proposal: cmd/go: fuzzing option to add the last few "interesting" inputs to the seed corpus

[myaaaaaaaaa]

Proposal Details

This is somewhat of a successor to this declined proposal, with the same motivation:

• proposal: cmd/go: go test should always fuzz a bit even without -fuzz #50969

I'm too lazy to come up with a seed corpus for otherwise succeeding tests.

Essentially, the last few "interesting" inputs are the most difficult ones to find. I think they're the best candidates for checking into the seed corpus.

Saving just those means that future fuzzing runs can reach maximum coverage much faster, with minimum repository bloat.

[Jorropo]

We already store all interesting cases in the fuzzcache, and the fuzzcache is loaded and re-tested in the gathering baseline coverage part when you start fuzzing.
Outside of using all test cases vs some amount of recently found one, how is your proposal different ?

[myaaaaaaaaa]

Adding them into the seed corpus means they would run automatically as part of go test. This improved determinism and repeatability is more compatible with CI. It also means that only one developer needs to pay the cost of fuzzing, and then other developers who don't fuzz during development also get the benefits.

Essentially, I am treating this as a more advanced form of table-driven tests. I would like to be able to write some property tests, and have the fuzzer automatically populate a test table for them. Use of coverage guiding means that this test table will contain very diverse test cases.

Presumably, some users would like to limit the size of this test table, and avoid checking every single interesting input into the seed corpus. Hence the "last few interesting inputs" part, which I think is better than using the first few interesting inputs.

[Jorropo]

So you want interesting inputs to be added to the testdata folder ?

[myaaaaaaaaa]

In short, yes