BUG (blocked the supervised live re-import): /import/execute threw
Npgsql 23503 FK_Litters_Gerbils_FatherId and rolled back. Root cause was
INSERT ORDERING — synthesized derived litters were SaveChanges()'d BEFORE
the parent gerbils (created later in the animal loop), so the litter's
Father/MotherId pointed at rows that didn't exist yet. The dry-run and the
EF in-memory test provider don't enforce FKs, so it slipped through.
Fix:
- Stage synthesized litters in the context but DON'T save them early; the
single SaveChanges after the animal loop lets EF order parents → litters →
offspring (all FKs nullable). Saving litters first was the fault.
- FK-integrity guard (god's spec): compute the persisted set (existing DB +
this run's loadable) and null out any litter parent FK not in it; SKIP a
derived litter whose BOTH parents are unresolvable (offspring loads with
LitterId=null). Quarantined parents already resolve to null via
ResolveParentGid; this is defense-in-depth + makes the invariant explicit.
- Report litterParentFksDropped + derivedLittersSkipped (LitterSummary) +
a German note — so a green dry-run (0/0) GUARANTEES execute won't FK-fault.
Tests: two SQLite-backed regressions (SQLite enforces FKs, unlike the
in-memory provider) — a derived litter with NEW chart parents executes
without throwing, and a quarantined parent leaves that FK null. 120 C#
tests + python green; no schema change (has-pending clean).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Gulum
fd48f7fc09