fix(voice+llm+admin): KI-200/201/202/203/204/205/206/207/208/210/211 bundle

THE LAST OPERATIONAL FIX BUNDLE — voice + recommendation routing + LLM
election + admin clarity. Docs cascade held for post-test confirmation.

KI-200 — Admin LLM Chain redesigned to 2 clean tables.
Table 1 "Currently In Use" (per use case + provider + tier).
Table 2 "All Eligible Models" with explicit Primary / 1st backup /
Last resort labels. Provider column color-coded (Google blue / NVIDIA
green / OpenRouter orange).

KI-201 — Election by chain order (not latency score).
backend/llm_health.py get_primary + get_backup now walk the chain
definition in priority order. Returns first eligible model. nemotron
only serves as last resort. Resolves "maverick elected over qwen
when all healthy" bug.

KI-202 — Utterance batching with 1.5s grace.
Mid-sentence pauses < 1.5s accumulate into one submission.
pendingUtteranceRef + pendingChunksRef + grace timer per onend.

KI-203 — Block recognition transcripts during TTS.
dropResultsRef gates onresult while audio plays + 300ms post-TTS
delay. Eliminates "flexibility" / "perfect days to get started"
echo into the input box.

KI-204 — Universal audio pause on send().
Every send() pauses + resets all <audio> elements before submitting.
Prior bot TTS shuts up the instant user starts a new turn.

KI-205 — Broadened recommendation-query detection.
Added "show me", "side by side", "three options", "shortlist",
"give me three", etc. Faithfulness gate no longer misfires on
natural recommendation phrasings.

KI-206 — Regulatory chunks excluded from chat citations + brain.
insurer_slug='regulatory' filtered from chunks BEFORE brain call
(for recommendation/comparison intents) AND from citations list.

KI-207 — Admin LLM Chain tab stripped to bare essentials.
Removed outer card wrapper + row-between flex wrapper + actions
wrapper. Only h2 + Refresh button + chains container remain.

KI-208 — Defensive regulatory filter on /api/policies/all pass-2.
Mirrors pass-1 KI-132 filter for the curated facts path.

KI-210 — Voice utterances NEVER dropped (wait-and-retry).
KI-202's 3 drop-on-text-racing sites converted to poll + 30s cap.
User speech captured during bot's thinking gap no longer silently
lost.

KI-211 — TTS volume default 0.6 → 0.3.
Quieter first-turn TTS so KI-189 barge-in VAD works pre-calibration.
Bot stays audible; mic-bleed minimized.

VERIFICATION:
py_compile clean (all backend files).
npx tsc --noEmit clean (frontend).
Election: chain order honored.
Voice path: no echo, no drop, barge-in capable from turn 1.

DOCS CASCADE: ADR-042 + CLAUDE.md + README.md stashed locally; will
land in a follow-up commit after live confirmation.

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

backend/llm_health.py

@@ -27,13 +27,16 @@ Records per model:
   - probe_history:    last PROBE_HISTORY_LEN (ok, latency_ms) tuples; powers
                       the success_rate signal in the election score.
 
-Election (KI-080):
-  - For each chain (brain / fast_brain / judge), compute a score per healthy
-    candidate: score = (1 / max(50, latency_ms)) * success_rate_last_5.
-  - CURRENT_PRIMARY = highest scorer.
-  - CURRENT_BACKUP  = highest scorer among candidates with a DIFFERENT
-    provider (NIM vs Groq vs OpenRouter) than primary; falls back to next-
-    best same-provider candidate when no cross-provider option qualifies.
+Election (KI-201, 2026-05-15 — supersedes KI-080/KI-087 score-based election):
+  - Walk the chain definition (BRAIN_CHAIN / FAST_BRAIN_CHAIN / JUDGE_CHAIN)
+    in priority order. CURRENT_PRIMARY = first election-eligible model.
+    CURRENT_BACKUP  = next election-eligible model after primary.
+  - Chain hierarchy IS the truth — nemotron-49b is LAST in BRAIN_CHAIN so
+    it only serves when qwen + mistral + maverick are all unavailable.
+  - Latency / success_rate still gate ELIGIBILITY (probe-fresh, not in
+    sin-bin, credits above water) but no longer drive ORDERING. Pre-KI-201
+    the (1/latency)×success_rate score let a chain-#3 model beat chain-#0
+    on a faster probe, violating the operator-defined hierarchy.
   - DEGRADED window: when chat() calls report_failure(model), that model is
     sidelined for either DEGRADED_WINDOW_SEC (transient, 30s) or
     DEGRADE_DURATION_LONG_S (rate-limit / HTTP 429, 1h — KI-084) so the
@@ -420,62 +423,59 @@ def _ranked_candidates(chain_name: str) -> list[ModelHealth]:
 
 
 def get_primary(chain_name: str) -> Optional[str]:
-    """Top-scoring election-eligible candidate, or None when no probe data
-    is fresh enough.
-
-    KI-087 (2026-05-15) — **NIM-first preference.** NIM is the strategic
-    free provider (ADR-019, single-key, $0 cost, 110+ models, no daily
-    cap). Groq has a hard 100K tokens/day free-tier cap and OpenRouter
-    charges a real USD balance. Both should serve as EMERGENCY fallback
-    only, not as primary just because their probe latency is lower.
-    Election therefore prefers ANY eligible NIM candidate over ALL
-    non-NIM candidates. Only when the NIM pool is empty (every NIM
-    model is down, throttled, or out of credits) does election fall
-    through to Groq/OpenRouter as primary. Within the NIM pool, the
-    standard latency × success_rate score still picks the fastest
-    healthy NIM model.
+    """KI-201 (2026-05-15) — elect by CHAIN ORDER, not latency score.
+
+    Walk the chain definition in priority order. Return the first model
+    that's election-eligible. Chain hierarchy IS the truth — e.g. in
+    BRAIN_CHAIN, nemotron-49b is LAST so it only serves when every
+    higher-priority model is unavailable (KI-175 last-resort rule).
+
+    Pre-KI-201 the elector picked by (1/latency) × success_rate, which
+    meant a chain-#3 model (maverick) could beat chain-#0 (qwen) on a
+    faster probe. Live evidence: all 4 NIM models healthy but elector
+    picked maverick over qwen. User-stated spec: "If a model is not
+    already in use, but at the backend everything is live, then it
+    should always suggest it according to our set hierarchy."
+
+    Latency is no longer used for primary/backup selection — eligibility
+    filtering still uses the rolling probe state, but ordering is
+    purely chain-positional.
     """
-    ranked = _ranked_candidates(chain_name)
-    if not ranked:
+    chain = _chain_for(chain_name)
+    if not chain:
         return None
-    # KI-087: NIM-first within the eligible set.
-    nim_pool = [h for h in ranked if provider_of(h.model) == "nim"]
-    if nim_pool:
-        return nim_pool[0].model
-    # No eligible NIM candidate — fall through to cross-provider primary.
-    return ranked[0].model
+    # _ranked_candidates returns the eligibility-filtered set (probe-fresh,
+    # not in sin-bin, credit-not-exhausted, healthy). Reduce to a set for
+    # O(1) lookup; we ignore its score-based ordering.
+    eligible_models = {h.model for h in _ranked_candidates(chain_name)}
+    for model in chain:
+        if model in eligible_models:
+            return model
+    return None  # nothing eligible
 
 
 def get_backup(chain_name: str) -> Optional[str]:
-    """Second-best election candidate. Prefers a DIFFERENT provider from
-    primary so a single provider's regional outage can't take out both.
-
-    KI-087 (2026-05-15) — when PRIMARY is NIM (the typical case under
-    NIM-first preference), BACKUP is the best NON-NIM candidate
-    (Groq / OpenRouter) — kept as the cross-provider safety net. When
-    NIM is wholly unavailable and PRIMARY is non-NIM, BACKUP is the
-    next-best non-NIM candidate or, failing that, the next-best
-    same-provider candidate.
+    """KI-201 (2026-05-15) — backup is the SECOND eligible model in chain
+    order, skipping the elected primary.
+
+    Walks the chain definition in priority order, skips the elected
+    primary, and returns the next eligible model. Same chain-as-truth
+    philosophy as get_primary: the chains in nvidia_nim_llm.py were
+    designed with family/provider diversity baked in (Qwen → Mistral →
+    Meta → NVIDIA), so walking past the primary in chain order naturally
+    preserves the cross-family backup invariant KI-087 used to enforce
+    via provider_of().
     """
-    ranked = _ranked_candidates(chain_name)
-    if len(ranked) < 2:
-        # No usable backup. Either zero candidates or only one (in which
-        # case the cold-start path in NimChainLLM falls back to chain[1]).
+    chain = _chain_for(chain_name)
+    if not chain:
         return None
+    eligible_models = {h.model for h in _ranked_candidates(chain_name)}
     primary = get_primary(chain_name)
-    if primary is None:
-        return None
-    primary_provider = provider_of(primary)
-    # Prefer cross-provider against the PRIMARY's provider.
-    for h in ranked:
-        if h.model == primary:
+    for model in chain:
+        if model == primary:
             continue
-        if provider_of(h.model) != primary_provider:
-            return h.model
-    # No cross-provider candidate — accept same-provider next-best.
-    for h in ranked:
-        if h.model != primary:
-            return h.model
+        if model in eligible_models:
+            return model
     return None
 
