diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0f14a149a..89172e489 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,28 @@
 
 ## [Unreleased]
 
+## [1.32.20] — 2026-07-24
+
+The medication-adherence insight now passes through the same safety screen as
+every other insight card. It was the one card of eight that showed the model's
+text without running that screen first. It already held back a reply that came
+back broken or cut off, but a dose-change instruction or an invented figure
+could still reach it. It now runs the same check the other seven cards run, and
+anything that trips the check falls back to the plain, data-grounded line the
+card shows when no model is set up.
+
+- **The medication-adherence card is screened like the rest.** A reply that
+  reads as a dose change, or that cites a made-up risk figure, is held back and
+  the card shows its plain fallback. A normal grounded summary still shows as
+  before.
+- **Two background notification lines can no longer print a raw model reply.**
+  The "just in" reaction line and the proactive coach nudge now catch a reply
+  that arrived as structured data or got cut off partway, and use their fixed
+  wording instead of printing it as is.
+- **Smaller cleanups in the same area.** A stray marker left behind by a
+  cut-off coach reply is now stripped from the text, and an unreachable branch
+  in the adherence figure was removed.
+
 ## [1.32.19] — 2026-07-24
 
 After you logged a dose, the Today view could keep showing it as still due
diff --git a/docs/api/openapi.yaml b/docs/api/openapi.yaml
index 61272e806..0ddcc2d0a 100644
--- a/docs/api/openapi.yaml
+++ b/docs/api/openapi.yaml
@@ -1,7 +1,7 @@
 openapi: 3.1.0
 info:
   title: HealthLog API
-  version: 1.32.19
+  version: 1.32.20
   description: >-
     Self-hosted personal-health-tracking PWA — public API surface for the iOS native client and external ingest.
 
diff --git a/package.json b/package.json
index 394273a58..6da91b935 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "healthlog",
-  "version": "1.32.19",
+  "version": "1.32.20",
   "description": "Self-hosted personal-health-tracking PWA with Withings integration, AI insights, and doctor-report PDF export.",
   "license": "PolyForm-Noncommercial-1.0.0",
   "homepage": "https://healthlog.dev",
diff --git a/public/sw.js b/public/sw.js
index 555646c48..2aede105f 100644
--- a/public/sw.js
+++ b/public/sw.js
@@ -36,7 +36,7 @@ try {
 // v1.4.38.4 → v1.4.42. Do not hand-edit; bump `package.json` and rebuild.
 const CACHE_VERSION =
   (typeof self !== "undefined" && self.__APP_VERSION__) ||
-  /* @sw-version-fallback */ "v1.32.19";
+  /* @sw-version-fallback */ "v1.32.20";
 const STATIC_CACHE = `healthlog-static-${CACHE_VERSION}`;
 const PAGE_CACHE = `healthlog-pages-${CACHE_VERSION}`;
 // v1.18.6 — read-only data cache for a curated allowlist of safe GET `/api/*`
diff --git a/src/lib/ai/coach/conversation-summary.ts b/src/lib/ai/coach/conversation-summary.ts
index 10ea0620f..074fc7c9d 100644
--- a/src/lib/ai/coach/conversation-summary.ts
+++ b/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 =
diff --git a/src/lib/ai/coach/suggest-action.ts b/src/lib/ai/coach/suggest-action.ts
index e466b55c0..cc55d8f9d 100644
--- a/src/lib/ai/coach/suggest-action.ts
+++ b/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);
diff --git a/src/lib/insights/__tests__/medication-compliance-status.test.ts b/src/lib/insights/__tests__/medication-compliance-status.test.ts
index 61aea6cd8..1e3a850b0 100644
--- a/src/lib/insights/__tests__/medication-compliance-status.test.ts
+++ b/src/lib/insights/__tests__/medication-compliance-status.test.ts
@@ -26,6 +26,7 @@ vi.mock("@/lib/medication-category", () => ({
 import { prisma } from "@/lib/db";
 import { runStatusCompletion } from "@/lib/insights/status-provider";
 import { getMedicationCategories } from "@/lib/medication-category";
+import { getNoKeyMedicationComplianceStatusText } from "@/lib/insights/no-key-fallbacks";
 import { generateMedicationComplianceStatusForUser } from "../medication-compliance-status";
 
 const dayMs = 24 * 60 * 60 * 1000;
@@ -218,6 +219,100 @@ describe("generateMedicationComplianceStatusForUser — cache-read skips a stub"
   });
 });
 
