fix(web): second sweep round - 20 confirmed findings across all surfaces

The agent sweep finished all nine surfaces (52 raw findings, 39 after
dedup) and the verify stage confirmed the set below against the tree.

Chat:
- New chat now orphans an in-flight ask via a run counter: the zombie
  stream can no longer write its answer, citations, status, or busy
  state into the next conversation.
- Citation chips in older messages open the source-page modal for
  their own evidence instead of highlighting the last turn's panel
  (the panel only ever shows the last retrieval).
- Stream/transport failures keep the partially streamed answer with an
  interruption note instead of replacing it with key-blaming boilerplate.
- Error and notice turns are excluded from the prior-turn history sent
  to condense/generation; error turns get visible styling.
- Figure thumbnails get real CSS (page PNGs rendered at natural size);
  the routing card reports the mode the turn ran under, not the
  current setting; agentic turns get their own pill; cited visual
  candidates show their citation number instead of IMG.

Layout/shell:
- The tweaks effect no longer clobbers the persisted layout on every
  reload, and the "focus" tweak maps to the layout value the state
  machine actually uses.
- ChatView stays mounted across tab switches - switching to Papers and
  back no longer destroys the conversation. The hash follows the tab,
  so stale deep links stop overriding the saved tab.
- KeyModal clears stale typed input on every open. routing_available
  must be positively confirmed by /health.

Score display:
- Relevance bars are min-max scaled within the result set in both the
  chat panel and Inspection (all-negative logit sets rendered every
  bar empty; mixed sets rendered the worst rows full).

Inspection/Papers/Figures/Why:
- The forced paper filter is passed to traces AND shown as a chip;
  example chips disable while a trace runs.
- The figures gallery surfaces docling table chunks (kind=table, ~30%
  of the gallery) as "table" with their own filter chip.
- Paper drawer captions hide internal [chunk_id] placeholders; the
  papers empty state explains itself; the dead "other" source filter
  only appears when the corpus mixes sources.
- The Why tab no longer renders "0/0" stats while its data loads, and
  its CTA stops promising live routing on routerless deployments.
- Orphaned CSS fixed or removed (.figthumb hover, .stage-dive); the
  bbox note no longer promises a region box that multi-page chunks
  never have; SSE streams flush a final unterminated frame.

Author: NorthernLightx

web/app/api.js

@@ -242,7 +242,12 @@
         : new RegExp(`^fig(?:ure)?\\.?\\s*${num}\\b`, "i");
       for (const paperId of papers) {
         const f = figureIndex.find((g) => g.paper_id === paperId && re.test(String(g.caption || "").trim()));
-        if (f && typeof f.page_number === "number") out.push({ paperId, page: f.page_number });
+        if (f && typeof f.page_number === "number") {
+          out.push({ paperId, page: f.page_number });
+          // One page per reference — matching the same "Figure 2" in a second
+          // paper would crowd out the question's other references.
+          break;
+        }
       }
     }
     return out.slice(0, 2);
@@ -359,6 +364,23 @@
         if (obj.usage) usage = obj.usage;
       }
     }
+    // Flush a final line that arrived without a trailing newline — the
+    // usage-bearing frame is often the last thing in the stream.
+    const tail = buf.trim();
+    if (tail.startsWith("data:")) {
+      const payload = tail.slice(5).trim();
+      if (payload && payload !== "[DONE]") {
+        try {
+          const obj = JSON.parse(payload);
+          if (obj.usage) usage = obj.usage;
+          const delta = obj.choices?.[0]?.delta?.content || "";
+          if (delta) {
+            acc += delta;
+            onDelta(delta);
+          }
+        } catch { /* partial frame — drop */ }
+      }
+    }
     return { text: acc, usage };
   }
 

web/app/app.jsx

@@ -201,6 +201,9 @@ const KEY_MODAL_COPY = {
 };
 function KeyModal({ open, onSave, onClose }) {
   const [val, setVal] = useState("");
+  // Reset on every open: a stale half-typed key from a previous open must not
+  // be one Enter press away from becoming the active key.
+  useEffect(() => { if (open) setVal(""); }, [open]);
   useEffect(() => {
     if (!open) return;
     const onEsc = (e) => { if (e.key === "Escape") onClose(); };
@@ -254,7 +257,12 @@ function App() {
 