 

backend/main.py

@@ -1933,6 +1933,13 @@ async def policies_all(session_id: Optional[str] = None):
             continue
         seen_policy_ids.add(curated_policy_id)
         slug = data.get("insurer_slug", "")
+        # KI-208 (2026-05-15) — defensive symmetry with pass-1 (line 1842): any
+        # curated_facts entry with insurer_slug=='regulatory' must NOT surface
+        # as a marketplace card. Today no curated regulatory docs exist, but
+        # adding the filter here closes a future-leak vector if an operator
+        # accidentally curates an IRDAI/NHA fact-sheet under 40-data/policy_facts.
+        if slug == "regulatory":
+            continue
         name, home = insurer_meta.get(slug, (slug, ""))
         # Insurer reviews for scorecard
         ir = None

backend/orchestrator.py

@@ -998,6 +998,37 @@ async def handle_turn(
         profile_name_slug=profile_slug_for_retrieve,
         session_id=session_id,
     )
+
+    # KI-205 (2026-05-15) — detect recommendation-shaped queries. The original
+    # KI-171 keyword list missed phrases like "show me three options side by
+    # side", which routed a recommendation turn through the faithfulness gate
+    # and into a generic refusal. Expanded to cover side-by-side/shortlist/
+    # options-comparison shapes. Computed BEFORE the regulatory filter below
+    # (KI-206) so both the post-retrieve filter and the judge-skip can reuse it.
+    _user_text_lc = (user_text or "").lower()
+    _is_recommendation = any(
+        kw in _user_text_lc for kw in (
+            "recommend", "suggest", "best polic", "top 3", "top three",
+            "top 5", "top five", "show me polic", "show me what", "fits me",
+            "right for me", "policies for me", "which polic", "good polic",
+            "what polic", "your suggestion", "your recommendation",
+            # KI-205 (2026-05-15) — user said "show me show me three options
+            # side by side" → wasn't matched → faithfulness gate misfired.
+            "show me", "show me three", "show me few", "show me options",
+            "show me a few", "show me some", "side by side", "side-by-side",
+            "three options", "few options", "some options", "compare options",
+            "shortlist", "give me options", "give me three",
+        )
+    )
+
+    # KI-206 (2026-05-15) — regulatory chunks (IRDAI consolidated, etc.) are
+    # master regulatory text, not actual policies. They shouldn't appear in
+    # user-facing recommendation citations. Drop them post-retrieve for
+    # recommendation/comparison intent. Non-recommendation queries (e.g. "what
+    # does IRDAI say about waiting periods") still see regulatory evidence.
+    if _is_recommendation or intent == "comparison":
+        chunks = [c for c in chunks if (c.insurer_slug or "").lower() != "regulatory"]
+
     context_str = format_for_llm_context(chunks)
 
     # 3. Pick brain
@@ -1099,15 +1130,9 @@ async def handle_turn(
     # answers stitch together evidence from many policies plus the user's
     # profile — the judge can't grade that fairly and currently blocks valid
     # recommendation flows after fact_find completes. Detect via query shape.
-    _user_text_lc = (user_text or "").lower()
-    _is_recommendation = any(
-        kw in _user_text_lc for kw in (
-            "recommend", "suggest", "best polic", "top 3", "top three",
-            "top 5", "top five", "show me polic", "show me what", "fits me",
-            "right for me", "policies for me", "which polic", "good polic",
-            "what polic", "your suggestion", "your recommendation",
-        )
-    )
+    # KI-205 (2026-05-15) — `_is_recommendation` is now computed once earlier
+    # (immediately after retrieve) so the same boolean drives both the
+    # regulatory-chunk post-filter (KI-206) and this judge-skip.
     _skip_judge = (intent == "fact_find") or bool(in_fact_find_continuation) or _is_recommendation
     if _skip_judge:
         skip_reason = "ki171_skip_on_recommendation" if _is_recommendation else "ki091_skip_on_fact_find"
@@ -1177,6 +1202,14 @@ async def handle_turn(
     # context about the user (age/dependents/income/etc.), but it is NOT a
     # "cited policy" and should not appear in the chat's "CITED POLICIES"
     # card. Filter by insurer_slug=='profile' (set by profile_rag.upsert).
+    # KI-206 (2026-05-15) — also exclude regulatory chunks (IRDAI consolidated,
+    # etc.). They're master regulatory text, not policies the user can buy, so
+    # showing them under "CITED POLICIES" misled users into thinking the bot
+    # had policy evidence when it had only regulator text. Non-recommendation
+    # queries that genuinely ask about regulations still surface regulatory
+    # chunks via context_str (the post-retrieve filter above only drops them
+    # for recommendation/comparison intent), but they never appear in the
+    # chat's policy-citation card.
     citations = [
         {
             "policy_id": c.policy_id,
@@ -1188,7 +1221,7 @@ async def handle_turn(
             "score": round(c.score, 3),
         }
         for c in chunks
-        if (c.insurer_slug or "").lower() != "profile"
+        if (c.insurer_slug or "").lower() not in ("profile", "regulatory")
     ]
 
     # 7. INDIC CASCADE — translate the English reply back into Hinglish/Hindi,

frontend/public/admin/llm-control.html

@@ -668,28 +668,14 @@
 
     <!-- Tab 3: LLM Chain — KI-164: stripped to ONLY the 2-table view. -->
     <section id="tab-chain" class="tabpane" role="tabpanel">
-      <div class="card" id="llm-health-card">
-        <div class="row-between" style="margin-bottom: 12px;">
-          <h2>LLM health <span class="small muted" id="llm-health-snapshot-ts"></span></h2>
-          <div class="actions">
-            <button id="btn-refresh-llm-health">Refresh</button>
-          </div>
-        </div>
-        <!-- KI-184 (2026-05-15) — Tier banner so the table below isn't
-             confused with the actual primary serving traffic. Real-world
-             call order: Tier 0 Google Gemini → Tier 1 NIM (shown below) →
-             Tier 2 OpenRouter free pool → NIM nemotron-49b last resort. -->
-        <div style="margin-bottom: 14px; padding: 10px 14px; border: 1px solid var(--border); border-radius: 8px; background: #0b0d12; font-size: 12px; line-height: 1.5;">
-          <strong style="color: var(--text);">3-tier brain chain (KI-179 + ADR-040):</strong>
-          <span class="muted">
-            Tier 0 — <strong>Google Gemini 2.5 Flash / Flash Lite</strong> (PRIMARY, free 1500/day)
-            &nbsp;·&nbsp; Tier 1 — NIM (shown below)
-            &nbsp;·&nbsp; Tier 2 — OpenRouter free pool
-            &nbsp;·&nbsp; Last resort — NIM Nemotron-49B
-          </span>
-        </div>
-        <div id="llm-health-chains" class="llm-simple-tables"></div>
-      </div>
+      <!-- KI-207 (2026-05-15) — per user instruction "There should be nothing else
+           in that LLM router part of the control panel." Only the 3 elements below
+           may live inside this section: the h2 header (with snapshot timestamp
+           span), the refresh button, and the chains container that holds the
+           two generated tables (Currently In Use + All Eligible Models). -->
+      <h2>LLM health <span class="small muted" id="llm-health-snapshot-ts"></span></h2>
+      <button id="btn-refresh-llm-health">Refresh</button>
+      <div id="llm-health-chains" class="llm-simple-tables"></div>
     </section>
   </div>
 
