The model reads the whole thought, and the citation still names the chunk

A document is cut into chunks so it can be indexed, and how long a chunk

may be is decided by the embedding model's input limit — not by how much

an answer needs. Those are different numbers, and left alone the second

loses: a heading section longer than one chunk is cut mid-thought and

nothing in the answer says it was cut.

So when a chunk wins the ranking, the prompt receives the whole section it

came from. Nothing is re-indexed and nothing is re-stored; the vault reads

more of a document it had already decrypted for that caller. It changes

which offsets are requested, not what was written, which is why it reaches

documents already in the vault rather than only new ones.

Two properties are load-bearing and neither is negotiable. A widened span

is clamped to the version it came from, so it can never join text from

another document — that would be a fabrication with a citation attached.

And the widened span travels alongside the matched chunk's own locator

rather than replacing it, so a citation names what the vault claims it

read. Widening changes what the model reads; it never changes what the

vault says it read.