fix(voice+recall): KI-214 + KI-217 + KI-218 bundle

KI-214 — Block recognition for ENTIRE request-in-flight window.
Previously dropResultsRef only blocked during <audio> playback (KI-203),
leaving a 1-3s gap between send() firing and bot audio rendering+playing
where mic captured bot voice from speakers. Now onresult also early-
returns when isTextRequestPendingRef is true. Eliminates the entire
echo-leak window from user submit until bot finishes speaking.

KI-217 — Fix first-phrase cutoff in utterance batching.
User showed "My name is Rohit Sarma" reduced to "last name" — start
of utterance dropped. Root cause: rec.onend reset finalsRef.current=[]
on every restart cycle. Chrome occasionally delivers buffered
isFinal=true events for the early speech AFTER onend (same dirty-
window quirk that motivated KI-203). Eager wipe meant late finals
landed in a freshly emptied ref, then got wiped again before submit.

Fix: introduced finalsConsumedRef monotonic cursor. onend reads
finalsRef.current.slice(consumed), appends new finals to pending,
bumps cursor. finalsRef itself only resets at the actual submit
boundary inside submitPendingUtterance + on user-toggled lifecycle
(start/stop). Late-delivered finals from the dying recognition
instance now stick around to be picked up on the next onend.

KI-218 — Richer welcome-back profile summary.
KI-196's "first buy" was sparse because the prior profile genuinely
only had primary_goal captured. Generalized the formatter to render
every captured slot in natural English: "age 29, single, metro city,
income ₹25L+, first-time buyer, no pre-existing conditions, no
existing cover". When the stored profile has fewer slots, the
formatter renders what's there. Module-level label dicts map enums
to readable phrases (dependents/income/location/goal/budget).

VERIFICATION:
py_compile clean (orchestrator).
npx tsc --noEmit clean (frontend).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/orchestrator.py

@@ -383,6 +383,111 @@ def _format_known_profile_summary(profile) -> str:
     return summary
 
 
+# KI-218 (2026-05-15) — readable welcome-back recall summary.
+# The original KI-196 implementation joined raw enum values with commas
+# ("first buy" alone). This formatter enumerates every captured slot in
+# natural English so the user can recognise their stored profile at a glance.
+_DEPENDENTS_LABEL = {
+    "self": "single",
+    "self+spouse": "with spouse",
+    "self+spouse+kids": "with spouse and kids",
+    "self+parents": "with parents",
+    "self+spouse+kids+parents": "with spouse, kids and parents",
+}
+_INCOME_LABEL = {
+    "under_5L": "income under ₹5L",
+    "5L-10L": "income ₹5-10L",
+    "10L-25L": "income ₹10-25L",
+    "25L+": "income ₹25L+",
+}
+_LOCATION_LABEL = {
+    "metro": "metro city",
+    "tier1": "tier-1 city",
+    "tier2": "tier-2 city",
+    "tier3": "tier-3 city",
+}
+_GOAL_LABEL = {
+    "first_buy": "first-time buyer",
+    "upgrade": "upgrading existing cover",
+    "compare_specific": "comparing specific policies",
+    "tax_planning": "tax-planning",
+}
+_BUDGET_LABEL = {
+    "under_15k": "budget under ₹15k",
+    "15k_30k": "budget ₹15-30k",
+    "30k_60k": "budget ₹30-60k",
+    "60k+": "budget ₹60k+",
+}
+
+
+def _format_recall_summary_phrase(summary: dict) -> str:
+    """Build a natural-English phrase enumerating captured profile slots.
+
+    Skips slots that are None / unset / empty. Maps enum values to
+    user-readable words (no JSON, no underscores). Returns a comma-joined
+    string suitable to slot into the welcome-back ask.
+
+    See KI-218 — the prior implementation showed only raw enum tokens for
+    the 4 of 8 slots it touched, leaving the user unable to recognise the
+    stored profile.
+    """
+    bits: list[str] = []
+
+    age = summary.get("age")
+    if age is not None:
+        bits.append(f"age {age}")
+
+    dependents = summary.get("dependents")
+    if dependents:
+        bits.append(_DEPENDENTS_LABEL.get(dependents, str(dependents).replace("_", " ")))
+
+    location = summary.get("location_tier")
+    if location:
+        bits.append(_LOCATION_LABEL.get(location, str(location)))
+
+    income = summary.get("income_band")
+    if income:
+        bits.append(_INCOME_LABEL.get(income, f"income {income}"))
+
+    goal = summary.get("primary_goal")
+    if goal:
+        bits.append(_GOAL_LABEL.get(goal, str(goal).replace("_", " ")))
+
+    # health_conditions: [] is a meaningful capture ("no PEDs") but the
+    # orchestrator's pre-filter drops empty lists, so [] won't reach us.
+    # We still handle it defensively in case future call sites pass it through.
+    if "health_conditions" in summary:
+        hc = summary.get("health_conditions")
+        if isinstance(hc, list):
+            if len(hc) == 0:
+                bits.append("no pre-existing conditions")
+            else:
+                bits.append("conditions: " + ", ".join(str(c) for c in hc))
+        elif hc:
+            bits.append(f"conditions: {hc}")
+
+    existing = summary.get("existing_cover_inr")
+    if existing is not None:
+        try:
+            n = int(existing)
+            if n == 0:
+                bits.append("no existing cover")
+            elif n >= 100000:
+                lakhs = n / 100000
+                lakh_str = f"{lakhs:.0f}" if lakhs.is_integer() else f"{lakhs:.1f}"
+                bits.append(f"₹{lakh_str}L existing cover")
+            else:
+                bits.append(f"₹{n:,} existing cover")
+        except (TypeError, ValueError):
+            bits.append(f"existing cover {existing}")
+
+    budget = summary.get("budget_band")
+    if budget:
+        bits.append(_BUDGET_LABEL.get(budget, f"budget {budget}"))
+
+    return ", ".join(bits) if bits else "your earlier captures"
+
+
 _HELPFUL_GAP_LABELS = {
     "age": "your age",
     "dependents": "who you're covering",
@@ -820,16 +925,9 @@ async def handle_turn(
                         # the user should see right now. Subsequent turns
                         # WILL go through the brain with the recall directive
                         # in the system prompt (see _build_system_prompt).
-                        _bits = []
-                        if summary.get("age") is not None:
-                            _bits.append(f"age {summary['age']}")
-                        if summary.get("location_tier"):
-                            _bits.append(str(summary["location_tier"]))
-                        if summary.get("dependents"):
-                            _bits.append(str(summary["dependents"]))
-                        if summary.get("primary_goal"):
-                            _bits.append(str(summary["primary_goal"]).replace("_", " "))
-                        _summary_phrase = ", ".join(_bits) if _bits else "your earlier captures"
+                        # KI-218 — rich, all-8-slot natural-English summary
+                        # so user can recognise their stored profile.
+                        _summary_phrase = _format_recall_summary_phrase(summary)
                         sb_result.reply_text = (
                             f"Welcome back, {session.profile.name} — I have a profile under your name "
                             f"from before: {_summary_phrase}. Continue from there or start fresh?"