@@ -1442,230 +1428,279 @@
         );
       }
 
-      function renderLlmSimpleTables(host, chains) {
-        // ---- Table 1: What's Live ----
+      // KI-200 (2026-05-15) — chain definitions mirror backend/sales_brain.py +
+      // backend/providers/tiered_brain_llm.py + nvidia_nim_llm.py chain order.
+      // Tier 0 = Google AI Studio (Gemini) — PRIMARY for fast_brain + brain.
+      // Tier 1 = NVIDIA NIM (chain order matters; last entry = "Last resort").
+      // Tier 2 = OpenRouter free pool — backstop inside NIM-failed path.
+      // Judge intentionally has NO Tier 0 (cross-family invariant, ADR-040).
+      var CHAIN_DEFINITIONS = {
+        fast_brain: {
+          tier0: { provider: 'Google AI Studio', model: 'google/gemini-2.5-flash-lite' },
+          tier1: [
+            'qwen/qwen3-next-80b-a3b-instruct',
+            'mistralai/mistral-large-3-675b-instruct-2512',
+            'meta/llama-4-maverick-17b-128e-instruct',
+            'nvidia/llama-3.3-nemotron-super-49b-v1.5'
+          ],
+          tier2: [
+            'nvidia/nemotron-3-super-120b-a12b:free',
+            'qwen/qwen3-next-80b-a3b-instruct:free',
+            'google/gemma-4-31b-it:free'
+          ]
+        },
+        brain: {
+          tier0: { provider: 'Google AI Studio', model: 'google/gemini-2.5-flash' },
+          tier1: [
+            'qwen/qwen3-next-80b-a3b-instruct',
+            'mistralai/mistral-large-3-675b-instruct-2512',
+            'meta/llama-4-maverick-17b-128e-instruct',
+            'nvidia/llama-3.3-nemotron-super-49b-v1.5'
+          ],
+          tier2: [
+            'nvidia/nemotron-3-super-120b-a12b:free',
+            'qwen/qwen3-next-80b-a3b-instruct:free',
+            'google/gemma-4-31b-it:free'
+          ]
+        },
+        judge: {
+          tier0: null,
+          tier1: [
+            'mistralai/mistral-large-3-675b-instruct-2512',
+            'meta/llama-4-maverick-17b-128e-instruct',
+            'nvidia/llama-3.3-nemotron-super-49b-v1.5'
+          ],
+          tier2: [
+            'qwen/qwen3-next-80b-a3b-instruct:free'
+          ]
+        }
+      };
+
+      // KI-187 — provider colors: blue for Google, green for NVIDIA NIM,
+      // orange for OpenRouter.
+      var PROVIDER_COLORS = {
+        'Google AI Studio': '#4285f4',
+        'NVIDIA NIM':       '#76b900',
+        'OpenRouter':       '#ffa657'
+      };
+
+      function providerForModel(model) {
+        if (!model) return 'NVIDIA NIM';
+        if (model.indexOf('google/gemini') === 0) return 'Google AI Studio';
+        if (model.indexOf(':free') !== -1) return 'OpenRouter';
+        return 'NVIDIA NIM';
+      }
+
+      // KI-200 — compute the model that would ACTUALLY serve right now for a
+      // given role. Strategy:
+      //   1. If Tier 0 exists for this role AND Gemini is available, return Tier 0.
+      //   2. Else fall back to NIM's elected primary from chains[role].current_primary.
+      //   3. Judge never uses Tier 0 (cross-family invariant) — always NIM primary.
+      // TODO(backend): expose top-level `gemini_available: bool` in
+      //   /api/admin/llm-health response. Until then we ASSUME Gemini is
+      //   available (true) — matches the deployed Tier-0-primary state.
+      function resolveCurrentlyInUse(role, chainsByRole, healthPayload) {
+        var def = CHAIN_DEFINITIONS[role];
+        var geminiAvailable = (healthPayload && typeof healthPayload.gemini_available === 'boolean')
+          ? healthPayload.gemini_available
+          : true;  // default-assume true; backend can override later
+        if (def && def.tier0 && geminiAvailable) {
+          return { model: def.tier0.model, provider: def.tier0.provider, tier: 'T0', isTier0: true };
+        }
+        var c = chainsByRole[role];
+        if (c && c.current_primary && !isRetiredModel(c.current_primary)) {
+          return { model: c.current_primary, provider: providerForModel(c.current_primary), tier: 'T1', isTier0: false };
+        }
+        return { model: '—', provider: '—', tier: '—', isTier0: false };
+      }
+
+      // KI-200 — build the unioned model list across all chains with
+      // per-chain role labels. Returns array of:
+      //   { model, provider, tier, roles: { fast_brain, brain, judge } }
+      // where each `roles[r]` is one of:
+      //   'Primary' | '1st backup' | '2nd backup' | '3rd backup' | 'Last resort' | null
+      function buildEligibleModelTable() {
+        var BACKUP_LABELS = ['Primary', '1st backup', '2nd backup', '3rd backup', '4th backup', '5th backup'];
+        var models = {};  // model -> entry
+
+        function ensureEntry(model, tier) {
+          if (!models[model]) {
+            models[model] = {
+              model: model,
+              provider: providerForModel(model),
+              tier: tier,
+              roles: { fast_brain: null, brain: null, judge: null },
+              // for sorting: minimum chain position across all chains (lower = more important)
+              minPos: 999,
+              anyPrimary: false
+            };
+          }
+          return models[model];
+        }
+
+        Object.keys(CHAIN_DEFINITIONS).forEach(function (role) {
+          var def = CHAIN_DEFINITIONS[role];
+          // Build the ordered chain for this role:
+          // [tier0?, ...tier1, ...tier2]. The LAST entry overall is "Last resort".
+          var ordered = [];
+          if (def.tier0) ordered.push({ model: def.tier0.model, tier: 'T0' });
+          (def.tier1 || []).forEach(function (m) { ordered.push({ model: m, tier: 'T1' }); });
+          (def.tier2 || []).forEach(function (m) { ordered.push({ model: m, tier: 'T2' }); });
+          var lastIdx = ordered.length - 1;
+          ordered.forEach(function (slot, idx) {
+            var entry = ensureEntry(slot.model, slot.tier);
+            // If a model appears in multiple chains/tiers, the tier sticks to
+            // its first-seen value — but Tier 0 wins (lowest number) since
+            // Gemini fast-lite/flash are unique per role and never collide.
+            if (slot.tier === 'T0') entry.tier = 'T0';
+            else if (entry.tier !== 'T0' && slot.tier === 'T1') entry.tier = 'T1';
+            // (T2 only overrides nothing.)
+            var label;
+            if (idx === lastIdx) {
+              label = 'Last resort';
+            } else if (idx < BACKUP_LABELS.length) {
+              label = BACKUP_LABELS[idx];
+            } else {
+              label = (idx + 1) + 'th backup';
+            }
+            entry.roles[role] = label;
+            if (idx < entry.minPos) entry.minPos = idx;
+            if (label === 'Primary') entry.anyPrimary = true;
+          });
+        });
+
+        // Sort: tier asc (T0 < T1 < T2), then primaries first, then by minPos.
+        var TIER_RANK = { 'T0': 0, 'T1': 1, 'T2': 2 };
+        var arr = Object.keys(models).map(function (k) { return models[k]; });
+        arr.sort(function (a, b) {
+          var ta = TIER_RANK[a.tier] != null ? TIER_RANK[a.tier] : 99;
+          var tb = TIER_RANK[b.tier] != null ? TIER_RANK[b.tier] : 99;
+          if (ta !== tb) return ta - tb;
+          if (a.anyPrimary !== b.anyPrimary) return a.anyPrimary ? -1 : 1;
+          if (a.minPos !== b.minPos) return a.minPos - b.minPos;
+          return a.model.localeCompare(b.model);
+        });
+        return arr;
+      }
+
+      function renderRoleLabelCell(label) {
+        var td = createEl('td', { className: 'role-cell' });
+        td.style.textAlign = 'center';
+        if (!label) {
+          td.textContent = '—';
+          td.style.color = 'var(--muted)';
+        } else if (label === 'Primary') {
+          td.textContent = 'Primary';
+          td.style.color = 'var(--green)';
+          td.style.fontWeight = '600';
+        } else if (label === 'Last resort') {
+          td.textContent = 'Last resort';
+          td.style.color = 'var(--muted)';
+          td.style.fontStyle = 'italic';
+        } else {
+          td.textContent = label;
+          td.style.color = 'var(--text)';
+        }
+        return td;
+      }
+
+      // KI-200 — accepts the full llm-health payload so we can read
+      // `gemini_available` (when backend exposes it). Caller signature
+      // updated below in renderLlmHealth().
+      function renderLlmSimpleTables(host, healthPayload) {
+        var chains = (healthPayload && healthPayload.chains) || [];
+        var chainsByRole = {};
+        chains.forEach(function (c) { if (c && c.role) chainsByRole[c.role] = c; });
+
+        // ---- Table 1: Currently In Use ----
         var liveBlock = createEl('div', { className: 'simple-table-block' });
-        liveBlock.appendChild(createEl('h3', { text: "What's Live" }));
+        liveBlock.appendChild(createEl('h3', { text: 'Currently In Use' }));
         var liveTable = createEl('table');
         var liveThead = createEl('thead');
         var liveHr = createEl('tr');
-        ['Use', 'Currently In Use', 'Probed'].forEach(function (h) {
-          liveHr.appendChild(createEl('th', { text: h }));
+        ['Use', 'Current Model', 'Provider', 'Tier'].forEach(function (h, i) {
+          var th = createEl('th', { text: h });
+          if (i >= 2) th.style.textAlign = 'center';
+          liveHr.appendChild(th);
         });
         liveThead.appendChild(liveHr);
         liveTable.appendChild(liveThead);
         var liveTbody = createEl('tbody');
 
