observer: add /relevance heuristic filter for adjacency pollution

Matrix retrieval often surfaces high-cosine chunks that are about
symbols the focus file IMPORTS but doesn't define. The reviewer LLM
then hallucinates those imported-crate internals as in-file content
("I see main.rs does X" when X lives in queryd::context).

mcp-server/relevance.ts — pure scorer with five signals:
  path_match      +1.0  chunk source/doc_id encodes focus path
  defined_match   +0.6  chunk text mentions focus.defined_symbols
  token_overlap   +0.4  jaccard of non-stopword tokens
  prefix_match    +0.3  shared first-2-segment prefix
  import_only    -0.5  mentions only imported symbols (pollution)

Default threshold 0.3 — tuned empirically on the gateway/main.rs case.

Also fixes a regex bug in the import extractor: the character class
was lowercase-only, so `use catalogd::Registry;` silently never
matched (regex backed off when it hit the uppercase R). Caught by
the test suite.

observer.ts — POST /relevance endpoint wraps filterChunks().
scrum_master_pipeline.ts — fetchMatrixContext gains optional
focusContent param; calls /relevance after collecting allHits and
before sort+top. Opt-out via LH_RELEVANCE_FILTER=0; threshold via
LH_RELEVANCE_THRESHOLD. Fall-open on observer failure.

9 unit tests, all green. Live probe on real shape correctly drops
a 0.7-cosine adjacency-pollution chunk while keeping in-focus hits.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/real-world/scrum_master_pipeline.ts

@@ -499,6 +499,7 @@ async function fetchMatrixContext(
   filePath: string,
   topPerCorpus = 3,
   topOverall = 8,
+  focusContent?: string,
 ): Promise<MatrixContext> {
   const t0 = Date.now();
   const corpora = MATRIX_CORPORA_FOR_TASK[taskClass] ?? [];
@@ -548,9 +549,48 @@ async function fetchMatrixContext(
     errors.pathway_memory = e.message;
   }
 
+  // Optional adjacency-pollution filter via observer /relevance.
+  // Drops chunks that scored well on cosine but are about symbols the
+  // focus file only IMPORTS (not defines). Opt-out with
+  // LH_RELEVANCE_FILTER=0 for A/B comparison.
+  const relevanceEnabled = process.env.LH_RELEVANCE_FILTER !== "0";
+  let filteredHits = allHits;
+  let droppedCount = 0;
+  if (relevanceEnabled && allHits.length > 0) {
+    try {
+      const r = await fetch(`${OBSERVER_URL}/relevance`, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({
+          focus_file: { path: filePath, content: focusContent ?? "" },
+          chunks: allHits.map(h => ({
+            source: h.source_corpus,
+            doc_id: h.doc_id,
+            text: h.text,
+            score: h.score,
+          })),
+          threshold: Number(process.env.LH_RELEVANCE_THRESHOLD ?? 0.3),
+        }),
+        signal: AbortSignal.timeout(5000),
+      });
+      if (r.ok) {
+        const fr: any = await r.json();
+        const keptKeys = new Set((fr.kept ?? []).map((c: any) => `${c.source}|${c.doc_id}`));
+        filteredHits = allHits.filter(h => keptKeys.has(`${h.source_corpus}|${h.doc_id}`));
+        droppedCount = (fr.dropped ?? []).length;
+      } else {
+        errors.relevance_filter = `HTTP ${r.status}`;
+      }
+    } catch (e: any) {
+      // Fall-open: filter failure must not block the pipeline.
+      errors.relevance_filter = e.message;
+    }
+  }
+
   // Sort all hits by score desc, take top N
-  allHits.sort((a, b) => b.score - a.score);
-  const topHits = allHits.slice(0, topOverall);
+  filteredHits.sort((a, b) => b.score - a.score);
+  const topHits = filteredHits.slice(0, topOverall);
+  if (droppedCount > 0) byCorpus._relevance_dropped = droppedCount;
 
   return {
     hits: topHits,
@@ -1349,7 +1389,7 @@ Respond with markdown. Be specific, not generic. Cite file-region + PRD-chunk-of
     // source so retrieval anchors against both the metadata and the
     // actual code being reviewed.
     const matrixQuery = `${taskClass} ${rel} ${content.slice(0, 500)}`;
-    const matrixCtx = await fetchMatrixContext(matrixQuery, taskClass, rel, 3, 8);
+    const matrixCtx = await fetchMatrixContext(matrixQuery, taskClass, rel, 3, 8, content);
     provenApproachesPreamble = buildMatrixPreamble(matrixCtx);
     const corporaSummary = Object.entries(matrixCtx.by_corpus)
       .map(([k, v]) => `${k.split("_v")[0]}=${v}`).join(" ");