fix(agent): send_chat matcher requires single-word recipient

Bug: 'tell me a haiku about flowcharts' matched send_chat with
target='me a haiku' / body='flowcharts', producing the bogus
"I don't know who 'me a haiku' is" reply and silently swallowing
all non-tool prompts that started with tell/ask. The LLM fallback
never ran, so [llm] telemetry from wave 3 never fired in tests.

Fix: tighten the matcher to require a single-word [A-Za-z]+ name
between the verb and that/about/:. Drop the substring search in the
directory; require an exact match. Now 'tell me a haiku about X'
falls through to llmFallback as intended.

qa/_llm_telemetry_check.mjs

@@ -0,0 +1,42 @@
+// Isolated LLM telemetry check — uses a non-tool prompt + waits for the
+// 12s deadline so the timeout-or-error branch fires deterministically.
+import { chromium } from "playwright";
+
+const b = await chromium.launch({ headless: true });
+const ctx = await b.newContext({ viewport: { width: 1440, height: 900 } });
+const p = await ctx.newPage();
+const msgs = [];
+p.on("console", (m) => msgs.push({ type: m.type(), text: m.text() }));
+await ctx.addInitScript(() => localStorage.setItem("fm.canvas.tour.seen", String(Date.now())));
+
+await p.goto("https://canvas.flow-master.ai/", { waitUntil: "domcontentloaded" });
+await p.waitForTimeout(2500);
+await p.locator(".persona-chip").filter({ hasText: /Mariana/ }).click();
+await p.waitForSelector(".topbar", { timeout: 30000 });
+await p.waitForTimeout(2000);
+await p.locator(".tab").filter({ hasText: /Assistant/i }).first().click();
+await p.waitForSelector(".agent-composer textarea", { timeout: 15000 });
+await p.waitForTimeout(2000);
+
+// Three different non-tool prompts to make sure we hit the LLM fallback
+const prompts = [
+  "tell me a haiku",
+  "summarize what you can do for me",
+  "is the sky blue today",
+];
+for (const q of prompts) {
+  const ta = p.locator(".agent-composer textarea").first();
+  await ta.fill(q);
+  await ta.press("Enter");
+  await p.waitForTimeout(14000); // > 12s LLM deadline
+}
+
+const llm = msgs.filter((m) => m.text.includes("[llm]"));
+console.log("[llm] entries:", llm.length);
+for (const e of llm) console.log(`  [${e.type}] ${e.text.slice(0, 200)}`);
+
+const replies = await p.locator(".agent-turn-body").allTextContents();
+console.log("\nlast 3 replies:");
+for (const r of replies.slice(-3)) console.log("  -", r.slice(0, 160));
+
+await b.close();

qa/_wave3_verify.mjs

@@ -55,11 +55,13 @@ else fail("wave3_studio_draft_lands_via_split_post", `posts=${JSON.stringify(flo
 
 console.log("\n=== watch-out 2: LLM telemetry shows in console ===");
 await p.locator(".tab").filter({ hasText: /Assistant/i }).first().click();
-await p.waitForTimeout(3000);
+await p.waitForSelector(".agent-composer textarea", { timeout: 15000 });
+await p.waitForTimeout(2000);
 const txt = p.locator(".agent-composer textarea").first();
-await txt.fill("Briefly: what is EA2?");
+// Genuinely non-tool prompt so routeAgentInput falls through to llmFallback.
+await txt.fill("tell me a haiku about flowcharts");
 await txt.press("Enter");
-await p.waitForTimeout(10000);
+await p.waitForTimeout(14000);
 const llmConsoleEntries = consoleMsgs.filter(m =>
   m.type === "warning" || m.type === "info"
 ).filter(m => m.text.includes("[llm]"));