docs: ARCHITECTURE.md has drifted from the relay code (search_index_tx, fan-out tiers, rate limiting)

Start with the three affected sections in ARCHITECTURE.md and verify each claim against the referenced relay, search, subscription, connection, state, and config files. Update the search-index description, fan-out tier count, and rate-limiting gap while leaving the other listed gaps intact. Done means ARCHITECTURE.md matches the cited code at the referenced commit.

Observed at commit 63496cc1d4c6f1b7c613801bdcc694169dcf391a (v0.5.2-5-g63496cc1). Three places where ARCHITECTURE.md describes a relay that no longer exists — each contradicted by an explicit statement in the code itself:

1. search_index_tx / separate FTS worker is gone

ARCHITECTURE.md still lists a search_index_tx field, but indexing is now the row insert itself: events.search_tsv is a GENERATED ALWAYS AS (to_tsvector(...)) STORED column with a GIN index (crates/buzz-search/src/lib.rs:5-15 — "every row write is the index update — there is no separate indexer, no mpsc queue, no reindex job"). The relay confirms in a comment at crates/buzz-relay/src/handlers/event.rs:502-506: "The old Typesense index_event worker and its search_index_tx mpsc are gone."

2. "Three-Tier Fan-Out" is now five tiers

ARCHITECTURE.md (~L294) describes three subscription index tiers; crates/buzz-relay/src/subscription.rs (~L50-58) defines five community-keyed maps: channel_kind_index, channel_wildcard_index, global_kind_index, global_p_kind_index, global_wildcard_index.

3. "No rate-limiting implementation" is stale

The known-gaps section (~L816-827) says rate limiting exists only as a trait with a test stub. WS admission rate limiting is now implemented and enforced on the hot path: enforce_ws_admission (crates/buzz-relay/src/connection.rs:594-650) runs Redis sliding-window checks before dispatch on every EVENT/REQ/COUNT, with four live limiters constructed in AppState::new (crates/buzz-relay/src/state.rs:584-597) and env-tunable limits (crates/buzz-relay/src/config.rs:301-314).

The other gaps listed there (WF-07 NotImplemented actions, WF-08 approval gates — see #3525 — missing huddle recording, no sqlx offline cache) still check out; it's specifically these three claims that have drifted.

Found while building code-derived documentation of the platform (with AI assistance) — the drift actively misled a docs-first pass until we re-verified against source. Happy to send a docs PR updating the three sections if that's welcome.