   const setTheme = (th) => {setThemeRaw(th);localStorage.setItem("sr-theme", th);};
   useEffect(() => {document.documentElement.setAttribute("data-theme", theme);}, [theme]);
-  useEffect(() => {localStorage.setItem("sr-tab", tab);}, [tab]);
+  useEffect(() => {
+    localStorage.setItem("sr-tab", tab);
+    // Keep the hash in sync — a stale deep-link hash would otherwise override
+    // the saved tab on every reload.
+    if ((location.hash || "").replace(/^#/, "") !== tab) history.replaceState(null, "", "#" + tab);
+  }, [tab]);
   useEffect(() => {localStorage.setItem("sr-layout", layout);}, [layout]);
 
   // Real corpus data: the paper list (feeds the paper filter) and whether page
@@ -263,7 +271,10 @@ function App() {
   useEffect(() => {
     window.RAG.loadPapers().then(setPapers);
     window.RAG.loadFigures().then(setFigures);
-    window.RAG.loadHealth().then((h) => { setPagesAvailable(!!h.pages_available); setDemoAvailable(!!h.demo_available); setRoutingAvailable(h.routing_available !== false); });
+    // routing_available must be POSITIVELY confirmed — a failed /health (or
+    // an older server without the field) should not leave routing controls
+    // offered on a deployment that can't honor them.
+    window.RAG.loadHealth().then((h) => { setPagesAvailable(!!h.pages_available); setDemoAvailable(!!h.demo_available); setRoutingAvailable(h.routing_available === true); });
   }, []);
 
   // apply tweaks → CSS
@@ -277,7 +288,15 @@ function App() {
   }, [t.accent, theme]);
   useEffect(() => {document.documentElement.setAttribute("data-density", t.density);}, [t.density]);
   useEffect(() => {document.documentElement.setAttribute("data-answerfont", t.answerFont);}, [t.answerFont]);
-  useEffect(() => {setLayout(t.defaultLayout); /* eslint-disable-next-line */}, [t.defaultLayout]);
+  // Apply the layout tweak only when it actually changes: running on mount
+  // would clobber the persisted sr-layout choice with the tweak default. The
+  // tweak speaks "focus"; the layout state machine speaks "single".
+  const layoutTweakFirst = useRef(true);
+  useEffect(() => {
+    if (layoutTweakFirst.current) { layoutTweakFirst.current = false; return; }
+    setLayout(t.defaultLayout === "focus" ? "single" : t.defaultLayout);
+    /* eslint-disable-next-line */
+  }, [t.defaultLayout]);
 
   const crumb = CRUMB[tab];
   const stats = { papers: papers.length, figures: figures ? figures.length : 0 };
@@ -308,11 +327,16 @@ function App() {
         </div>
 
         <div className="view">
-          {tab === "chat" && <ChatView settings={settings} set={set} layout={layout} apiKey={apiKey} model={model} papers={papers} figures={figures} pagesAvailable={pagesAvailable} demoAvailable={demoAvailable} routingAvailable={routingAvailable} onNeedKey={(reason) => setKeyModalOpen(reason || "quota")} />}
+          {/* ChatView stays mounted across tab switches — unmounting would
+              destroy the conversation while the chat itself points users at
+              the Papers and Figures tabs. */}
+          <div style={{ display: tab === "chat" ? "contents" : "none" }}>
+            <ChatView settings={settings} set={set} layout={layout} apiKey={apiKey} model={model} papers={papers} figures={figures} pagesAvailable={pagesAvailable} demoAvailable={demoAvailable} routingAvailable={routingAvailable} onNeedKey={(reason) => setKeyModalOpen(reason || "quota")} />
+          </div>
           {tab === "inspection" && <InspectionView settings={settings} apiKey={apiKey} model={model} papers={papers} pagesAvailable={pagesAvailable} routingAvailable={routingAvailable} />}
           {tab === "papers" && <PapersView setTab={setTab} papers={papers} figures={figures} />}
           {tab === "figures" && <FiguresView figures={figures} />}
-          {tab === "why" && <WhyView setTab={setTab} />}
+          {tab === "why" && <WhyView setTab={setTab} routingAvailable={routingAvailable} />}
         </div>
       </main>
 