+describe("generateMedicationComplianceStatusForUser — outbound screen (v1.32.20)", () => {
+  /**
+   * The medication-adherence card is the highest-leverage surface for the
+   * outbound screen — a dose-change imperative or a fabricated figure lands
+   * exactly here. v1.32.20 routes it through `finalizeStatusSummary`, so it now
+   * WITHHOLDS the same dose / risk / causal violations its seven sibling status
+   * cards do (previously it only withheld unparseable envelopes). WITHHOLD =
+   * serve the deterministic no-key line, persist no model text, `updatedAt`
+   * stays null; the per-medication placeholder rows still render.
+   */
+  function setupCard(now: Date) {
+    vi.mocked(prisma.auditLog.findFirst).mockResolvedValue(null);
+    vi.mocked(prisma.medication.findMany).mockResolvedValue([
+      medFixture(now),
+    ] as never);
+    vi.mocked(prisma.medicationIntakeEvent.findMany).mockResolvedValue(
+      [] as never,
+    );
+    vi.mocked(prisma.auditLog.create).mockResolvedValue({
+      createdAt: now,
+    } as never);
+  }
+
+  it("withholds a dose-change imperative instead of rendering the raw prose", async () => {
+    const now = new Date();
+    setupCard(now);
+    stubCompletion('{"summary":"Increase your dose to 10 mg next week."}');
+
+    const result = await generateMedicationComplianceStatusForUser("user-1", {
+      locale: "en",
+    });
+
+    expect(result.summary).toBe(getNoKeyMedicationComplianceStatusText("en"));
+    expect(result.summary).not.toContain("10 mg");
+    expect(result.summary).not.toContain("Increase");
+    expect(result.updatedAt).toBeNull();
+    // The per-medication placeholder rows still surface for the active med.
+    expect(result.medications).toHaveLength(1);
+    // No cache row ever persisted the screened model text.
+    for (const call of vi.mocked(prisma.auditLog.create).mock.calls) {
+      const details = JSON.parse(
+        (call[0] as { data: { details: string } }).data.details,
+      ) as { summary?: string };
+      expect(details.summary ?? "").not.toContain("10 mg");
+    }
+  });
+
+  it("withholds a fabricated risk figure instead of rendering the raw prose", async () => {
+    const now = new Date();
+    setupCard(now);
+    stubCompletion(
+      '{"summary":"Your 10-year cardiovascular risk is about 14%."}',
+    );
+
+    const result = await generateMedicationComplianceStatusForUser("user-1", {
+      locale: "en",
+    });
+
+    expect(result.summary).toBe(getNoKeyMedicationComplianceStatusText("en"));
+    expect(result.summary).not.toContain("14%");
+    expect(result.updatedAt).toBeNull();
+    expect(result.medications).toHaveLength(1);
+  });
+
+  it("still renders a clean grounded compliance summary", async () => {
+    const now = new Date();
+    setupCard(now);
+    stubCompletion(
+      '{"summary":"Adherence held steady at 96% across the last 30 days."}',
+    );
+
+    const result = await generateMedicationComplianceStatusForUser("user-1", {
+      locale: "en",
+    });
+
+    expect(result.summary).toBe(
+      "Adherence held steady at 96% across the last 30 days.",
+    );
+    expect(result.updatedAt).not.toBeNull();
+    expect(result.cached).toBe(false);
+    // The clean assessment was persisted (not the negative stub).
+    const persisted = vi
+      .mocked(prisma.auditLog.create)
+      .mock.calls.map(
+        (call) =>
+          JSON.parse(
+            (call[0] as { data: { details: string } }).data.details,
+          ) as { summary?: string },
+      )
+      .find((details) => (details.summary ?? "").includes("96%"));
+    expect(persisted).toBeTruthy();
+  });
+});
+
 describe("generateMedicationComplianceStatusForUser — token-leak hardening (v1.4.27 F16)", () => {
   it("strips metric: tokens out of the cached summary", async () => {
     const now = new Date();
diff --git a/src/lib/insights/general-status.ts b/src/lib/insights/general-status.ts
index 50791e585..e294ed998 100644
--- a/src/lib/insights/general-status.ts
+++ b/src/lib/insights/general-status.ts
@@ -284,7 +284,12 @@ export async function prepareGeneralStatusForUser(
     .sort(([a], [b]) => a.localeCompare(b))
     .map(([day, value]) => ({
       measuredAt: new Date(`${day}T12:00:00.000Z`),
-      value: value.total > 0 ? round((value.taken / value.total) * 100, 1) : 0,
+      // A map entry is only ever created after `bucket.total += 1`, so
+      // `value.total` is always >= 1 here — the former `: 0` guard branch was
+      // unreachable. v1.32.20 removed it: a real zero-dose day never reaches
+      // this fold, so a synthesised `0` would have been a fabricated rate, not
+      // an honest absence (the #582 silent-wrong-absence class).
+      value: round((value.taken / value.total) * 100, 1),
     }));
   const adherenceSeries = applyPayloadBudget(adherenceRecords, {
     now,
diff --git a/src/lib/insights/medication-compliance-status.ts b/src/lib/insights/medication-compliance-status.ts
index 5fc2ba088..f7528dc2e 100644
--- a/src/lib/insights/medication-compliance-status.ts
+++ b/src/lib/insights/medication-compliance-status.ts
@@ -30,7 +30,7 @@ import { buildMetricSignal } from "@/lib/insights/metric-signal";
 import { degradeStatusSnapshotToBudget } from "@/lib/insights/graded-series";
 import {
   type SupportedLocale,
-  extractAssessmentSummary,
