fix(insights): screen the medication-adherence card and background lines uniformly

Route the medication-compliance status card through finalizeStatusSummary so
it enforces the same dose / risk / causal outbound contracts its seven sibling
status cards do, not only the unparseable-envelope withhold it already had. A
dose-change imperative or a fabricated figure is now withheld to the controlled
fallback on this surface too.

Add the status cards' structured-envelope detection to the two background hero
lines (arrival reaction, coach nudge): a brace-led or fenced JSON reply, or a
truncated envelope under the char cap, now resolves to the deterministic lead
rather than rendering raw.

Fold in the residuals from the same sweep: strip a truncated SUGGEST-ACTION
marker from the Coach prose tail, drop an unreachable zero-rate branch in the
adherence fold, and document why the internal conversation summary stays off
the render-boundary screen.

src/lib/ai/coach/conversation-summary.ts

@@ -205,6 +205,16 @@ export async function refreshConversationSummary(
   if (trimmed.length === 0) {
     return { status: "skipped" };
   }
+  // v1.32.20 — deliberately NOT routed through `screenModelOutput`. This
+  // summary is internal memory only: it is stored encrypted and read back
+  // solely as `priorSummary` to prime the NEXT summary/chat prompt
+  // (`persistence.ts` decrypt → `chat/route.ts` priorSummary). It is never
+  // rendered to a user surface. Any dose/risk/causal phrasing it might carry
+  // cannot reach a person here, and the next Coach reply that could echo it is
+  // itself screened (`screenCoachReply` in the chat route) before it ships.
+  // Screening here would only risk withholding a legitimate internal summary
+  // (memory loss) with no user-facing safety gain — so the screen stays on the
+  // rendering boundaries, not this internal-memory write.
   // Backstop the stored length even if the provider over-ran the prompt's
   // "short paragraph" instruction.
   const text =

src/lib/ai/coach/suggest-action.ts

@@ -213,11 +213,35 @@ function resolveActionParams(
  * invalid param, drops the action (`malformed: true`) while still removing the
  * raw marker from the prose.
  */
+/**
+ * Strip a TRUNCATED (partial) open sentinel left at the tail of the model's
+ * output. A reply cut off mid-marker — e.g. "…worth a look.\n---SUGGEST-ACTIO"
+ * — never contains the full `OPEN_SENTINEL`, so the block strip below misses it
+ * and the fragment renders raw (the #608 malformed-output class). Remove any
+ * trailing run that is a non-empty proper prefix of the open sentinel of at
+ * least `---S`, so a legitimate bare "---" separator is left untouched.
+ */
+function stripTruncatedOpenSentinel(prose: string): string {
+  const trimmedEnd = prose.replace(/\s+$/u, "");
+  for (let len = OPEN_SENTINEL.length - 1; len >= 4; len--) {
+    const partial = OPEN_SENTINEL.slice(0, len);
+    if (trimmedEnd.endsWith(partial)) {
+      return trimmedEnd.slice(0, trimmedEnd.length - partial.length).trimEnd();
+    }
+  }
+  return prose;
+}
+
 export function parseSuggestAction(raw: string): SuggestActionParseResult {
   if (!raw) return { prose: "", action: null, malformed: false };
 
   const openIdx = raw.indexOf(OPEN_SENTINEL);
-  if (openIdx === -1) return { prose: raw, action: null, malformed: false };
+  if (openIdx === -1) {
+    // No full block — but a truncated open marker at the tail must still be
+    // scrubbed so the raw fragment never reaches the user's prose.
+    const prose = stripTruncatedOpenSentinel(raw);
+    return { prose, action: null, malformed: prose !== raw };
+  }
 
   const before = raw.slice(0, openIdx);
   const afterOpen = raw.slice(openIdx + OPEN_SENTINEL.length);