fix(web): friendlier answers for corpus-wide and multi-paper figure asks

Two wording gaps from the answer-quality pass: corpus-meta questions
("how many papers discuss X") got the bare refusal phrase - honest but
unhelpful - and a vague "show me a figure" listed figures from several
papers without saying which paper each came from. The prompt now has
the model explain the top-k limitation and point to the Papers and
Figures tabs for corpus-wide asks, and attribute figures to their
papers when an answer spans more than one.

Author: NorthernLightx

web/app/api.js

@@ -245,6 +245,8 @@
       '- Use only the provided context for factual claims. If the user asks a factual question the context cannot answer, say exactly "Not stated in the provided context." — do not speculate.',
       "- Out-of-domain questions (e.g. about a topic completely unrelated to the chunks) must be refused with the exact phrase above. Do not produce a generic summary of the chunks instead.",
       "- Not every message is a question. If the user is reacting to your previous answer — challenging it, asking what you checked, or making conversation — reply naturally using the prior turns: explain what evidence you used (the chunks and page images you cited) and offer to check something specific. Never use the refusal phrase for such messages.",
+      "- Questions about the corpus as a whole (how many papers, which papers cover a topic) cannot be answered from the few retrieved excerpts you see. Say that plainly and point to the Papers and Figures tabs for corpus-wide browsing — do not use the refusal phrase and do not guess a count.",
+      "- When one answer mentions figures from more than one paper, say which paper each figure belongs to.",
       "- If the user asks to see a figure, plot, or graph and the retrieved chunks don't contain one matching the question, say so plainly — do not claim you cannot display images.",
       "- Cite specific chunk IDs when making factual claims by wrapping the literal id in square brackets. Example: if a chunk header is [2604.22753v1::p5::c24], cite it as [2604.22753v1::p5::c24] — NOT [chunk_id 2604.22753v1::p5::c24] and NOT [chunk 24]. Use only ids that appear in the provided context.",
       "- Attached page images are labeled with their own id, e.g. [page image 2604.22753v1::p5::page]. When you describe what a figure, plot, table, or diagram shows based on looking at a page image, cite that page id (e.g. [2604.22753v1::p5::page]) — not a text chunk. Cite text chunk ids only for claims supported by the chunk text itself.",