+  finalizeStatusSummary,
   normalizeLocale,
   normalizeSummaryText,
   round,
@@ -577,15 +577,22 @@ export async function prepareMedicationComplianceStatusForUser(
       // does not emit per-medication text. The per-medication cards carry a
       // placeholder so the UI surfaces a row per active medication; the
       // overall `summary` is the model-authored assessment.
-      // A malformed / truncated / summary-less structured envelope must never
-      // persist as the day's assessment (the raw-JSON class of bug). WITHHOLD:
+      //
+      // v1.32.20 — route the completion through the shared status chokepoint,
+      // so this card enforces the exact contracts its seven siblings do rather
+      // than only withholding unparseable envelopes. `finalizeStatusSummary`
+      // parses the `{ summary }` envelope, scrubs chart tokens + whitespace,
+      // then SCREENS the prose against `INSIGHTS_CONTRACTS` (dose / risk /
+      // causal). This is the highest-leverage surface for the outbound screen —
+      // a medication card is exactly where a dose-change imperative or a
+      // fabricated figure lands. WITHHOLD on any block or unparseable envelope:
       // serve the deterministic no-key line, write a short negative stub, and
       // persist no model text. The per-medication rows still render.
-      const extracted = extractAssessmentSummary(outcome.content);
-      if (extracted.kind === "unparseable") {
+      const finalized = finalizeStatusSummary(outcome.content, locale);
+      if (!finalized.ok) {
         annotate({
           action: { name: "insights.status.outbound_blocked" },
-          meta: { cacheAction, reason: "unparseable_output" },
+          meta: { cacheAction, reason: finalized.reason },
         });
         returnTimeoutFallback({
           cacheAction,
@@ -602,7 +609,7 @@ export async function prepareMedicationComplianceStatusForUser(
           updatedAt: null,
         };
       }
-      const summary = normalizeSummaryText(extracted.text);
+      const summary = finalized.text;
       if (!summary) {
         throw new Error(
           "Medication-compliance-status summary was empty after normalization",
diff --git a/src/lib/jobs/__tests__/hero-line-json-envelope.test.ts b/src/lib/jobs/__tests__/hero-line-json-envelope.test.ts
new file mode 100644
index 000000000..1204fc953
--- /dev/null
+++ b/src/lib/jobs/__tests__/hero-line-json-envelope.test.ts
@@ -0,0 +1,67 @@
+import { describe, expect, it, vi } from "vitest";
+
+// The sanitisers are pure, but their modules import the db client at load time.
+// Stub it so the unit under test never reaches for a connection.
+vi.mock("@/lib/db", () => ({ prisma: {} }));
+
+import { sanitiseReactionLine } from "@/lib/jobs/reaction-line";
+import { sanitiseAiBody } from "@/lib/jobs/coach-nudge-ai";
+
+/**
+ * v1.32.20 — both background hero lines (the arrival reaction and the coach
+ * nudge body) now run the status cards' structured-envelope detection
+ * (`extractAssessmentSummary`). A hero line is one plain sentence; a brace-led
+ * or fenced `{"line":"…"}` reply, or a truncated envelope under the char cap,
+ * is a broken contract and must resolve to the deterministic lead/template
+ * rather than render raw JSON. Genuine prose still ships.
+ */
+describe("sanitiseReactionLine — JSON-envelope detection", () => {
+  it("rejects a well-formed JSON envelope so the deterministic lead stands", () => {
+    expect(
+      sanitiseReactionLine(
+        '{"line":"Nice work today, your numbers are holding steady."}',
+        "en",
+      ),
+    ).toBeNull();
+  });
+
+  it("rejects a truncated brace-led envelope", () => {
+    expect(sanitiseReactionLine('{"line":"Nice work today', "en")).toBeNull();
+  });
+
+  it("rejects a fenced JSON envelope", () => {
+    expect(
+      sanitiseReactionLine('```json\n{"line":"Steady week."}\n```', "en"),
+    ).toBeNull();
+  });
+
+  it("still renders a plain-text line verbatim", () => {
+    expect(
+      sanitiseReactionLine(
+        "Nice work today, your numbers are holding steady.",
+        "en",
+      ),
+    ).toBe("Nice work today, your numbers are holding steady.");
+  });
+});
+
+describe("sanitiseAiBody — JSON-envelope detection", () => {
+  it("rejects a well-formed JSON envelope so the template stands", () => {
+    expect(
+      sanitiseAiBody(
+        '{"body":"A gentle check-in on your routine this week."}',
+        "en",
+      ),
+    ).toBeNull();
+  });
+
+  it("rejects a truncated brace-led envelope", () => {
+    expect(sanitiseAiBody('{"body":"A gentle check-in', "en")).toBeNull();
+  });
+
+  it("still renders a plain-text body verbatim", () => {
+    expect(
+      sanitiseAiBody("A gentle check-in on your routine this week.", "en"),
+    ).toBe("A gentle check-in on your routine this week.");
+  });
+});
diff --git a/src/lib/jobs/coach-nudge-ai.ts b/src/lib/jobs/coach-nudge-ai.ts
index f5c6e82b0..756ca1e02 100644
--- a/src/lib/jobs/coach-nudge-ai.ts
+++ b/src/lib/jobs/coach-nudge-ai.ts
@@ -48,6 +48,7 @@ import {
 import { AI_BUDGETS } from "@/lib/ai/ai-budgets";
 import { singleUserTurn } from "@/lib/ai/types";
 import { screenCoachReply } from "@/lib/ai/coach/outbound-guard";
+import { extractAssessmentSummary } from "@/lib/insights/status-shared";
 import { annotate } from "@/lib/logging/context";
 