-        // Index chains by role for stable Brain Fast / Brain Main / Judge ordering.
-        var chainsByRole = {};
-        (chains || []).forEach(function (c) { if (c && c.role) chainsByRole[c.role] = c; });
-
         SIMPLE_USE_ORDER.forEach(function (role) {
-          var c = chainsByRole[role];
+          var live = resolveCurrentlyInUse(role, chainsByRole, healthPayload);
           var tr = createEl('tr');
+          if (live.isTier0) {
+            // KI-200 — subtle green wash to signal "real primary serving traffic"
+            tr.style.background = 'rgba(118, 185, 0, 0.06)';
+          }
           tr.appendChild(createEl('td', { className: 'use-label', text: SIMPLE_USE_LABELS[role] }));
 
           var modelCell = createEl('td', { className: 'mono' });
-          if (c && c.current_primary && !isRetiredModel(c.current_primary)) {
-            if (c.current_primary_available === false) {
-              var dot = createEl('span', { className: 'health-dot bad' });
-              dot.style.marginRight = '6px';
-              modelCell.appendChild(dot);
-            }
-            modelCell.appendChild(document.createTextNode(c.current_primary));
-          } else {
-            modelCell.textContent = '—';
-          }
+          modelCell.textContent = live.model;
           tr.appendChild(modelCell);
 
-          var probedCell = createEl('td', { className: 'probed-cell' });
-          if (c && c.last_probe_age_seconds != null) {
-            probedCell.textContent = fmtDurationShort(c.last_probe_age_seconds) + ' ago';
-          } else {
-            probedCell.textContent = '—';
-          }
-          tr.appendChild(probedCell);
+          var provCell = createEl('td', { className: 'provider-cell' });
+          provCell.style.textAlign = 'center';
+          provCell.style.fontWeight = '600';
+          provCell.style.color = PROVIDER_COLORS[live.provider] || 'var(--muted)';
+          provCell.textContent = live.provider;
+          tr.appendChild(provCell);
+
+          var tierCell = createEl('td', { className: 'tier-cell' });
+          tierCell.style.textAlign = 'center';
+          tierCell.style.fontWeight = '600';
+          tierCell.style.color = live.tier === 'T0' ? 'var(--green)' : 'var(--muted)';
+          tierCell.textContent = live.tier;
+          tr.appendChild(tierCell);
+
           liveTbody.appendChild(tr);
         });
         liveTable.appendChild(liveTbody);
         liveBlock.appendChild(liveTable);
         host.appendChild(liveBlock);
 
-        // ---- Table 2: What's Available ----
-        // Walk every chain_members[] across the 3 chains, dedupe by model,
-        // and collect: healthy boolean + used-by-role list with primary-first ordering.
-        var modelMap = {};  // model -> { healthy, roleEntries: [{role, isPrimary, idx}] }
-        (chains || []).forEach(function (c) {
-          if (!c || !Array.isArray(c.chain_members)) return;
-          c.chain_members.forEach(function (mi, idx) {
-            if (!mi || !mi.model || isRetiredModel(mi.model)) return;
-            if (!modelMap[mi.model]) {
-              modelMap[mi.model] = {
-                healthy: !!mi.available_for_calls,
-                roleEntries: []
-              };
-            } else {
-              // If any chain reports the model as available, treat it as healthy.
-              if (mi.available_for_calls) modelMap[mi.model].healthy = true;
-            }
-            modelMap[mi.model].roleEntries.push({
-              role: c.role,
-              isPrimary: !!mi.is_current_primary,
-              idx: idx
-            });
-          });
-        });
-
+        // ---- Table 2: All Eligible Models ----
         var availBlock = createEl('div', { className: 'simple-table-block' });
-        availBlock.appendChild(createEl('h3', { text: "What's Available" }));
+        availBlock.appendChild(createEl('h3', { text: 'All Eligible Models' }));
         var availTable = createEl('table');
         var availThead = createEl('thead');
         var availHr = createEl('tr');
-        // KI-186 (2026-05-15) — added "Tier" column so Gemini + OR rows can
-        // be shown alongside NIM rows. Tier 0 = Google AI Studio, Tier 1 =
-        // NIM (probed), Tier 2 = OpenRouter free pool.
-        // KI-187 (2026-05-15) — added explicit "Provider" column ("Google
-        // AI Studio" / "NVIDIA NIM" / "OpenRouter") so users don't have to
-        // infer provider from the model-id suffix (e.g., qwen3-next-80b
-        // appears in BOTH NIM and OR; only `:free` suffix disambiguated).
-        var availCols = ['Model', 'Provider', 'Tier', 'Status'].concat(SIMPLE_USE_ORDER.map(function (r) { return SIMPLE_USE_LABELS[r]; }));
+        var availCols = ['Model', 'Provider', 'Tier'].concat(SIMPLE_USE_ORDER.map(function (r) { return SIMPLE_USE_LABELS[r]; }));
         availCols.forEach(function (h, i) {
           var th = createEl('th', { text: h });
-          if (i >= 4) th.style.textAlign = 'center';
-          if (i === 1 || i === 2 || i === 3) th.style.textAlign = 'center';
+          if (i >= 1) th.style.textAlign = 'center';
           availHr.appendChild(th);
         });
         availThead.appendChild(availHr);
         availTable.appendChild(availThead);
         var availTbody = createEl('tbody');
 
-        // KI-186 — Tier 0 (Google) + Tier 2 (OR) are STATIC rows; they're
-        // not in llm_health's probe registry but they're real members of
-        // the production chain per sales_brain.py + tiered_brain_llm.py.
-        // Status defaults to 'Healthy' (we don't have probe data for these
-        // tiers; they fail-loud at call time per the fail-loud > silent-
-        // garbage invariant).
-        var STATIC_TIER_ROWS = [
-          // Tier 0 — Google AI Studio (Gemini, free 1500/day)
-          { model: 'google/gemini-2.5-flash',          provider: 'Google AI Studio', tier: 'T0', healthy: true,  brainFast: '—',       brainMain: 'primary', judge: '—' },
-          { model: 'google/gemini-2.5-flash-lite',     provider: 'Google AI Studio', tier: 'T0', healthy: true,  brainFast: 'primary', brainMain: '—',       judge: '—' },
-          // Tier 2 — OpenRouter free pool (server-side fallback inside OR)
-          { model: 'nvidia/nemotron-3-super-120b-a12b:free', provider: 'OpenRouter', tier: 'T2', healthy: true, brainFast: '✓', brainMain: '✓', judge: '—' },
-          { model: 'qwen/qwen3-next-80b-a3b-instruct:free',  provider: 'OpenRouter', tier: 'T2', healthy: true, brainFast: '✓', brainMain: '—', judge: '✓' },
-          { model: 'google/gemma-4-31b-it:free',              provider: 'OpenRouter', tier: 'T2', healthy: true, brainFast: '✓', brainMain: '—', judge: '—' },
-        ];
-        // KI-187 — provider colors: blue for Google (matches their brand),
-        // green for NVIDIA NIM (matches NVIDIA), orange for OpenRouter
-        // (distinct from the others).
-        var PROVIDER_COLORS = {
-          'Google AI Studio': '#4285f4',   // Google blue
-          'NVIDIA NIM':       '#76b900',   // NVIDIA green
-          'OpenRouter':       '#ffa657',   // OR orange
-        };
-        STATIC_TIER_ROWS.forEach(function (row) {
+        var rows = buildEligibleModelTable();
+        rows.forEach(function (entry) {
           var tr = createEl('tr');
-          tr.appendChild(createEl('td', { className: 'model-name', text: row.model }));
-          // Provider badge — KI-187
+          tr.appendChild(createEl('td', { className: 'model-name', text: entry.model }));
+
           var provCell = createEl('td', { className: 'provider-cell' });
           provCell.style.textAlign = 'center';
           provCell.style.fontWeight = '600';
-          provCell.style.color = PROVIDER_COLORS[row.provider] || 'var(--muted)';
-          provCell.textContent = row.provider;
+          provCell.style.color = PROVIDER_COLORS[entry.provider] || 'var(--muted)';
+          provCell.textContent = entry.provider;
           tr.appendChild(provCell);
-          // Tier badge
+
           var tierCell = createEl('td', { className: 'tier-cell' });
           tierCell.style.textAlign = 'center';
           tierCell.style.fontWeight = '600';
-          tierCell.style.color = row.tier === 'T0' ? 'var(--green)' : 'var(--muted)';
-          tierCell.textContent = row.tier;
+          tierCell.style.color = entry.tier === 'T0' ? 'var(--green)' : 'var(--muted)';
+          tierCell.textContent = entry.tier;
           tr.appendChild(tierCell);
-          // Status cell
-          var statusCell = createEl('td', { className: 'status-cell' });
-          var hd = createEl('span', { className: 'health-dot ' + (row.healthy ? 'ok' : 'bad') });
-          statusCell.appendChild(hd);
-          statusCell.appendChild(document.createTextNode(row.healthy ? ' Healthy' : ' Down'));
-          tr.appendChild(statusCell);
-          // Per-role cells
-          [row.brainFast, row.brainMain, row.judge].forEach(function (val) {
-            var td = createEl('td', { className: 'role-cell' });
-            td.style.textAlign = 'center';
-            td.textContent = val;
-            if (val === '—') { td.style.color = 'var(--muted)'; }
-            else if (val === 'primary') { td.style.color = 'var(--green)'; td.style.fontWeight = '600'; }
-            tr.appendChild(td);
-          });
-          availTbody.appendChild(tr);
-        });
 