frontend/src/lib/useStreamingVoice.ts

@@ -193,6 +193,15 @@ export function useStreamingVoice(
 
   const recognitionRef = useRef<SpeechRecognitionInstance | null>(null);
   const finalsRef = useRef<string[]>([]);
+  // KI-217 (2026-05-15) — track how many entries of finalsRef have already
+  // been drained to pendingUtteranceRef. Each onend reads the slice from
+  // `finalsConsumedRef.current` to end, then bumps the cursor. finalsRef
+  // itself is NOT reset between restart cycles — only after the grace-timer
+  // submit (when onFinalRef fires) or on user-toggled start/stop. This
+  // prevents a Chrome quirk where late-delivered isFinal results arriving
+  // after onend on a mid-utterance restart cycle would land in a freshly
+  // wiped finalsRef and get dropped on the NEXT onend cycle's drain.
+  const finalsConsumedRef = useRef<number>(0);
   const wantRunningRef = useRef(false); // mirrors `enabled` for handler closures
   const restartTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
   const errorBackoffRef = useRef(0);
@@ -410,8 +419,11 @@ export function useStreamingVoice(
       // TTS audio ("perfect days to get started Rohit") was leaking into
       // the user input field between `audio.play()` firing and our abort()
       // actually taking effect.
-      if (dropResultsRef.current) {
-        console.debug("[useStreamingVoice] KI-203 dropping recognition result during/after TTS");
+      if (dropResultsRef.current || isTextRequestPendingRef.current) {
+        console.debug("[useStreamingVoice] KI-203/214 dropping recognition result", {
+          drop: dropResultsRef.current,
+          textPending: isTextRequestPendingRef.current,
+        });
         return;
       }
       let interim = "";
@@ -460,8 +472,14 @@ export function useStreamingVoice(
 
     rec.onend = () => {
       onListeningRef.current(false);
-      const webSpeechText = finalsRef.current.join(" ").trim();
-      finalsRef.current = [];
+      // KI-217 — drain only the NEW finals (everything past the consumed
+      // cursor). DO NOT reset finalsRef here: a late-delivered isFinal
+      // chunk arriving after onend would otherwise be wiped before the
+      // next onend cycle can pick it up. finalsRef is reset on actual
+      // utterance submit (grace-timer flush) and on user start/stop.
+      const newFinals = finalsRef.current.slice(finalsConsumedRef.current);
+      const webSpeechText = newFinals.join(" ").trim();
+      finalsConsumedRef.current = finalsRef.current.length;
 
       // KI-168 PHASE 2 — race guard: if a typed-text turn is in flight,
       // drop both transcripts on the floor (text wins). Don't start a
@@ -576,6 +594,11 @@ export function useStreamingVoice(
         const accumulatedChunks = pendingChunksRef.current;
         pendingUtteranceRef.current = "";
         pendingChunksRef.current = [];
+        // KI-217 — the utterance is now being submitted; safe to wipe
+        // finalsRef + reset the consumed cursor. Any late results that
+        // arrive after this point are for a NEW utterance.
+        finalsRef.current = [];
+        finalsConsumedRef.current = 0;
         console.debug("[useStreamingVoice] KI-202 grace window elapsed — submitting", {
           textLen: accumulatedText.length,
           chunkCount: accumulatedChunks.length,
@@ -694,6 +717,7 @@ export function useStreamingVoice(
       recognitionRef.current = buildRecognition();
     }
     finalsRef.current = [];
+    finalsConsumedRef.current = 0;
     // Kick off audio capture in parallel with recognition. If it fails we
     // degrade to Web Speech-only — onend handles the fallback path.
     void ensureAudioCapture();
@@ -713,6 +737,7 @@ export function useStreamingVoice(
     }
     teardownAudio();
     finalsRef.current = [];
+    finalsConsumedRef.current = 0;
     // KI-202 — drop any pending utterance so toggling voice off mid-grace
     // doesn't auto-submit a stale half-sentence next time voice comes on.
     if (pendingSubmitTimerRef.current !== null) {