Query page — retrieval and synthesis in two explicit stages

The query page never blurs together "what the vault found" and "what

the model said about it". Stage one retrieves document chunks through

the vault's own gated search — the same search every other surface

uses, bounded by the caller's compartments. Stage two asks the locally

configured model to answer a question over exactly the chunk SET stage

one handed back, and nothing else: the client sends references, not text,

and the vault re-resolves each one itself before the model sees anything.

Each resolved passage is then widened to its enclosing section, so the

model reads more of those same documents than stage one displayed —

inside the same versions, under the same permissions, and the citation

still names the passage that matched. A reference to content the caller cannot reach — whether

fabricated outright or copied from somewhere real but out of reach —

resolves to nothing, the same honest "doesn't cover" answer the vault

gives when nothing was asked for at all. Nothing about that answer

reveals which case actually happened.