 /** Per-call upstream timeout — kept inside the spec's ≤8–10 s window. */
@@ -138,9 +139,16 @@ Rephrase it warmly and naturally in your own words, following every rule. Return
  * unusable (empty / over-length / unsafe). Strips wrapping quotes and collapses
  * whitespace; rejects rather than mid-word-clamping an over-long reply.
  */
-function sanitiseAiBody(raw: string, locale: Locale): string | null {
+export function sanitiseAiBody(raw: string, locale: Locale): string | null {
   let text = (raw ?? "").trim();
   if (!text) return null;
+  // v1.32.20 — the same structured-envelope detection the status cards run
+  // (`extractAssessmentSummary`). A nudge body is one plain sentence; a
+  // brace-led or fenced `{"line":"…"}` / `{"body":"…"}` reply — or a truncated
+  // envelope under the char cap — is a broken contract and must resolve to the
+  // deterministic template rather than render raw JSON. Only genuine prose
+  // (`kind === "prose"`) passes.
+  if (extractAssessmentSummary(text).kind !== "prose") return null;
   // Strip a single layer of wrapping quotes the model sometimes adds.
   text = text
     .replace(/^["“”'`]+/, "")
diff --git a/src/lib/jobs/reaction-line.ts b/src/lib/jobs/reaction-line.ts
index b0bca6138..413c8fcc0 100644
--- a/src/lib/jobs/reaction-line.ts
+++ b/src/lib/jobs/reaction-line.ts
@@ -44,6 +44,7 @@ import {
 } from "@/lib/ai/coach/budget";
 import { AI_BUDGETS } from "@/lib/ai/ai-budgets";
 import { screenCoachReply } from "@/lib/ai/coach/outbound-guard";
+import { extractAssessmentSummary } from "@/lib/insights/status-shared";
 import { encryptToBytes } from "@/lib/ai/coach/bytes-codec";
 import { fenceUserText } from "@/lib/ai/coach/data-fence";
 import { singleUserTurn, type CompletionResult } from "@/lib/ai/types";
@@ -113,6 +114,14 @@ export function sanitiseReactionLine(
 ): string | null {
   let text = (raw ?? "").trim();
   if (!text) return null;
+  // v1.32.20 — the same structured-envelope detection the status cards run
+  // (`extractAssessmentSummary`). A hero line is one plain sentence; a
+  // brace-led or fenced `{"line":"…"}` reply — or a truncated envelope under
+  // the char cap — is a broken contract, not a sentence, and must resolve to
+  // the deterministic lead rather than render raw JSON. Only genuine prose
+  // (`kind === "prose"`) passes; a parsed `{ summary }`/`{ line }` object or an
+  // unparseable brace-led attempt both reject here.
+  if (extractAssessmentSummary(text).kind !== "prose") return null;
   text = text
     .replace(/^["“”'`]+/, "")
     .replace(/["“”'`]+$/, "")