perf(relay,desktop): bulk latest-message-per-channel lookup for get_channels

[tlongwell-block]

What

get_channels populated last_message_at with one limit:1 filter per channel — N top-1 DB queries per sidebar refresh, re-run every 60s (bounded-concurrent since #1457, but still O(N) queries). This replaces the N filters with one bridge extension filter and one indexed DB round trip. This is the DISTINCT ON follow-up deliberately split out of #1457.

• buzz-db: get_latest_event_per_channel — unnest($channels) CROSS JOIN LATERAL (... ORDER BY created_at DESC, id ASC LIMIT 1) riding idx_events_community_channel_created.
• bridge /query: new latest_per_channel: true extension field (same raw-JSON pattern as before_id / depth_limit / feed_types), handled in a dedicated phase ahead of the catch-all.
• desktop get_channels: N filters → {"kinds":[9,40002],"#h":[...all ids],"latest_per_channel":true}. Rust-only change; no collision with perf(desktop): instant channel switching — non-blocking first paint, persisted snapshots #1452's TS surface.

Correctness

• Same winner per channel. The LATERAL subquery's ordering (created_at DESC, id ASC LIMIT 1) is byte-identical to query_events with limit:1, which is what each per-channel filter executed before. A Postgres-backed equivalence test asserts bulk winners == serial winners, including the same-second id ASC tie-break, soft-delete exclusion, wrong-kind exclusion, and empty-channel omission.
• Same access enforcement. Requested #h channels are intersected with accessible_channels before the query (same outcome as the catch-all's access-scope skip), and every returned event passes the identical four per-event gates: event_in_accessible_channel, filters_match, reader_authorized_for_event, is_author_only_event.
• Why the multi-#h hazard doesn't apply. Plain multi-#h pushdown is rejected in this codebase because quieter channels get silently dropped under a shared limit budget. Here top-1-per-channel is guaranteed by construction (LATERAL per unnest row), so there is no shared budget to starve.
• Kinds semantics follow NIP-01: kinds absent = any kind; empty = match nothing.
• Opt-in only. Filters without latest_per_channel: true are handled exactly as before; the WS REQ path is untouched.

Plan shape

EXPLAIN shows ordered index scans on (community_id, channel_id, created_at DESC, id) under a Limit node on every partition — no sort. LATERAL was chosen over DISTINCT ON specifically because the per-group LIMIT 1 cannot degrade to a full sort as channel count/history grows.

Verified

• Postgres-backed equivalence test: cargo test -p buzz-db latest_event_per_channel -- --ignored → green
• cargo test -p buzz-relay -p buzz-db: 451 + 79 pass, 0 fail (re-run after merging main e42dae3f on top)
• Desktop tauri tests: 864 pass, 0 fail
• cargo clippy --workspace --all-targets -- -D warnings: clean