-        var modelNames = Object.keys(modelMap);
-        if (!modelNames.length) {
-          // After KI-186 we always have static Tier 0 + Tier 2 rows, so a
-          // "no models" message would be misleading even if NIM has none.
-        } else {
-          // Sort: anyone's-primary first, then alphabetical by model name.
-          modelNames.sort(function (a, b) {
-            var aPrim = modelMap[a].roleEntries.some(function (e) { return e.isPrimary; });
-            var bPrim = modelMap[b].roleEntries.some(function (e) { return e.isPrimary; });
-            if (aPrim !== bPrim) return aPrim ? -1 : 1;
-            return a.localeCompare(b);
+          SIMPLE_USE_ORDER.forEach(function (role) {
+            tr.appendChild(renderRoleLabelCell(entry.roles[role]));
           });
-          modelNames.forEach(function (model) {
-            var info = modelMap[model];
-            var tr = createEl('tr');
-            tr.appendChild(createEl('td', { className: 'model-name', text: model }));
-
-            // KI-187 — Provider badge (NIM-derived rows all come from NVIDIA NIM)
-            var provCell = createEl('td', { className: 'provider-cell' });
-            provCell.style.textAlign = 'center';
-            provCell.style.fontWeight = '600';
-            provCell.style.color = PROVIDER_COLORS['NVIDIA NIM'];
-            provCell.textContent = 'NVIDIA NIM';
-            tr.appendChild(provCell);
-
-            // KI-186 — Tier badge (NIM models are Tier 1)
-            var tierCell = createEl('td', { className: 'tier-cell' });
-            tierCell.style.textAlign = 'center';
-            tierCell.style.fontWeight = '600';
-            tierCell.style.color = 'var(--muted)';
-            tierCell.textContent = 'T1';
-            tr.appendChild(tierCell);
-
-            // Status cell: dot + word.
-            var statusCell = createEl('td', { className: 'status-cell' });
-            var hd = createEl('span', { className: 'health-dot ' + (info.healthy ? 'ok' : 'bad') });
-            statusCell.appendChild(hd);
-            statusCell.appendChild(document.createTextNode(info.healthy ? ' Healthy' : ' Down'));
-            tr.appendChild(statusCell);
-
-            // Group role entries by role for per-column lookup.
-            var entriesByRole = {};
-            info.roleEntries.forEach(function (e) {
-              if (!entriesByRole[e.role]) entriesByRole[e.role] = [];
-              entriesByRole[e.role].push(e);
-            });
-            // Per-role cell: (primary) / ✓ / —.
-            SIMPLE_USE_ORDER.forEach(function (role) {
-              var roleEntries = entriesByRole[role] || [];
-              var td = createEl('td', { className: 'role-cell' });
-              td.style.textAlign = 'center';
-              if (!roleEntries.length) {
-                td.textContent = '—';
-                td.style.color = 'var(--muted)';
-              } else if (roleEntries.some(function (e) { return e.isPrimary; })) {
-                td.textContent = 'primary';
-                td.style.color = 'var(--green)';
-                td.style.fontWeight = '600';
-              } else {
-                td.textContent = '✓';
-                td.style.color = 'var(--text)';
-                td.title = 'Backup in this chain';
-              }
-              tr.appendChild(td);
-            });
 
-            availTbody.appendChild(tr);
-          });
-        }
+          availTbody.appendChild(tr);
+        });
         availTable.appendChild(availTbody);
         availBlock.appendChild(availTable);
         host.appendChild(availBlock);
@@ -1810,7 +1845,8 @@
         }
 
         clearChildren(chainsHost);
-        renderLlmSimpleTables(chainsHost, STATE.llmHealth.chains || []);
+        // KI-200 — pass the full payload so renderer can read gemini_available
+        renderLlmSimpleTables(chainsHost, STATE.llmHealth);
 
         renderLlmHealthCandidates(STATE.llmHealth.candidates || []);
         renderLlmHealthRecent(STATE.llmHealth.recent_turns || []);

frontend/src/app/page.tsx

@@ -329,6 +329,20 @@ export default function Page() {
 
   async function send(text: string) {
     if (!text.trim() || busy) return;
+    // KI-204 (2026-05-15) — silence any prior bot TTS BEFORE submitting.
+    // User starting a new turn always takes precedence over the bot's
+    // current reply audio. Covers typed sends, voice barge-in, manual Send
+    // button, programmatic submits — every path through send() gets this.
+    if (typeof document !== "undefined") {
+      document.querySelectorAll("audio").forEach((el) => {
+        try {
+          (el as HTMLAudioElement).pause();
+          (el as HTMLAudioElement).currentTime = 0;
+        } catch {
+          // ignore — element may be in a state that disallows pause
+        }
+      });
+    }
     setBusy(true);
     // KI-165 (2026-05-15) — flip the text-in-flight flag so the voice hook
     // (useLiveConversation) discards any captures that close during this

frontend/src/lib/useStreamingVoice.ts

@@ -61,7 +61,12 @@ const BARGE_IN_BASE_THRESHOLD = 0.005;
 // residual mic bleed (after AEC) and the user's normal-volume speech,
 // making barge-in trivial. 0.6 is loud enough to hear clearly on
 // headphones and laptop speakers without overpowering user speech.
-const VOICE_MODE_TTS_VOLUME = 0.6;
+// KI-211 (2026-05-15) — was 0.6; lowered to 0.3 because first-turn barge-in
+// fails when adaptive calibration (KI-195) hasn't sampled user_speech_rms yet.
+// 0.3 is loud enough to hear clearly on speakers + mic bleed is well under
+// the static BARGE_IN_RMS_THRESHOLD, so users can talk over the bot on the
+// first turn without needing prior calibration.
+const VOICE_MODE_TTS_VOLUME = 0.3;
 // KI-195 (2026-05-15) — adaptive TTS volume calibration relative to user's
 // own measured speech level. Architecture: while user speaks (recorder
 // active, NOT TTS) we sample mic RMS and track a rolling peak in
@@ -78,6 +83,21 @@ const VOLUME_CALIB_TICK_MS = 300;              // calibration sample period duri
 const VOLUME_CALIB_DUCK_FACTOR = 0.8;          // multiply el.volume by this per tick if too loud
 const VOLUME_CALIB_FLOOR = 0.15;               // never drop bot below this — must stay audible
 
