[Aschenbath]

Closing a Review fact before its start currently stores an inverted interval and removes the item from review. Reject these closes with HTTP 422 interval_inverted, and show the localized error in all three Review cards.

• Check both endpoints at their own precision in close_superseded, after locking the live fact row and before any rewrite. This also covers a conflict closed at the new fact's start by default; close_stated and the ingestion callers stay unchanged.
• Add English and Chinese error text and error toasts for fact, conflict, and violation decisions, making empty_state_span refusals visible too.
• Exercise all three HTTP routes, default conflict closes, unchanged facts and open queues after refusals, and a successful retry with a later date normalized to its precision.

Validation on Linux x86_64 against local PostgreSQL 16 with pgvector and Poppler (UTOPIA_TEST_REQUIRE_DB=1, UTOPIA_TEST_REQUIRE_PDFTOTEXT=1), on dev 2bcc763:

• cargo fmt --all --check and cargo clippy --workspace --all-targets -- -D warnings: passed.
• cargo test --workspace: 1,622 passed, 0 failed, 6 existing opt-in tests ignored.
• cargo test --workspace --bin utopia-server review_closes_refuse: failed before the fix (route returned 200 {"ok":true}) and passed after it.
• web/: pnpm build passed, pnpm test 286/286 passed, node --test scripts/bench/lib.test.mjs 15/15 passed.

The commit is DCO signed off.

Closes #1294

[WaylandYang]

Thank you, Aschenbath, clean and well targeted. The check sits in close_superseded under the row lock, compares both ends after truncating them with the same truncate_to #1295 uses, and covers the conflict's default close at the new fact's start, while close_stated and the graph.rs callers stay untouched. The three cards now toast the localized refusal, which also brings today's silent empty_state_span to the surface, and the route test covers all three paths plus the default close, confirms nothing is written and the items stay in their queues, and that a later date still goes through.

Thanks also to @Maya-Kid for the report and the failing cases on #1294.

Merging.