Allow one-shot document attachments in offline Chat mode

Local (offline) Chat mode previously hard-blocked any document
attachment outright, on an earlier real test report that the full
12,000-char server-side cap produced garbled, unusable replies from
the tiny local 1B model. Re-enabled with a much smaller cap
(MAX_ONE_SHOT_DOC_CHARS_LOCAL, 4000 chars combined) instead of
removing the safeguard entirely, so a document can't crowd out the
system prompt and conversation history. Multi-file selection already
combines into one attachment for this path, so it works there too.
Images remain blocked in local mode — that's a real capability gap
(text-only engine), not a size-tunable limit.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

index.html

@@ -374,7 +374,7 @@
       <div class="cap-title">🖥️ Local AI mode — <span class="info-term" id="localLimitedTerm" tabindex="0" role="button" aria-expanded="false" aria-describedby="localLimitedTooltip">limited<div class="tooltip-bubble" id="localLimitedTooltip" role="tooltip">
         <ol>
           <li><strong>Smaller model</strong> — it's a 1B-parameter model, chosen specifically to run smoothly on phones and slower devices (WebGPU memory is often tight on mobile). A real capability gap versus the server's best ("Smart", 8B), but a much smaller, faster download than a bigger model would be.</li>
-          <li><strong>No images, no one-off documents</strong> — it's text-only, and can't analyze a document in a single pass. Switch to RAG mode to work with a document offline instead — that's fully supported, indexed and searched entirely on your device.</li>
+          <li><strong>No images</strong> — it's text-only. One-off documents in Chat mode are read, but truncated much shorter than on the server (this tiny model's usable context is far smaller) — for longer documents, RAG mode works better offline: fully supported, indexed and searched entirely on your device.</li>
           <li><strong>No live data</strong> — the server injects real-time weather, news, currency rates, holidays, and general web facts into its answers. Local mode has none of that — it's just the raw model guessing from its own training, so it'll be wrong or evasive on anything time-sensitive ("what's the weather," "who won yesterday," etc.).</li>
           <li><strong>No date awareness</strong> — the server also tells the model today's actual date; local mode doesn't get that either, so it may not know what day it is.</li>
           <li><strong>Speed/quality depend on your device</strong> — no server-grade GPU behind it, just whatever's in your laptop/phone.</li>
@@ -767,6 +767,13 @@ const MAX_ATTACHMENT_BYTES = 4 * 1024 * 1024; // 4MB
 // mode is to not need a single-prompt size ceiling at all, so this stays
 // scoped to pendingOneShotDoc specifically, not a general document limit.
 const MAX_ONE_SHOT_DOC_CHARS = 12000;
+// The local engine (Llama-3.2-1B) runs with a much smaller practical
+// context window than the server model, and a smaller, far less capable
+// model besides — the full 12,000-char cap was tried here before and
+// produced unusable, garbled replies. A tighter cap doesn't guarantee
+// good output from a 1B model, but it keeps a document from crowding out
+// the system prompt and conversation history entirely.
+const MAX_ONE_SHOT_DOC_CHARS_LOCAL = 4000;
 
 let pendingImage = null; // { base64, name }
 // A document attached while in "Chat" mode — a one-shot analysis, folded
@@ -903,14 +910,17 @@ fileInput.addEventListener('change', async () => {
   if (!docs.length) { fileInput.value = ''; return; }
 
   if (attachmentMode === 'rag') {
+    // Works the same whether the engine is server or local — RAG mode
+    // never folds full document text into a message at all, it retrieves
+    // small, relevant excerpts per question instead, so it isn't subject
+    // to the tiny local model's context-window limits the way a one-shot
+    // dump is.
     await ingestDocumentsForRag(docs);
-  } else if (localModeActive) {
-    alert("Local AI mode can't analyze a one-off document — switch to RAG mode to index it offline, or switch back to the server.");
-    if (!images.length) clearAttachment();
   } else {
     try {
       const texts = await Promise.all(docs.map(async (f) => `[Attached file: ${f.name}]\n\n${await extractFileText(f)}`));
-      const combined = texts.join('\n\n').trim().slice(0, MAX_ONE_SHOT_DOC_CHARS);
+      const cap = localModeActive ? MAX_ONE_SHOT_DOC_CHARS_LOCAL : MAX_ONE_SHOT_DOC_CHARS;
+      const combined = texts.join('\n\n').trim().slice(0, cap);
       pendingOneShotDoc = { text: combined, name: docs.length === 1 ? docs[0].name : `${docs.length} files` };
       showAttachmentPreview(pendingOneShotDoc.name);
     } catch (err) {
@@ -2786,17 +2796,18 @@ async function requestReply(imageBase64) {
   const userMessageText = [...history].reverse().find((m) => m.role === 'user')?.content || '';
 
   if (localModeActive) {
-    // Local mode still can't handle images or a one-shot document dump
-    // (a chat-mode attachment folded into the message as a "[Attached
-    // file: ...]" block with the full extracted text) — the tiny local
-    // model isn't remotely equipped for that (real report: it produced
-    // total gibberish when handed a document this way). RAG-mode
-    // documents are fine now, though: those get retrieved from the
-    // offline (IndexedDB) index inside requestLocalReply itself, not
-    // stuffed into the message here at all.
-    if (imageBase64 || /\[Attached file:/.test(userMessageText)) {
+    // Local mode still can't read images — the engine loaded here is
+    // text-only, not a technical choice that could be relaxed by capping
+    // input size the way the one-shot document case below was. A
+    // one-shot document dump WAS blocked here too, on an earlier real
+    // test report of unusable, garbled replies from the tiny 1B model —
+    // now allowed again, but capped much tighter for local mode
+    // specifically (MAX_ONE_SHOT_DOC_CHARS_LOCAL) than the server path's
+    // budget, to keep a document from crowding out the system prompt and
+    // conversation history entirely.
+    if (imageBase64) {
       loadingBubble.parentElement.classList.remove('loading');
-      loadingBubble.textContent = "Local AI mode can't read images or one-off documents — remove the attachment, switch to RAG mode to index it offline, or switch back to the server.";
+      loadingBubble.textContent = "Local AI mode can't read images — remove the attachment, or switch back to the server.";
       scrollToBottom();
       return;
     }