+// KI-202 (2026-05-15) — utterance batching grace window.
+// Web Speech API's `onend` fires after ~1.5s silence, which means a natural
+// mid-sentence pause ("So it will be just [pause] me") triggers TWO separate
+// onend events and the user's sentence is submitted in two halves. We delay
+// the actual submission by UTTERANCE_GRACE_MS after onend; if recognition
+// re-fires (next word burst) before the timer expires, we append the new
+// text/audio chunks and reset the timer. Only after a full UTTERANCE_GRACE_MS
+// of true silence do we submit.
+const UTTERANCE_GRACE_MS = 1500;
+// KI-203 (2026-05-15) — post-TTS result-drop window.
+// `recognition.abort()` doesn't immediately stop result delivery — onresult
+// events from the now-abandoned recognition can keep arriving for a beat
+// afterwards. Keep dropping results for this many ms after TTS ends.
+const POST_TTS_DROP_MS = 300;
+
 // Minimal types for the Web Speech API since lib.dom.d.ts ships them under
 // `webkitSpeechRecognition` only and the standard `SpeechRecognition` symbol
 // is still vendor-prefixed in most browsers as of 2026-05.
@@ -177,6 +197,24 @@ export function useStreamingVoice(
   // Tracked via a MutationObserver + per-element play/pause/ended hooks.
   const isTtsPlayingRef = useRef(false);
   const ttsAudioElementsRef = useRef<Set<HTMLAudioElement>>(new Set());
+  // KI-203 (2026-05-15) — silently discard SpeechRecognition.onresult events
+  // while this flag is true. Flipped on the instant TTS playback starts
+  // (closes the ~100-300ms window between `audio.play()` and our abort()
+  // taking effect, during which bot voice was being transcribed as user
+  // input). Flipped back ~POST_TTS_DROP_MS after TTS ends so any in-flight
+  // results from the dying recognition pipeline are still suppressed.
+  const dropResultsRef = useRef(false);
+  const dropResultsClearTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  // KI-202 (2026-05-15) — utterance-batching state.
+  // pendingUtteranceRef accumulates the Web Speech transcript across multiple
+  // onend events separated by sub-grace-window pauses. pendingChunksRef does
+  // the same for MediaRecorder blobs so the Sarvam POST sees the WHOLE
+  // utterance, not just the tail after the last pause. pendingSubmitTimerRef
+  // is the grace-window setTimeout; it gets reset every time onend appends
+  // more content.
+  const pendingUtteranceRef = useRef<string>("");
+  const pendingChunksRef = useRef<Blob[]>([]);
+  const pendingSubmitTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   // ----------------------------------------------------------------------
   // KI-168 PHASE 2 — Sarvam authoritative-transcript layer.
@@ -207,6 +245,23 @@ export function useStreamingVoice(
     }
   }, []);
 
+  // KI-210 (2026-05-15) — wait for an in-flight text turn to clear instead of
+  // dropping the accumulated voice utterance. Polls isTextRequestPendingRef
+  // every 300ms; resolves true once the flag clears, or false if the
+  // maxWaitMs cap elapses first (we then proceed anyway rather than leak the
+  // utterance forever on a stuck text request).
+  const waitForTextClear = useCallback(async (maxWaitMs = 30000): Promise<boolean> => {
+    const startTs = Date.now();
+    while (isTextRequestPendingRef.current) {
+      if (Date.now() - startTs > maxWaitMs) {
+        console.debug("[useStreamingVoice] KI-210 wait timed out, submitting anyway");
+        return false; // gave up waiting — proceed anyway
+      }
+      await new Promise((r) => setTimeout(r, 300));
+    }
+    return true; // text cleared, ok to proceed
+  }, [isTextRequestPendingRef]);
+
   const safeStart = useCallback(() => {
     const rec = recognitionRef.current;
     if (!rec) return;
@@ -341,6 +396,17 @@ export function useStreamingVoice(
     };
 
     rec.onresult = (ev: SpeechRecognitionEventLike) => {
+      // KI-203 (2026-05-15) — early-return while TTS is playing (or within
+      // the POST_TTS_DROP_MS window after TTS ends). recognition.abort()
+      // doesn't immediately stop result delivery, so we silently discard
+      // every chunk that arrives during the dirty window. Without this, bot
+      // 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");
+        return;
+      }
       let interim = "";
       // Walk every result; finals get pushed onto finalsRef, interims get
       // concatenated into a running string that's displayed in the input.
@@ -423,109 +489,191 @@ export function useStreamingVoice(
         return drained;
       };
 
-      // If there's no audio recorder, or text is racing, fall back to the
-      // Phase 1 behaviour: submit the Web Speech transcript and bail.
-      if (!recorderActiveRef.current || textRacing) {
-        if (webSpeechText && !textRacing) {
-          onFinalRef.current(webSpeechText);
-        }
-        // Best-effort: clear any partial audio so the next utterance isn't
-        // contaminated with the previous one's tail.
-        if (recorderActiveRef.current) drainChunks();
-        scheduleRestart();
-        return;
+      // KI-202 (2026-05-15) — utterance batching. Web Speech's onend fires
+      // after ~1.5s of silence, so a natural mid-sentence pause splits one
+      // utterance into two onend events and the user's sentence gets
+      // submitted in halves ("First word getting cut off. Cutoff is the
+      // biggest issue. Auto-submitting without capturing the first half
+      // or the second half"). Instead of submitting immediately, we
+      // append THIS onend's text + audio chunks to pendingUtterance*Ref
+      // buffers, then start (or reset) a UTTERANCE_GRACE_MS timer. If
+      // recognition restarts (auto-restart picks up the next word burst)
+      // within the grace window, the next onend appends more content +
+      // resets the timer. Only after a FULL UTTERANCE_GRACE_MS of true
+      // silence does the timer fire and submit the accumulated buffer.
+      //
+      // Pauses < 1.5s merge into one turn (intended fix).
+      // Pauses > 1.5s split (intended — that IS a new turn).
+
+      // Drain the CURRENT onend's chunks now so the recorder keeps capturing
+      // the next word burst without contamination across pending utterances.
+      const drainedThisEnd = recorderActiveRef.current ? drainChunks() : [];
+      if (webSpeechText) {
+        pendingUtteranceRef.current = pendingUtteranceRef.current
+          ? `${pendingUtteranceRef.current} ${webSpeechText}`
+          : webSpeechText;
       }
+      if (drainedThisEnd.length > 0) {
+        pendingChunksRef.current.push(...drainedThisEnd);
+      }
+      console.debug("[useStreamingVoice] KI-202 onend appended to pending utterance", {
+        thisTextLen: webSpeechText.length,
+        thisChunkCount: drainedThisEnd.length,
+        pendingTextLen: pendingUtteranceRef.current.length,
+        pendingChunkCount: pendingChunksRef.current.length,
+        textRacing,
+      });
 
-      // Sarvam path. Fire-and-forget so we don't block the recognition
-      // restart loop on the network round-trip.
-      void (async () => {
-        // Snapshot user-visible interim so the input area doesn't go blank
-        // while Sarvam is in flight. The page-side input still shows the
-        // Web Speech transcript; we'll overwrite it via onFinalTranscript
-        // once Sarvam returns.
-        if (webSpeechText) onInterimRef.current(webSpeechText);
-
-        // We need to stop the recorder to get the final dataavailable
-        // chunk; then we re-arm a new recorder for the next utterance.
-        await stopRecorder();
-        const drained = drainChunks();
-        const totalSize = drained.reduce((n, b) => n + b.size, 0);
-        console.debug("[useStreamingVoice] silence-detect", {
-          webSpeechLen: webSpeechText.length,
-          chunkCount: drained.length,
-          blobBytes: totalSize,
-        });
-
-        // Re-arm audio capture for the next utterance (don't block on it).
-        teardownAudio();
-        // Only restart the audio pipeline if the user still wants the mic
-        // live. Fire-and-forget; recognition restart is scheduled below.
-        if (wantRunningRef.current) {
-          void ensureAudioCapture();
+      // Mic restart happens immediately regardless of grace window — we
+      // WANT recognition to come back online so it can pick up the next
+      // word burst within the grace window and append to pending.
+      scheduleRestart();
+
+      // KI-210 (2026-05-15) — DO NOT drop pending utterance when text is
+      // racing. Previously we cleared pendingUtteranceRef + pendingChunksRef
+      // here, which silently lost any voice the user spoke during the bot's
+      // text-submit/TTS-thinking gap. The downstream wait-and-retry inside
+      // `submitPendingUtterance` (timer fire) + the post-await wait inside
+      // the Sarvam fire-and-forget now hold the buffer until the text turn
+      // clears, then submit. We leave `textRacing` as a debug breadcrumb in
+      // the log above and continue accumulating.
+
+      // KI-210 — refactor the grace-timer body into a named async function
+      // so it can re-schedule itself (wait-and-retry) when text is in flight
+      // instead of dropping the utterance. Capped at 30s total wait so a
+      // stuck text request can't leak the timer forever; if the cap fires
+      // we proceed with submission anyway (better to submit than drop).
+      const SUBMIT_WAIT_CAP_MS = 30000;
+      const submitStartTsRef = { ts: 0 };
+      const submitPendingUtterance = async () => {
+        pendingSubmitTimerRef.current = null;
+
+        // KI-210 — if text is still in flight when the grace window fires,
+        // wait instead of dropping. Re-schedule a 300ms retry until either
+        // text clears or we hit the 30s cap.
+        if (isTextRequestPendingRef.current) {
+          if (submitStartTsRef.ts === 0) submitStartTsRef.ts = Date.now();
+          if (Date.now() - submitStartTsRef.ts > SUBMIT_WAIT_CAP_MS) {
+            console.debug("[useStreamingVoice] KI-210 timer wait cap reached; submitting anyway");
+            // fall through and submit
+          } else {
+            console.debug("[useStreamingVoice] KI-210 timer fired but text in flight; waiting 300ms");
+            pendingSubmitTimerRef.current = setTimeout(() => {
+              void submitPendingUtterance();
+            }, 300);
+            return;
+          }
         }
 
