Chat embeddings: retrieve heading/harvest chunks, not whole-node blobs

Task

Chat and import embeddings are still one vector per node, often the whole page. That is why AI Exchange (and similar long markdown) comes back as a single blob. After #1075 returns top-k hits, those hits are still coarse unless we store section-level embeddings_content.

Harvester already has markdown_heading / Docling strategies (application/utils/harvester/chunker.py). Chat does not read harvest_input / chunk records; it reads the embeddings table.

Success criteria (all must pass)

• Document the current path: PromptHandler.generate_embeddings → embeddings.embeddings_content (whole node / fetched URL) vs Module A chunks
• Design (in-issue or short RFC comment): either (a) embed heading-sized slices into embeddings with a stable embeddings_url fragment, or (b) chat retrieval may read harvested chunk text when a node has OIE chunks — pick one, do not run two indexes
• AI Exchange (and other long MD standards) no longer injects an entire page as the only context for a narrow question
• embeddingsUrl / citations still resolve to a section, not just the document root
• Heroku: no torch / LlamaIndex on prod slug. Chunking for import may stay on workers / requirements-dev.txt
• Tests for the chosen slice boundary (heading vs fixed-size) and for citation URL fragments
• make lint / make mypy / targeted tests green

Context

• Depends on / follows #1075 (multi-hit chat is wasted if each hit is the whole AIX page)
• Harvester: application/utils/harvester/chunker.py, docling_chunker.py, repos.yaml
• Embed path: application/prompt_client/prompt_client.py (generate_embeddings, normalize_embeddings_content)
• Related: #962, OIE Module A

Constraints

• In scope: finer-grained chat/import embeddings for long standards
• Out of scope: replacing chat with a notebook product; PDF-only RAGFlow; changing Module C’s CRE hub identity
• Do not duplicate Module A’s LlamaIndex Docling onto the web dyno

Start with application/prompt_client/prompt_client.py, especially generate_embeddings and normalize_embeddings_content, then compare the existing strategies in application/utils/harvester/chunker.py and docling_chunker.py. Choose one indexing approach and document the design before implementing it. Done means section-level retrieval and citation fragments work for long Markdown standards, with boundary and URL-fragment tests plus make lint, make mypy, and targeted tests passing.