A document's request history should go with the document (and a memory's with the memory) #6098

#5948 deletes a bank's llm_requests together with the bank. delete_document and memory invalidation leave theirs behind. A client that offers its users a real "forget this conversation" can delete the document, its memories and its observations, but the raw prompt and answer of every model call that read the text stay in llm_requests until the retention sweep runs. That covers retain extraction, consolidation and refreshes.

The rows are already addressable:

• retain rows carry metadata.document_id and metadata.memory_ids;
• consolidation rows carry memory_ids and source_memory_ids;
• GET /v1/default/banks/{bank_id}/llm-requests already filters on document_id and memory_id.

Nothing can delete them, though. There is no delete endpoint, and no admin command.

Proposal, in three independent parts:

1. A document's request history goes with the document. delete_document(bank_id, document_id) also runs the following, in the same transaction, with $3 the unit ids it is deleting (it already has them), and gated by _llm_requests_persistable() like fix(banks): delete a bank's LLM request history with the bank #5948:

   DELETE FROM llm_requests
   WHERE bank_id = $1
     AND (metadata->>'document_id' = $2
          OR metadata->'memory_ids' ?| $3
          OR metadata->'source_memory_ids' ?| $3)

2. A delete endpoint with the GET's filters. DELETE /v1/default/banks/{bank_id}/llm-requests takes the same filters as the list route (document_id, memory_id, trace_id, operation, scope, start_date, end_date), with at least one required. It runs under a new BankWriteOperation.DELETE_LLM_REQUESTS, so an operation validator can gate it, and returns {"deleted": n}. This lets a client purge the traces of a memory it invalidates rather than deletes.
3. Every row that quotes a memory is reachable by memory_id. Today refresh_mental_model, reflect and rerank rows carry no link. The refresh and reflect paths would call record_source_memory_ids() with the memories their prompt quoted, and the rerank trace would carry the candidate memory ids beside the texts.

Rough size: part 1 is about 15 lines beside the #5948 block. Part 2 is about 80 lines: an engine method reusing list_llm_requests' filter builder, plus a route and a test. Part 3 is about 20 lines. Tracing stays on by default, so debugging keeps working; only the forgotten text goes. We'd be glad to send a PR for any part if that helps.