-        // Skip submit when there's effectively no audio or no Web Speech
-        // text. ~3 KB is the empirical noise floor used by the PTT path's
-        // KI-134 silence guard (page.tsx uses 1 KB; we're stricter here
-        // because Live mode auto-fires on every silence pause).
-        const MIN_BLOB_BYTES = 3000;
-        if (!webSpeechText && totalSize < MIN_BLOB_BYTES) {
-          console.debug("[useStreamingVoice] skipping submit — no text and tiny blob");
-          scheduleRestart();
-          return;
-        }
+        const accumulatedText = pendingUtteranceRef.current.trim();
+        const accumulatedChunks = pendingChunksRef.current;
+        pendingUtteranceRef.current = "";
+        pendingChunksRef.current = [];
+        console.debug("[useStreamingVoice] KI-202 grace window elapsed — submitting", {
+          textLen: accumulatedText.length,
+          chunkCount: accumulatedChunks.length,
+        });
 
-        // If Web Speech got nothing, but we have a real blob, still try
-        // Sarvam — Web Speech occasionally drops short utterances on noisy
-        // mics that Sarvam handles fine.
-        if (!webSpeechText && totalSize < MIN_BLOB_BYTES) {
-          scheduleRestart();
+        // No-recorder path: just submit Web Speech text.
+        if (!recorderActiveRef.current || accumulatedChunks.length === 0) {
+          if (accumulatedText) {
+            onFinalRef.current(accumulatedText);
+          }
           return;
         }
 
-        // Race-check again after the await — text turn may have started
-        // while we were stopping the recorder.
-        if (isTextRequestPendingRef.current) {
-          console.debug("[useStreamingVoice] text turn started mid-await; dropping voice turn");
-          scheduleRestart();
-          return;
-        }
+        // Sarvam path. Fire-and-forget so we don't block recognition.
+        void (async () => {
+          // Snapshot user-visible interim so the input area doesn't go blank
+          // while Sarvam is in flight. The page-side input still shows the
+          // Web Speech transcript; we'll overwrite it via onFinalTranscript
+          // once Sarvam returns.
+          if (accumulatedText) onInterimRef.current(accumulatedText);
+
+          // We need to stop the recorder to get the final dataavailable
+          // chunk for the LAST burst (anything mid-recording when the grace
+          // window opened is in chunksRef, which we now flush into our
+          // accumulated set before posting).
+          await stopRecorder();
+          const tailChunks = drainChunks();
+          const allChunks = [...accumulatedChunks, ...tailChunks];
+          const totalSize = allChunks.reduce((n, b) => n + b.size, 0);
+          console.debug("[useStreamingVoice] KI-202 batched submit", {
+            webSpeechLen: accumulatedText.length,
+            chunkCount: allChunks.length,
+            blobBytes: totalSize,
+          });
+
+          // Re-arm audio capture for the next utterance (don't block on it).
+          teardownAudio();
+          if (wantRunningRef.current) {
+            void ensureAudioCapture();
+          }
+
+          // Skip submit when there's effectively no audio or no Web Speech
+          // text. ~3 KB is the empirical noise floor used by the PTT path's
+          // KI-134 silence guard.
+          const MIN_BLOB_BYTES = 3000;
+          if (!accumulatedText && totalSize < MIN_BLOB_BYTES) {
+            console.debug("[useStreamingVoice] KI-202 skipping submit — no text and tiny blob");
+            return;
+          }
+
+          // KI-210 — wait-and-retry instead of dropping. If a text turn
+          // started during the await above, hold the utterance until it
+          // clears (capped at 30s) instead of throwing it away.
+          await waitForTextClear();
 
-        let authoritativeText = webSpeechText;
-        if (drained.length > 0 && totalSize >= MIN_BLOB_BYTES) {
-          const blob = new Blob(drained, { type: recorderMimeRef.current || "audio/webm" });
-          const controller = new AbortController();
-          const timeoutId = setTimeout(() => controller.abort(), 8000);
-          try {
-            console.debug("[useStreamingVoice] POST /api/transcribe", { bytes: blob.size, mime: blob.type, lang: language });
-            const sarvam = await postTranscribe(blob, language, controller.signal);
-            const sarvamText = (sarvam.text || "").trim();
-            if (sarvamText) {
-              authoritativeText = sarvamText;
-              console.debug("[useStreamingVoice] Sarvam OK", {
-                latency_ms: sarvam.latency_ms,
-                webSpeechLen: webSpeechText.length,
-                sarvamLen: sarvamText.length,
-              });
-            } else {
-              console.debug("[useStreamingVoice] Sarvam returned empty; using Web Speech fallback");
+          let authoritativeText = accumulatedText;
+          if (allChunks.length > 0 && totalSize >= MIN_BLOB_BYTES) {
+            const blob = new Blob(allChunks, { type: recorderMimeRef.current || "audio/webm" });
+            const controller = new AbortController();
+            const timeoutId = setTimeout(() => controller.abort(), 8000);
+            try {
+              console.debug("[useStreamingVoice] POST /api/transcribe", { bytes: blob.size, mime: blob.type, lang: language });
+              const sarvam = await postTranscribe(blob, language, controller.signal);
+              const sarvamText = (sarvam.text || "").trim();
+              if (sarvamText) {
+                authoritativeText = sarvamText;
+                console.debug("[useStreamingVoice] Sarvam OK", {
+                  latency_ms: sarvam.latency_ms,
+                  webSpeechLen: accumulatedText.length,
+                  sarvamLen: sarvamText.length,
+                });
+              } else {
+                console.debug("[useStreamingVoice] Sarvam returned empty; using Web Speech fallback");
+              }
+            } catch (err) {
+              console.debug("[useStreamingVoice] Sarvam failed; using Web Speech fallback", err);
+            } finally {
+              clearTimeout(timeoutId);
             }
-          } catch (err) {
-            console.debug("[useStreamingVoice] Sarvam failed; using Web Speech fallback", err);
-          } finally {
-            clearTimeout(timeoutId);
           }
-        }
 
-        if (authoritativeText && !isTextRequestPendingRef.current) {
-          onFinalRef.current(authoritativeText);
-        }
-        scheduleRestart();
-      })();
+          // KI-210 — final wait-and-retry after Sarvam round-trip. Don't
+          // drop the now-authoritative transcript if text raced us during
+          // the network call.
+          if (authoritativeText) {
+            await waitForTextClear();
+            onFinalRef.current(authoritativeText);
+          }
+        })();
+      };
+
+      // (Re)start the grace-window timer. Every onend resets it, so as long
+      // as the user keeps starting new word bursts within 1.5s of the last
+      // silence, the timer never fires and the utterance keeps growing.
+      if (pendingSubmitTimerRef.current !== null) {
+        clearTimeout(pendingSubmitTimerRef.current);
+      }
+      submitStartTsRef.ts = 0;
+      pendingSubmitTimerRef.current = setTimeout(() => {
+        void submitPendingUtterance();
+      }, UTTERANCE_GRACE_MS);
     };
 
     return rec;
-  }, [language, isTextRequestPendingRef, clearRestartTimer, safeStart, stopRecorder, teardownAudio, ensureAudioCapture]);
+  }, [language, isTextRequestPendingRef, clearRestartTimer, safeStart, stopRecorder, teardownAudio, ensureAudioCapture, waitForTextClear]);
 
   const start = useCallback(() => {
     if (!isSupported) {
@@ -558,6 +706,14 @@ export function useStreamingVoice(
     }
     teardownAudio();
     finalsRef.current = [];
+    // 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) {
+      clearTimeout(pendingSubmitTimerRef.current);
+      pendingSubmitTimerRef.current = null;
+    }
+    pendingUtteranceRef.current = "";
+    pendingChunksRef.current = [];
     onListeningRef.current(false);
   }, [clearRestartTimer, teardownAudio]);
 
