EDI date rejection follow-ups from #3604 (durable evidence, full interchange, 944/945 TLC) #3623

Follow-ups from the CodeRabbit review of #3604. They were deferred because each one is broader than the fix for preserving the EDI event date.

• Durable storage for EDI rejection evidence (routes.py:341, fix(ingestion): keep real EDI event time instead of ingest time #3604 (comment)). _preflight_event_date returns a rejection_id for a record that lives only in the capped, process-local rejection store, so eviction or a restart loses it. This predates fix(ingestion): keep real EDI event time instead of ingest time #3604, since fsma_validation_failed uses the same store. It needs a DB-backed append-only adapter or table. Also add the rejection reason and the rejected date to the warning log.
• Keep the whole interchange when one transaction set's date is rejected (routes.py:1169, fix(ingestion): keep real EDI event time instead of ingest time #3604 (comment)). _preflight_all_sets stops at the first bad set, and the rejection record holds only that set's extracted fields. Keep the full source document, or every extracted set, in durable rejection evidence so the earlier valid CTEs can be recovered.
• 944/945 date rejections shouldn't carry the form TLC (routes.py:1173, fix(ingestion): keep real EDI event time instead of ingest time #3604 (comment)). Make warehouse date rejections document-level at the top level. Add the warehouse tlc_source_reference to the extracted snapshot before the preflight, and keep the line TLCs under extracted.line_items.