web/app/chat.jsx

@@ -190,9 +190,12 @@ function RetrievalPanel({ turn, highlight, settings, paperTitle, routingAvailabl
     );
   }
   const cands = turn.candidates;
-  // Rerank scores are logits (often > 1); scale bars to the set's best score
-  // so they stay comparative instead of all clamping to 100%.
-  const maxScore = cands.reduce((m, c) => Math.max(m, c.score || 0), 0);
+  // Rerank scores are logits (any sign, any magnitude) — min-max scale the
+  // bars within the set so they stay comparative; all-negative sets would
+  // otherwise render every bar empty.
+  const scores = cands.map((c) => c.score || 0);
+  const sMax = Math.max(...scores), sMin = Math.min(...scores);
+  const relScore = (s) => (sMax === sMin ? 1 : (s - sMin) / (sMax - sMin));
   const vis = cands.filter((c) => c.kind === "visual");
   const total = cands.length;
   const visShare = total ? Math.round((vis.length / total) * 100) : 0;
@@ -217,7 +220,7 @@ function RetrievalPanel({ turn, highlight, settings, paperTitle, routingAvailabl
             </div>
           ) : (
             <div className="route-card">
-              <div className="gate"><RoutePill route={turn.route || "text"} /><span className="cand-src" style={{ marginLeft: "auto" }}>mode: {settings.route}</span></div>
+              <div className="gate"><RoutePill route={turn.route || "text"} /><span className="cand-src" style={{ marginLeft: "auto" }}>mode: {turn.mode || settings.route}</span></div>
               <div className="route-bars">
                 <div className="rb"><span className="lbl">text</span><div className="scorebar"><i style={{ width: txtShare + "%" }}></i></div><span className="pct">{txtShare}%</span></div>
                 <div className="rb"><span className="lbl">visual</span><div className="scorebar visual"><i style={{ width: visShare + "%" }}></i></div><span className="pct">{visShare}%</span></div>
@@ -235,11 +238,11 @@ function RetrievalPanel({ turn, highlight, settings, paperTitle, routingAvailabl
               <div key={c.chunk_id || i} className={"cand row-clickable" + (hl ? " hl" : "")}
                 onClick={() => setPageItem(c)} title="View source region on page">
                 <div className="cand-top">
-                  <span className={"cand-num" + (c.kind === "visual" ? " visual" : "")}>{c.kind === "visual" ? "IMG" : (num || "·")}</span>
+                  <span className={"cand-num" + (c.kind === "visual" ? " visual" : "")}>{num || (c.kind === "visual" ? "IMG" : "·")}</span>
                   <span className="cand-src">{c.paper} · p.{c.page}</span>
                   <span className="cand-score">{c.score.toFixed(3)}</span>
                 </div>
-                <ScoreBar score={maxScore > 0 ? Math.min(c.score / maxScore, 1) : 0} kind={c.kind} />
+                <ScoreBar score={relScore(c.score || 0)} kind={c.kind} />
                 {c.text && <div className="cand-quote">{previewQuote(c.text)}</div>}
                 <div className="cand-meta">
                   <span className="tag">{previewQuote(paperTitle(c.paper), 30)}</span>
@@ -311,28 +314,37 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
     return next;
   });
 
+  // Incremented by New chat: an in-flight ask compares its captured value and
+  // stops writing, so a zombie stream can't leak into the next conversation.
+  const runSeq = useRef(0);
+
   const ask = useCallback(async (q) => {
     if (busy) return;
+    const myRun = ++runSeq.current;
+    const live = (fn) => { if (runSeq.current === myRun) fn(); };
+    const upd = (patch) => live(() => updateLast(patch));
     setHighlight(null);
     setStatus("");
     setBusy(true);
 
     const priorTurns = turnsRef.current
-      .filter((t) => t.role === "user" || (t.role === "assistant" && t.answer))
+      // Error and notice turns are UI copy ("add your key…"), not assistant
+      // answers — feeding them to condense/generation pollutes the history.
+      .filter((t) => t.role === "user" || (t.role === "assistant" && t.answer && !t.error && !t.notice))
       .map((t) => ({ role: t.role, text: t.role === "user" ? t.text : t.answer }));
 
