feat(coach): verify Coach prose numbers against fetched figures

The Daily Briefing strips any number absent from its server-computed
figures; the Coach's tool path had no equivalent, so a transcription or
paraphrase drift (a tool returns systolic 128, the reply says ~138) could
ship. Collect the numeric leaves from every present tool result this turn,
extract the numbers the reply asserts, and soft-correct any that match no
fetched figure to a neutral placeholder — annotated, non-blocking, and a
no-op on a qualitative turn or the no-tools path where there is nothing to
grade against. The bounded retrieval loop now returns the present results'
payloads so the check has an authoritative figure set.

src/app/api/insights/chat/__tests__/route-tool-mode.test.ts

@@ -154,6 +154,7 @@ const { buildCoachDataInventory, renderDataInventory, runCoachToolLoop } =
       totalTokens: 80,
       rounds: 2,
       toolTrace: [{ name: "get_metric_series", present: true }],
+      toolResults: [],
     })),
   }));
 vi.mock("@/lib/ai/coach/tools", () => ({
@@ -183,6 +184,9 @@ vi.mock("@/lib/validations/coach-prefs", () => ({
 }));
 
 const { appendMessage } = await import("@/lib/ai/coach/persistence");
+const { parseKeyValuesSentinel } = await import("@/lib/ai/coach/keyvalues");
+const { parseSuggestReminder } =
+  await import("@/lib/ai/coach/suggest-reminder");
 
 vi.mock("@/lib/sse/create-stream", () => ({
   createSseStream: (
@@ -260,6 +264,50 @@ describe("coach chat — tool-mode routing (F1)", () => {
     ).toEqual([{ name: "get_metric_series", present: true }]);
   });
 
+  it("soft-strips a prose number the tools never returned (P6)", async () => {
+    resolveProviderChain.mockResolvedValue([
+      { providerType: "anthropic", instance: {} },
+    ]);
+    // Echo the real prose through the sentinel + suggest parsers so the
+    // verifier sees the model's actual numbers (the default mocks return a
+    // fixed string).
+    const drift = "Your systolic averaged about 138 lately.";
+    (parseKeyValuesSentinel as ReturnType<typeof vi.fn>).mockReturnValue({
+      prose: drift,
+      keyValues: [],
+      malformed: false,
+      malformedEntries: [],
+    });
+    (parseSuggestReminder as ReturnType<typeof vi.fn>).mockReturnValue({
+      prose: drift,
+    });
+    // The tool returned systolic 128; the prose drifts to 138.
+    (runCoachToolLoop as ReturnType<typeof vi.fn>).mockImplementation(
+      async () => ({
+        result: {
+          content: "Your systolic averaged about 138 lately.",
+          tokensUsed: 80,
+          model: "m",
+        },
+        workingProviderType: "anthropic",
+        totalTokens: 80,
+        rounds: 2,
+        toolTrace: [{ name: "get_metric_series", present: true }],
+        toolResults: [
+          { present: true, data: { aggregate: { avgSys30: 128 } } },
+        ],
+      }),
+    );
+    await post(chatReq({ message: "How is my BP?" }));
+    const calls = (appendMessage as ReturnType<typeof vi.fn>).mock.calls;
+    const assistantCall = calls.find(
+      (c) => (c[0] as { role: string }).role === "assistant",
+    );
+    const content = (assistantCall?.[0] as { content: string }).content;
+    expect(content).toContain("[unverified]");
+    expect(content).not.toContain("138");
+  });
+
   it("falls back to the snapshot-stuffing path when a provider lacks tools", async () => {
     resolveProviderChain.mockResolvedValue([
       { providerType: "anthropic", instance: {} },

src/app/api/insights/chat/route.ts

@@ -84,6 +84,10 @@ import {
 } from "@/lib/ai/coach/tools";
 import type { AiMessage } from "@/lib/ai/types";
 import { parseKeyValuesSentinel } from "@/lib/ai/coach/keyvalues";
+import {
+  findUnverifiedCoachNumbers,
+  stripUnverifiedNumbers,
+} from "@/lib/ai/coach/coach-prose-grounding";
 import { parseSuggestReminder } from "@/lib/ai/coach/suggest-reminder";
 import { gateSuggestion } from "@/lib/ai/coach/suggest-gate";
 import {
@@ -518,6 +522,9 @@ Reply now as the assistant, in ${locale === "de" ? "German" : "English"}.`;
   let result: CompletionResult;
   let workingProviderType: string;
   let toolTrace: CoachToolTrace[] = [];
+  // v1.21.0 (P6) — the present tool-result payloads this turn, for the post-hoc
+  // prose number-verifier. Empty on the no-tools path.
+  let toolResultPayloads: unknown[] = [];
   let totalTokensSpent: number;
   try {
     if (toolMode) {
@@ -556,6 +563,7 @@ Reply now as the assistant, in ${locale === "de" ? "German" : "English"}. Fetch
       result = loop.result;
       workingProviderType = loop.workingProviderType;
       toolTrace = loop.toolTrace;
+      toolResultPayloads = (loop.toolResults ?? []).map((r) => r.data);
       totalTokensSpent = loop.totalTokens;
     } else {
       const fallback = await runRawCompletionWithFallback({
@@ -701,6 +709,38 @@ Reply now as the assistant, in ${locale === "de" ? "German" : "English"}. Fetch
     });
   }
 
+  // v1.21.0 (P6 / C2-5) — post-hoc numeric verifier on the Coach prose. On the
+  // tool path, cross-check every number the model cited against the figures the
+  // tools actually returned this turn; an unmatched number (transcription /
+  // paraphrase drift) is soft-stripped to "[unverified]" and annotated. Cheap,
+  // non-blocking, and a no-op when no tool returned figures (a qualitative turn
+  // or the no-tools path) — the prompt-level grounding rule remains the
+  // backstop there, exactly like the briefing's "no signals → skip". A blocked
+  // turn already carries canned fallback prose, so skip it.
+  if (!outbound.block && toolResultPayloads.length > 0) {
+    const unverified = findUnverifiedCoachNumbers(
+      replyText,
+      toolResultPayloads,
+    );
+    if (unverified.length > 0) {
+      const { prose: corrected, stripped } = stripUnverifiedNumbers(
+        replyText,
+        unverified,
+      );
+      replyText = corrected;
+      annotate({
+        action: { name: "coach.prose.number_unverified" },
+        meta: {
+          flagged: unverified.length,
+          stripped,
+          // No raw values — just the count + truncated tokens for ops triage.
+          tokens: unverified.slice(0, 6).map((u) => u.source),
+          promptVersion: PROMPT_VERSION,
+        },
+      });
+    }
+  }
+
   let surfacedSuggestion: CoachSuggestion | null = null;
   if (!outbound.block && suggestParse.cadence) {
     const decision = await gateSuggestion({