feat(r42): improve KB grounding accuracy and add clear-history

R42 was missing knowledge-graph information (e.g. pension questions)
because retrieval and context-building dropped relevant facts:

- retrieval: exact-token TF-IDF could not match Dutch compound words,
  so a "pensioen" query scored 0 against "pensioenregeling" /
  "partnerpensioen" and never retrieved them. Add a compound-word
  fallback (shared >=6-char stem or containment, 0.4x weight) alongside
  exact matching.
- rag: deep article content was only injected for verbatim-mentioned
  topics; retrieved topics contributed just a 200-char description.
  Inject ~1000 chars of content for up to 5 topics (mentions first,
  then top-ranked retrieved) and widen the description snippet to 320.
- prompts: add a NAUWKEURIGHEID block (use all relevant facts, call
  lookup_topic before giving up) and relax the 4-sentence cap for
  detail/list answers so complete facts aren't summarised away.

Also add a clear-history control: a trash button in the chat header
(confirm dialog) wipes chat🧵{userId} and reseeds the greeting
via clearThread() in useChat.

Tests: compound-word matching + rag deep-content injection. Spec updated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/lib/retrieval.js

@@ -63,6 +63,29 @@ export function buildIndex(topics) {
   return index;
 }
 
+// Compound-word matching. Dutch is heavily compounding, so a user's word
+// (`pensioenafspraken`) is a *different* token than the graph's labels
+// (`pensioenregeling`, `partnerpensioen`), even though they share the stem
+// `pensioen`. Exact TF-IDF scores those pairs at 0, so the relevant topics are
+// never retrieved. These heuristics recover that recall at a reduced weight,
+// so exact matches still dominate the ranking.
+const PARTIAL_MIN_QUERY_LEN = 6; // only expand meaty query tokens
+const PARTIAL_MIN_OVERLAP = 6;   // shared stem / substring must be this long
+const PARTIAL_WEIGHT = 0.4;      // discount vs. an exact term hit
+
+/** True when two distinct tokens share a long stem or one contains the other. */
+function partialMatch(q, d) {
+  if (q === d) return false;
+  const shorter = q.length <= d.length ? q : d;
+  const longer = q.length <= d.length ? d : q;
+  if (shorter.length < PARTIAL_MIN_OVERLAP) return false;
+  if (longer.includes(shorter)) return true;
+  let n = 0;
+  const m = shorter.length;
+  while (n < m && q[n] === d[n]) n++;
+  return n >= PARTIAL_MIN_OVERLAP;
+}
+
 export function retrieveTopK(index, query, k = 10) {
   if (!index || !index.N || !query) return [];
   const qTokens = tokenize(query);
@@ -80,8 +103,19 @@ export function retrieveTopK(index, query, k = 10) {
     let s = 0;
     for (const t of qTokens) {
       const f = tf.get(t);
-      if (!f) continue;
-      s += (1 + Math.log(f)) * idf(t);
+      if (f) {
+        s += (1 + Math.log(f)) * idf(t);
+        continue;
+      }
+      // No exact hit — try a compound-word match against this doc's terms.
+      if (t.length < PARTIAL_MIN_QUERY_LEN) continue;
+      let best = 0;
+      for (const [term, tf2] of tf) {
+        if (!partialMatch(t, term)) continue;
+        const w = PARTIAL_WEIGHT * (1 + Math.log(tf2)) * idf(term);
+        if (w > best) best = w;
+      }
+      s += best;
     }
     scores[i] = s;
   }