Make diagnostics a first class citizen in LSIF

[felixfbecker]

I am very interested in the potential for LSIF to track code smells in a code base. However, right now diagnostics feel like a second-class citizen, because their ranges are in a big array instead of using Range vertexes. It would be great if eventually we could include precomputed fixes for the diagnostics in LSIF too (through code actions), which would then want to share Ranges with diagnostics.

Is there any benefit to have the diagnostic ranges isolated in the array?
The only difficulty I can think of is that diagnostic ranges might overlap with symbol ranges in ways that the spec currently does not allow.

[dbaeumer]

The major motivation was because of:

1.) LSIF dumps are usually produced on error free code so there will be only little diagnostices
1.) they might not overlap with the sensitive ranges.

However I do agree that they are the same in many cases so we might be able to think about a more optimized storage format for this.

[felixfbecker]

I created an LSIF exporter for ESLint to explore this: https://github.com/sourcegraph/eslint-formatter-lsif

My next steps would be to try it with a different Diagnostic format (using Range vertices), then running the validator over it to see if any ranges are overlapping.