rivet vadlidate and rivet validate --direct show different results

[avrabe]

some with errors in my example and the other without.

[avrabe]

Triage from the scheduled routine.

This is the same shape as #627 (which the routine is also working today): rivet validate and rivet validate --direct share library code but disagree on which links close which rules. #627 pins one specific divergence (self-satisfying links count as closing a rule in --direct but not in coverage); your "some with errors, some without" sounds like the more general statement of the same class of bug.

To pick this one up as its own focused PR rather than folding it into #627, the routine needs a single reproducible test case the fix can pin. Suggested AC shape:

## Acceptance Criteria
- [ ] Minimal repro: a fixture project (`tests/fixtures/validate-direct-divergence/`) where `rivet validate` and `rivet validate --direct` produce different diagnostic sets — fixture is checked in with a README naming the artifact + the disagreeing diagnostic.
- [ ] Root cause identified — which phase (lifecycle / structural / traceability) of which path emits the diagnostic the other doesn't, named in the PR body with a code pointer.
- [ ] Regression test in `rivet-core` asserts the two paths emit identical diagnostic sets on the fixture (sorted by `(rule, artifact_id, severity)` before comparing).
- [ ] On rivet's own artifacts: `rivet validate` ≡ `rivet validate --direct` diagnostic-by-diagnostic.
- [ ] **Out of scope:** the specific self-link case (already covered by #627 / #631); any *intended* divergences must be explicitly listed in the PR.

If your project is small enough to share, pasting the disagreeing diagnostic pair (the one each path emits that the other doesn't) into this issue would be enough for the routine to convert that into the fixture above. Refs #627, #631.

Not implemented this run — #627 is the only issue with explicit testable AC and the routine is already working it.

---

Generated by Claude Code