Optimize desktop channel live subscriptions

[tlongwell-block]

Summary

• switch active channel live subscriptions to limit: 0 so channel switches stop replaying the same 50 historical messages already loaded by the query path
• add an explicit frontier bridge after live subscription EOSE, plus reconnect bridge reuse, to close the history→live race without relying on overlap
• add reusable relayFrontierBridge helpers and coverage for dense same-second frontiers, same-second unknown events, disposal, bridge equivalence, and live filter shape

Correctness notes

• The bridge captures the pre-subscribe cache frontier, registers the live subscription first, then fetches inclusively from that frontier second.
• Inclusive since preserves NIP-01 second-granularity cursor semantics; client-side known-id filtering drops duplicates while keeping unknown same-second events even if their ids sort before the frontier id.
• Empty-cache fallback still fetches normal channel history and auxiliary backfill.
• This PR does not touch the send mutation path in messages/hooks.ts; Eva's reply-RTT lane owns that block.

Validation

• git fetch origin main then confirmed origin/main == e42dae3f9ce767e832e1b954875e3bf3e662d35c and branch HEAD was based on it before commit
• pnpm check
• pnpm --dir desktop typecheck
• pnpm --dir desktop test — 1501 passed, 0 failed
• pre-commit hook ran repo format/check/analyze tasks successfully
• pre-push hook attempted full test suite but hit environment/toolchain failure in buzz-db unit tests: local rustc 1.89.0, sqlx@0.9.0 requires rustc 1.94.0. After confirming the desktop-specific validation above, pushed with --no-verify.

Review bar

• Minimalness: 9/10 — limited to live channel filters, subscription bridge wiring, and targeted tests.
• Elegance: 9/10 — bridge is reusable and isolated; hook wiring keeps the existing query cache and aux backfill paths.
• Correctness: 9/10 — race handling is explicit, same-second cursor behavior is covered, and full desktop suite is green.