Danish statute + EUR-Lex RAG pipeline with clickable inline citations

The retrieval logic in lawRetrieval.ts goes beyond vanilla embedding search. It detects structured references like "§ 42, stk. 3" or "artikel 5 DORA" via regex (backed by a LAW_ALIASES map), runs a metadata lookup for that exact section in parallel with the cosine similarity search, then merges and deduplicates the results. That hybrid pattern - metadata fallback alongside embedding search - is the most reusable piece here, applicable to any domain with structured citation syntax.

Ingestion targets Danish primary-source statutes (hvidvaskloven, funktionærloven, selskabsloven) and EU regulations (DORA 32022R2554, PRIIPs 32014R1286), driven by env vars (RETSINFORMATION_LAWS, EURLEX_REGULATIONS). The backend migration adds a document_chunks table with an IVFFlat cosine index. There's a minor inconsistency: 001_document_chunks.sql creates the index with lists=100, then 002_reindex_chunks.sql immediately rebuilds it at lists=20 to reduce memory pressure.

Two other fixes come bundled in this commit set and are worth pulling on their own. normalizeCitation was silently dropping citations with page:null - the fix treats null as page=0. And the stop button was showing while the invisible <CITATIONS> SSE block was still streaming; showStopButton = isLoading && !isLoadingCitations fixes that so users can't accidentally cancel before citations arrive.

The law context is appended unconditionally to every chat turn, which will eat tokens on non-legal queries. The ingestion scripts call live endpoints with no visible rate-limiting.