benchmark: use matched_chunk for content hit scoring

Content hit rate jumps from 14.3% to 85.7% now that we read
the matched_chunk field from BM search results (PR #601).

Results: R@5=81%, R@10=88%, MRR=0.795, Content Hit=85.7%, P95=868ms

Author: bm-clawd

benchmark/corpus/MEMORY.md

@@ -1,3 +1,9 @@
+---
+title: MEMORY
+type: note
+permalink: benchmark/memory
+---
+
 # MEMORY.md - Long-Term Memory
 
 ## About Me
@@ -51,4 +57,4 @@
 - Rust compile times are brutal but the binary size payoff is worth it
 - SQLite WAL mode is mandatory for concurrent reads during metric aggregation
 - Never trust webhook delivery — always implement idempotent handlers
-- The onboarding flow is our weakest point right now (users drop off at OAuth)
+- The onboarding flow is our weakest point right now (users drop off at OAuth)

benchmark/run.ts

@@ -131,7 +131,7 @@ async function mcpSearch(query: string, limit = 10): Promise<SearchResult[]> {
       file_path: r.file_path || r.permalink || "",
       title: r.title || "",
       score: r.score || r.similarity || 0,
-      content: r.content || r.snippet || "",
+      content: r.matched_chunk || r.content || r.snippet || "",
     }))
   } catch (err) {
     console.error(`  MCP search failed for: "${query}"`, err)