff detection stuff that got fixed a few months ago instead of defining a new constant here?

I think my gut reaction here is sort of a "maybe". The diff stuff is large enough that it has quite a few different magic constants which are somewhat similar to this one (e.g., intradiff bailout in T1246, which must be in arcanist/; the highlighting limit in T7895, which is used by both Differential and Diffusion; the "ENORMOUS" stuff discussed in T13143; various other limits like "maximum number of things we show in the web UI before degrading"). I think this stuff is spread out enough that it probably doesn't necessarily make sense to try to group it all together -- it tends to affect different workflows and different areas of the code, even though many things are implementing similar ideas.

There's some kind of argument we could make that we should have no hard-coded magic numbers and all constants should be configurable, but the number of actual cases we run into where installs want to change one of these numbers is very small. Many of these constants can be difficult to fully understand the implications of, and mis-configuring them can cause big messes which are hard to support (e.g., PHI944 took an hour or so of digging to end up here), and the cases where installs do want to change these numbers often feel questionable (e.g., the driving use case is somewhat sketchy or abusive). And these constants are ultimately configurable since you can just edit them if you really want, as long as you're accepting the cost/risk of maintaining a one-line patch which adjusts a number.

The thing I like least about this patch is really that we don't explicitly tell the user we skipped the detection. I think a "better" version of this patch would probably write a metadata property like copy:lines:skipped-because-it-is-too-big = 2243921 (see inline) and then the UI would say "Copied code detection was skipped for for this diff because it is very large (2,243,921 lines). To improve performance, copied code is detected only for small revisions (65,535 or fewer lines)." However, I think this bailout is very likely above the threshold where anyone will ever actually notice or care, and adding that notice would make the change a bit larger and more complex, and the notice itself is probably just clutter in close to 100% of the cases where it would appear.