Chat mode, phase 1: grounded streaming chat over your notes

Entry points (no sixth tab): Ask in the Library header (desktop button,
phone icon), Ask about this deck in the deck menu, Ask a question on
study/homework results (compact result travels as context), and a Chat
row in the desktop sidebar. Phones get a full-screen route with back +
New chat; the tab bar steps aside.

- Client-side retrieval: chunk scoped decks (~800 tokens), BM25 against
  the question + recent turns, top chunks under ~7k tokens, each
  labelled with deck + section. Notes come from the deck's stored
  sourceText (already kept for Regenerate); decks without it fall back
  to generated content and the UI says so.
- Stateless server: /api/chat validates against the CHAT_MODELS
  allowlist, trims the history to ~6k tokens, caps context at ~9k and
  streams token/thinking/done/error NDJSON. /api/chat/extract reduces
  attachments to text through the existing vision step.
- UI: user bubbles + full-width assistant text, streaming with stop,
  auto-scroll that pauses with a jump-to-bottom button, safe markdown
  (React nodes, no HTML injection; KaTeX dynamically imported only when
  a message has math), Thinking row for reasoning, copy/regenerate/
  edit-and-resend, model + context pickers (last model remembered),
  composer with visualViewport keyboard handling, safe-area, 16px
  textarea, Enter-to-send on desktop.
- Usage-limit failures keep the draft and show the shared message with
  Settings/Ollama actions; stopping maps the reader abort to a clean
  stopped state; offline disables sending with a reason.
- Settings: Keep notes text for chat toggle; deck menu: Delete notes
  text. RATE_LIMIT_CHAT + CHAT_MODELS in .env.example and README.
- Tests: chat suite (63 checks, stub Ollama) wired into run-all.
  Full set: 12/12 suites, 487 checks. SW -> freezmo-v32.

Also: tests/model-compare.mjs (diagnostic) benchmarks two models on the
same notes (latency, repair rounds, activity types) so the gpt-oss vs
gemma default can be decided on numbers; defaults unchanged.

app/api/chat/extract/route.ts

@@ -0,0 +1,107 @@
+import { z } from "zod";
+
+import { extractNotes, notesToText } from "@/lib/ai";
+import { authErrorMessage, authErrorKind, getServerConfig, resolveAuth } from "@/lib/config";
+import { friendlyErrorMessage, OllamaError } from "@/lib/ollama";
+import { rejectCrossOrigin, rejectOversizedBody, requestLocale } from "@/lib/http-guard";
+import { clientKey, rateLimit, rateLimitResponse } from "@/lib/rate-limit";
+import { translate } from "@/lib/i18n-dict";
+
+export const runtime = "nodejs";
+export const dynamic = "force-dynamic";
+export const maxDuration = 300;
+
+/** Same ceiling as deck generation: up to 16 resized photos. */
+const MAX_BODY_BYTES = 33554432;
+const MAX_IMAGE_CHARS = 8_000_000;
+
+const requestSchema = z.object({
+  images: z.array(z.string().min(32).max(MAX_IMAGE_CHARS)).max(16).optional(),
+  text: z.string().max(120_000).optional(),
+});
+
+/**
+ * Turns chat attachments (photos, scanned pages, pasted document text) into
+ * plain text with the existing vision pipeline, so any chat model — including
+ * text-only ones — can use them. Never stores anything.
+ */
+export async function POST(request: Request) {
+  const locale = requestLocale(request);
+  const crossOrigin = rejectCrossOrigin(request);
+  if (crossOrigin) return crossOrigin;
+  const oversized = rejectOversizedBody(request, MAX_BODY_BYTES);
+  if (oversized) return oversized;
+
+  const config = getServerConfig();
+
+  const limit = rateLimit(`chat:${clientKey(request)}`, config.rateLimits.chat, config.rateLimits.windowMs);
+  if (!limit.allowed) return rateLimitResponse(limit, locale);
+
+  const auth = resolveAuth(config, request);
+  if ("error" in auth) {
+    return Response.json(
+      { error: authErrorMessage(locale, auth.error), kind: authErrorKind(auth.error) },
+      { status: 401 },
+    );
+  }
+
+  let payload: unknown;
+  try {
+    payload = await request.json();
+  } catch {
+    return Response.json({ error: translate(locale, "server.bodyNotJson") }, { status: 400 });
+  }
+
+  const parsed = requestSchema.safeParse(payload);
+  if (!parsed.success) {
+    return Response.json(
+      { error: translate(locale, "server.invalidRequest", { detail: parsed.error.issues[0]?.message ?? "" }) },
+      { status: 400 },
+    );
+  }
+
+  const images = (parsed.data.images ?? []).map(stripDataUrlPrefix).filter(Boolean);
+  const text = parsed.data.text?.trim() ?? "";
+  if (images.length === 0 && !text) {
+    return Response.json({ error: translate(locale, "server.needContent") }, { status: 400 });
+  }
+
+  const abort = new AbortController();
+  const timer = setTimeout(() => abort.abort(), config.timeoutMs);
+  try {
+    if (images.length === 0) {
+      // Plain text needs no model call at all.
+      return Response.json({ text: text.slice(0, 48_000) });
+    }
+    const extraction = await extractNotes({
+      model: config.visionModel,
+      repairModel: config.textModel,
+      auth: auth.auth,
+      images,
+      text: text || undefined,
+      signal: abort.signal,
+    });
+    const extracted = notesToText(extraction.notes).slice(0, 48_000);
+    if (!extracted.trim()) {
+      return Response.json({ error: translate(locale, "chat.attachEmpty"), kind: "bad_response" }, { status: 422 });
+    }
+    return Response.json({ text: extracted });
+  } catch (error) {
+    console.error("[api/chat/extract]", error instanceof OllamaError ? `${error.kind}: ${error.message}` : error);
+    return Response.json(
+      {
+        error: friendlyErrorMessage(error, auth.auth.baseUrl, locale),
+        kind: error instanceof OllamaError ? error.kind : "unknown",
+      },
+      { status: 502 },
+    );
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+function stripDataUrlPrefix(value: string): string {
+  const commaIndex = value.indexOf(",");
+  if (value.startsWith("data:") && commaIndex !== -1) return value.slice(commaIndex + 1);
+  return value;
+}