Reapply "precompile: fail in (closer to) linear time and linear error messages" with fixes

[vtjnash]

With thanks to Ian and Kristoffer for pointing out the needed fixes

• Will need: runtests(ALL, configuration = (goal=:load,), vs_configuration = (goal=:load,)) before merge

[vtjnash]

@nanosoldier runtests(ALL, configuration = (goal=:load,), vs_configuration = (goal=:load,))

[nanosoldier]

The package evaluation job you requested has completed - possible new issues were detected.
The full report is available.

Report summary

✖ Packages that failed

327 packages failed only on the current version.

• Package fails to precompile: 213 packages
• Illegal method overwrites during precompilation: 108 packages
• Package requires a missing binary dependency: 2 packages
• Test duration exceeded the time limit: 4 packages

949 packages failed on the previous version too.

~ Packages that at least loaded

2 packages successfully loaded only on the current version.

• Other: 2 packages

8941 packages successfully loaded on the previous version too.

➖ Packages that were skipped altogether

1 packages were skipped only on the current version.

• Package could not be installed: 1 packages

889 packages were skipped on the previous version too.

[vtjnash]

We may need to completely rethink how extensions to stdlibs work again, since this is now just a hack job to only be partly badly implemented, since the trigger sets being used for computing the safe operation are entirely wrong if any of them got added into a custom sysimage.

[IanButterworth]

That sounds like a merge blocker? Stdlibs are very commonly added to custom sysimages?

[vtjnash]

I mention it because it is a pre-existing mistake, not because it is introduced here

[KristofferC]

I don't know if #52841 is relevant but it was trying to make the set of packages in the sysimage irrelevant to what extensions are loaded. But it was buggy.

[vtjnash]

Yep, it was basically the same as that issue. But also making the observation that the resulting hack there (#55589) could be badly broken in the case of trying to compute the dependency graph of something that got added to the sysimg because it may refuse to load extensions for subsets of dependencies if all of those are in the sysimg (because that would incur a dependency cycle and not be solvable / loadable)