Merge AI client + coach cache fixes into release/v1.20.0

Send the OpenAI JSON response format only when JSON is requested and no
tools are in play, so the Coach prose path (and the tool loop's final
round) is no longer corrupted into a JSON blob. Resolve the Anthropic
finish reason to tool_calls whenever tool_use blocks are present. Memoize
the per-request feature extraction and prefs read and project the feature
query, so the coach tool fan-out shares one read instead of six.

src/lib/insights/features.ts

@@ -681,6 +681,18 @@ export async function extractFeatures(
       ? { userId, measuredAt: { gte: sinceCutoff }, deletedAt: null }
       : { userId, deletedAt: null },
     orderBy: { measuredAt: "asc" },
+    // Project only the columns every downstream consumer reads (`byType`,
+    // `summarize`, BP pairing, and `reconstructSleepNights`'s `SleepStageRow`).
+    // The PULSE / glucose windows are 200k-row-class; pulling every column
+    // (notes, externalId, …) is pure wasted I/O on the shared Prisma pool.
+    select: {
+      type: true,
+      value: true,
+      measuredAt: true,
+      sleepStage: true,
+      source: true,
+      deviceType: true,
+    },
   });
 
   // v1.18.11 P1 — when the bulk read is bounded to a recent window, the