     setTurns((ts) => [
       ...ts,
       { role: "user", text: q },
-      { role: "assistant", q, answer: "", streaming: true, candidates: null, citations: [], route: null },
+      { role: "assistant", q, answer: "", streaming: true, candidates: null, citations: [], route: null, mode: settings.route },
     ]);
 
     try {
       // Agentic search can't run keyless: the server-side agent spends the
       // caller's OpenRouter key. Stop before retrieval with a notice instead
       // of letting the thrown error render as a generic failure.
       if (settings.route === "agentic" && !(apiKey && apiKey.trim())) {
-        updateLast({
+        upd({
           answer: "Agentic search runs a search agent server-side on your OpenRouter key, so it needs one — add yours (top-right) to try it. The standard retrieval modes work without a key.",
           streaming: false,
           notice: true,
@@ -359,10 +371,10 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
         } else if (demoAvailable) {
           setStatus("Condensing the follow-up into a search query…");
           searchQuery = await window.RAG.condenseDemo(priorTurns, q);
-          setStatus("");
+          live(() => setStatus(""));
         }
       }
-      updateLast({ searchedFor: searchQuery });
+      upd({ searchedFor: searchQuery });
 
       // Retrieve.
       const t0 = performance.now();
@@ -373,17 +385,23 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
         paperId: settings.paper || "",
         dci: settings.route === "agentic",
         apiKey,
-        onStatus: setStatus,
+        onStatus: (s) => live(() => setStatus(s)),
       });
-      setStatus("");
+      live(() => setStatus(""));
       const tRetrieve = performance.now() - t0;
       const candidates = results.map(toCand);
       // With no router on the deployment the route label is always "text" —
       // noise, not information. Suppress the per-message pill there.
-      updateLast({ candidates, route: routingAvailable === false ? null : window.RAG.routeLabel(routing), routing, trace });
+      // Agentic turns are their own path (DCI), not a router decision; the
+      // pill should say so rather than defaulting to "text route".
+      upd({
+        candidates,
+        route: settings.route === "agentic" ? "agentic" : routingAvailable === false ? null : window.RAG.routeLabel(routing),
+        routing, trace,
+      });
 
       if (results.length === 0) {
-        updateLast({ answer: "No chunks retrieved. The corpus may not cover this query.", streaming: false, latencyMs: Math.round(tRetrieve) });
+        upd({ answer: "No chunks retrieved. The corpus may not cover this query.", streaming: false, latencyMs: Math.round(tRetrieve) });
         return;
       }
 
