[OM] Delay assertions until worklist drained

[seldridge]

Fix an issue with the OM evaluator where asserts were evaluated too early.
This could cause the getStrippedValue reference chasing to hit a null
value and assert. By delaying assertions until the worklist has drained
this avoids running into a null reference and also allows all assertions
to be evaluated together, providing a complete picture of all failures to
a user.

Is an attempt at #10264. This fixes the assertion issue, but not the actual issue. (h/t @uenoku).

Assisted-by: OpenCode:claude-sonnet-4-6

[fabianschuiki]

Results of circt-tests run for d3b7833 compared to results for 6273dde: no change to test results.

[uenoku]

LGTM for assertion queue. Note that the original test case of #10264 still fails (since it doesn't have an assertion).

I believe the root cause is that %4 is evaluated before %3 when referenced value %3 still being nullptr.

om.class @Domain(%in: !om.string)  -> (out: !om.string) {
  om.class.fields %in : !om.string
}
om.class @Foo_Class(%basepath: !om.frozenbasepath)  -> (test: i1) {
  %0 = om.constant "A" : !om.string
  %1 = om.object @Domain(%0) : (!om.string) -> !om.class.type<@Domain>
  %2 = om.object.field %1, [@out] : (!om.class.type<@Domain>) -> !om.string
  %3 = om.object @Domain(%2) : (!om.string) -> !om.class.type<@Domain>
  %4 = om.object.field %3, [@out] : (!om.class.type<@Domain>) -> !om.string
  %5 = om.constant "B" : !om.string
  %6 = om.prop.eq %4, %5 : !om.string
  om.class.fields %6 : i1
}

[seldridge]

Thanks for pointing out this doesn't fix it. 🫣

The real motivation was getting the assertions working for situations where these are lowered to intermediary objects.

Also, I'm fine with abandoning this (and reverting 6273dde) if this can be fully subsumed by your later PR.

[seldridge]

And by your "later" PR (#10265) I clearly mean "earlier".

[uenoku]

No even with #10265 it still ignores assertions that are not immediately evaluatable, so this PR (and logging) look great to me. #10265 should fix scheduling of nested references, but it still requires a bit more of unification between isFullyEvaluated and Pending state.