[unRekable]

What changed and why

The roadmap's message search starts here: the cache is searched, over the socket and from the
text shell. A query returns the newest hits, each with where it is and enough of it to choose
by. A later change puts the same hits in the switcher and lands on the message; after that the
networks themselves are asked (the spec's N12, staged).

• Request::SearchMessages (filters optional), Response::MessageHits and MessageHit,
  added by the protocol's optional-field rules, so an older frame still reads.
• Store::search_messages: LIKE over the message text, joined to the conversation for its
  title, newest first, clamped to PAGE_LIMIT.
• tinyclient --text --search <query>: the hits as lines, so the change is usable without the
  window.
• architecture.md's request table and the README's not-built-yet line follow the change.

Decisions taken in the change, with their reasons

• Message requests are not searched, nor hidden conversations, nor messages that never
  went. The product keeps requests out of everything that counts and promises "opening one
  reads nothing"; a search that surfaced one would get around that. One SQL condition, tested.
• No paging. Spec N12 is per keystroke; who needs more narrows the query or the
  conversation. A cursor can come later as an optional field, which costs nothing then.
• LIKE, not FTS5. Measured on a 40,000-message store (20 conversations x 2,000, what
  retention holds): 8 to 16 ms per search in a release build, holding the store's lock. It
  grows with the cache, so an honest ceiling exists. FTS5 would raise the schema version and
  with it delete every user's cache, and would add index bookkeeping to ingest, deletes, prune
  and forget. When a real, large cache shows the need, that change can stand on its own.

Test

• Store: newest-first order, conversation and network filters, hidden, request and pending
  excluded, case-insensitive, %/_/\ mean themselves, and the snippet window with its
  ... markers.
• Protocol: the exact wire format of the request and the answer.
• Measured against the seeded demo over the socket: tinyclient --text --search lichen
  returns the two hits with titles and snippets.

Numbers

40,000 messages, release build, the store's lock held while searching: 15.5 ms (4,000 hits),
10.4 ms (common word), 7.9 ms (full scan, no hits). Grows with the cache; the fix, when a real
cache needs it, is FTS5 as its own change.

Tested on Windows with --no-default-features (no protoc here for the signal feature);
cargo check -p tinyclient --features gui passes. The chatd suites that fail on this machine
(discord's spawn of true, the mediad process lookups in e2e, the release-json test in
tinyclient) fail the same on clean main; the blobs picker test is flaky under parallel load
and passes alone.

[itz4blitz]

Cached message search matches the protocol and store rules from review: bound LIKE with escaping, requests/hidden/pending excluded, and the README keeps the Signal Note to Self paragraph. macOS, Linux, and Windows CI are green on 4433107.