@@ -393,8 +411,8 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
       // retrieval with the bring-a-key notice.
       const hasKey = !!(apiKey && apiKey.trim());
       if (!hasKey && !demoAvailable) {
-        setStatus("Add your OpenRouter key (top-right) to generate a cited answer.");
-        updateLast({
+        live(() => setStatus("Add your OpenRouter key (top-right) to generate a cited answer."));
+        upd({
           answer: "Retrieved the chunks shown on the right. Add your OpenRouter key (top-right) to generate a cited answer from them.",
           streaming: false,
           notice: true,
@@ -406,10 +424,10 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
       // The demo chain is all vision-capable models, so keyless turns always
       // attach page images when pages are served.
       const useImages = pagesAvailable && (hasKey ? window.RAG.supportsVision(model) : true);
-      if (useImages) setStatus(hasKey ? "Reading the retrieved page images…" : "Free demo model is reading the retrieved pages…");
+      if (useImages) live(() => setStatus(hasKey ? "Reading the retrieved page images…" : "Free demo model is reading the retrieved pages…"));
       const messages = await window.RAG.buildMessages(priorTurns, q, results, useImages, figures);
       const tGen = performance.now();
-      const onDelta = (delta) => updateLast((prev) => ({ answer: prev.answer + delta }));
+      const onDelta = (delta) => upd((prev) => ({ answer: prev.answer + delta }));
       const { text, usage } = hasKey
         ? await window.RAG.streamChat(apiKey, model, messages, onDelta)
         : await window.RAG.streamDemoChat(messages, onDelta);
@@ -426,7 +444,7 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
         const pages = c ? c.page_numbers || [] : [];
         return { n: i + 1, id, paper: c ? c.paper_id : id, page: pages[0], quote: c ? previewQuote(c.text) : null, kind: c && c.source === "visual" ? "visual" : "text" };
       });
-      updateLast({
+      upd({
         answer: newText,
         streaming: false,
         citations,
@@ -439,33 +457,41 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
         // The shared free quota ran out for today: hand off to the key modal
         // instead of rendering it as a failure — retrieval still worked.
         onNeedKey && onNeedKey();
-        updateLast({
+        upd({
           answer: "Today's free demo answers are used up, but retrieval still works — the chunks are on the right. Add your own OpenRouter key (top-right) to keep generating answers.",
           streaming: false,
           notice: true,
         });
       } else if (err && (err.code === "demo_down" || err.code === "stream_error")) {
         // Upstream model failure, not the visitor's fault — say so without
-        // blaming a key they may not even have.
-        updateLast({
-          answer: `${(err && err.message) || "The model provider failed."} Retrieval still worked — the chunks are on the right.`,
+        // blaming a key they may not even have, and keep any partial answer.
+        upd((prev) => ({
+          answer: prev.answer
+            ? `${prev.answer}\n\nGeneration stopped early: ${(err && err.message) || "the model provider failed."}`
+            : `${(err && err.message) || "The model provider failed."} Retrieval still worked — the chunks are on the right.`,
           streaming: false,
           notice: true,
-        });
+        }));
       } else {
-        updateLast({
-          answer: `Request failed: ${(err && err.message) || err}. Either the server isn't reachable, or your OpenRouter key is invalid.`,
+        // Keep whatever streamed before the failure — wiping a half-answer is
+        // worse than showing it with an honest interruption note.
+        upd((prev) => ({
+          answer: prev.answer
+            ? `${prev.answer}\n\nGeneration interrupted: ${(err && err.message) || err}`
+            : `Request failed: ${(err && err.message) || err}. Either the server isn't reachable, or your OpenRouter key is invalid.`,
           streaming: false,
-          error: true,
-        });
+          error: !prev.answer,
+          notice: !!prev.answer,
+        }));
       }
-      setStatus("");
+      live(() => setStatus(""));
     } finally {
-      setBusy(false);
+      live(() => setBusy(false));
     }
   }, [busy, apiKey, model, settings, figures, pagesAvailable, demoAvailable, routingAvailable, onNeedKey]);
 
-  const newChat = () => { setTurns([]); setHighlight(null); setBusy(false); setStatus(""); };
+  // Bumping runSeq orphans any in-flight ask: its guarded writes become no-ops.
+  const newChat = () => { runSeq.current += 1; setTurns([]); setHighlight(null); setBusy(false); setStatus(""); };
   // The retrieval panel is hidden in focus layout and at phone width, so a
   // highlight there would be invisible — open the source-page modal instead.
   // Page-image citations always open the modal: they have no panel row.
@@ -476,11 +502,15 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
       setPageItem({ chunk_id: cit.id, paper: cit.paper, page: cit.page, pages: [cit.page], kind: "visual", bbox: null, text: "", page_cite: true });
       return;
     }
+    // The panel only ever shows the LAST turn's evidence, so a highlight is
+    // wrong for citations in older messages — open the modal for those too.
     const panelHidden = layout === "single" || (window.matchMedia && window.matchMedia("(max-width: 760px)").matches);
-    if (panelHidden && cit && msg.candidates) {
+    const isLastTurn = msg === lastAssistant;
+    if ((panelHidden || !isLastTurn) && cit && msg.candidates) {
       const cand = msg.candidates.find((cd) => cd.chunk_id === cit.id);
       if (cand) { setPageItem(cand); return; }
     }
+    if (!isLastTurn) return;
     setHighlight(tag);
   };
   const openFig = (cand) => setPageItem(cand);
@@ -515,7 +545,7 @@ function ChatView({ settings, set, layout, resetSignal, apiKey, model, papers, f
                   <div className="msg msg-user rise" key={i}><div className="bubble">{t.text}</div></div>
                 ) : (
                   <AiMessage key={i} msg={t} onCite={(tag) => onCite(tag, t)} onFig={openFig} paperTitle={paperTitle}
-                    pendingLabel={status || undefined} />
+                    pendingLabel={status || (routingAvailable === false ? "retrieving…" : undefined)} />
                 )
               )}
             </div>