@@ -944,6 +1100,16 @@ export function useStreamingVoice(
         // TTS just started — abort any in-flight recognition so it stops
         // transcribing the bot voice.
         console.debug("[useStreamingVoice] KI-188 TTS started — pausing recognition");
+        // KI-203 (2026-05-15) — flip the result-drop flag the INSTANT TTS
+        // starts. abort() below has a ~100-300ms tail during which onresult
+        // can still fire with bot-voice transcripts; the flag closes that
+        // window unconditionally.
+        if (dropResultsClearTimerRef.current !== null) {
+          clearTimeout(dropResultsClearTimerRef.current);
+          dropResultsClearTimerRef.current = null;
+        }
+        dropResultsRef.current = true;
+        console.debug("[useStreamingVoice] KI-203 dropResultsRef=true (TTS start)");
         const rec = recognitionRef.current;
         if (rec) {
           try { rec.abort(); } catch { /* ignore */ }
@@ -982,6 +1148,19 @@ export function useStreamingVoice(
         // TTS just ended — let the heartbeat/visibility listeners revive.
         // Trigger immediately too so the user doesn't wait ~4s.
         console.debug("[useStreamingVoice] KI-188 TTS ended — resuming recognition");
+        // KI-203 (2026-05-15) — keep dropping recognition results for
+        // POST_TTS_DROP_MS after TTS ends. The recognition pipeline we
+        // abort()'d at TTS-start can still deliver buffered events for a
+        // beat; without this delayed clear, the tail of the bot's TTS
+        // leaks into the input box as the user starts speaking.
+        if (dropResultsClearTimerRef.current !== null) {
+          clearTimeout(dropResultsClearTimerRef.current);
+        }
+        dropResultsClearTimerRef.current = setTimeout(() => {
+          dropResultsRef.current = false;
+          dropResultsClearTimerRef.current = null;
+          console.debug("[useStreamingVoice] KI-203 dropResultsRef=false (post-TTS window over)");
+        }, POST_TTS_DROP_MS);
         stopBargeInLoop();
         // KI-195 — freeze the per-element calibrated volume and resume
         // learning the user's speech RMS for the next turn.
@@ -1075,6 +1254,13 @@ export function useStreamingVoice(
       });
       ttsAudioElementsRef.current.clear();
       isTtsPlayingRef.current = false;
+      // KI-203 — clear the post-TTS drop-results window timer so a
+      // disabled-then-re-enabled voice mode doesn't inherit a stale flag.
+      if (dropResultsClearTimerRef.current !== null) {
+        clearTimeout(dropResultsClearTimerRef.current);
+        dropResultsClearTimerRef.current = null;
+      }
+      dropResultsRef.current = false;
       // KI-189 — release AnalyserNode + AudioContext on unmount / disable.
       teardownAnalyser();
     };
@@ -1132,6 +1318,13 @@ export function useStreamingVoice(
       }
       recognitionRef.current = null;
       teardownAudio();
+      // KI-202 — clear pending utterance grace timer on unmount.
+      if (pendingSubmitTimerRef.current !== null) {
+        clearTimeout(pendingSubmitTimerRef.current);
+        pendingSubmitTimerRef.current = null;
+      }
+      pendingUtteranceRef.current = "";
+      pendingChunksRef.current = [];
     };
   }, [clearRestartTimer, teardownAudio]);
 

tests/test_credits_election.py

@@ -174,16 +174,19 @@ class TestElectionCreditGate(unittest.TestCase):
                          "Quota-exhausted Groq should be skipped despite faster latency.")
 
     def test_nim_preferred_over_faster_groq_when_eligible(self) -> None:
-        """KI-087 NIM-first preference: even when both candidates are healthy
-        and Groq has plenty of credits AND is 3x faster on latency, NIM must
-        still be elected as PRIMARY. Groq + OpenRouter are emergency-fallback
-        only, not picked-on-latency. Replaces the pre-KI-087
-        `test_groq_above_water_picked_in_election` which asserted the
-        opposite (Groq wins on latency)."""
+        """KI-201 chain-order: chain head wins regardless of latency.
+
+        Under KI-201 the chain-head model is always primary when eligible.
+        Pre-KI-201 (KI-087) the elector applied a NIM-first preference on
+        top of latency-score; that preference is now redundant because the
+        canonical chains in nvidia_nim_llm.py list NIM models first by
+        construction. This test exercises the new contract: with NIM as
+        chain-head and Groq slower in the chain, NIM wins on POSITION
+        not on latency-vs-NIM-preference."""
         groq_model = "groq:llama-3.3-70b-versatile"
         nim_model = "qwen/qwen3-next-80b-a3b-instruct"
         gh = _healthy_now(groq_model)
-        gh.latency_ms = 100  # faster
+        gh.latency_ms = 100  # faster (irrelevant under KI-201)
         gh.credits_remaining = 10000.0  # well above water
         gh.credits_unit = "tokens_day"
         gh.credits_low_water = GROQ_TOKENS_LOW_WATER
@@ -191,25 +194,33 @@ class TestElectionCreditGate(unittest.TestCase):
         llm_health._STATE[groq_model] = gh
 
         nh = _healthy_now(nim_model)
-        nh.latency_ms = 300  # slower but NIM
+        nh.latency_ms = 300  # slower but chain-head
         llm_health._STATE[nim_model] = nh
 
+        # NIM is chain-head; under chain-order election it wins regardless
+        # of Groq being faster.
         with mock.patch.object(
-            llm_health, "_chain_for", return_value=[groq_model, nim_model]
+            llm_health, "_chain_for", return_value=[nim_model, groq_model]
         ):
             primary = llm_health.get_primary("brain")
         self.assertEqual(
             primary, nim_model,
-            "KI-087: NIM must beat Groq as PRIMARY even when Groq is faster + "
-            "has credits. NIM is the strategic free provider; Groq + OpenRouter "
-            "are emergency fallback only.",
+            "KI-201: chain-head wins regardless of latency. NIM is listed "
+            "first in BRAIN_CHAIN by construction (KI-175); chain-order "
+            "election picks it without needing a separate NIM-first rule.",
         )
 
     def test_groq_picked_when_nim_pool_empty(self) -> None:
-        """KI-087 fallthrough: when NO eligible NIM candidate exists (all NIM
-        models down / out of credits / not in chain), election falls through
-        to the best non-NIM candidate as PRIMARY. Locks in the safety net so
-        a full NIM regional outage still produces a working brain call."""
+        """KI-201 chain-order fallthrough: when NO eligible NIM candidate
+        exists (all NIM models down / out of credits / not in chain),
+        election walks the chain in order and picks the FIRST eligible
+        non-NIM candidate. Locks in the safety net so a full NIM regional
+        outage still produces a working brain call.
+
+        Pre-KI-201 (KI-087) this test asserted Groq won via latency-score
+        even when listed second in the chain. Under chain-order election
+        the chain is the truth, so we list Groq first to exercise the
+        same fallthrough behaviour."""
         groq_model = "groq:llama-3.3-70b-versatile"
         or_model = "openrouter:openai/gpt-oss-120b"
         gh = _healthy_now(groq_model)
@@ -221,18 +232,18 @@ class TestElectionCreditGate(unittest.TestCase):
         llm_health._STATE[groq_model] = gh
 
         oh = _healthy_now(or_model)
-        oh.latency_ms = 800  # slower
+        oh.latency_ms = 800  # slower (irrelevant under KI-201 chain-order)
         llm_health._STATE[or_model] = oh
 
-        # Note the chain has NO NIM candidates.
+        # Chain has NO NIM candidates; Groq is chain-head.
         with mock.patch.object(
-            llm_health, "_chain_for", return_value=[or_model, groq_model]
+            llm_health, "_chain_for", return_value=[groq_model, or_model]
         ):
             primary = llm_health.get_primary("brain")
         self.assertEqual(
             primary, groq_model,
-            "KI-087 fallthrough: no NIM eligible → election picks the highest-"
-            "scored non-NIM candidate (Groq's 100ms beats OpenRouter's 800ms).",
+            "KI-201 chain-order: no NIM eligible → election picks the "
+            "FIRST eligible candidate in chain order (Groq is chain-head).",
         )
 
     def test_none_credits_is_permissive(self) -> None: