Measure what a query costs as the corpus grows

[vmillet-dev]

Part of #21, and deliberately not closing it: the ticket asks for a measurement before anything is built, so this is the measurement and nothing else. The numbers and the reading are in the issue; the decision between its three fixes is not mine to make.

What it adds

cargo bench --bench corpus_size seeds 1 000, 5 000 and 20 000 notes and times each twice: the whole query, and the same query stopped before view::build and serde. The gap between the two is what a lighter wire shape could take away and what a LIMIT could not — which is the question that decides between the ticket's options.

It also prints the serialised payload per size, because half of what #21 is about is what crosses the bridge, and a size is not a duration.

Why a target of its own

commands keeps one corpus size on purpose — its baseline is the table in docs/architecture.md and has to stay comparable from run to run. And seeding 26 000 notes of ~13 kB would make a plain cargo bench unusable. So: [[bench]] name = "corpus_size", run on its own.

Shape

• corpus::build() becomes build_of(notes) with build() calling it at NOTES, so every existing group is untouched.
• run_query moves from commands.rs into corpus.rs: two targets needed it, and a copy in each is the duplication this repository has spent the day removing.
• ⚠️ #![allow(dead_code)] on corpus.rs, with the reason: each bench target compiles its own copy of the module and uses a part of it — commands never calls build_of, corpus_size never reads note_ids.