Fix Buzz live fan-out across multiple channels

[olegbrok]

What changed

• Open one long-lived Buzz REQ per authorized channel, each with a unique subscription ID, kinds: [9, 20002], one #h value, and no wire since.
• Keep the shared subscription floor enforced client-side before cache, authorization, or durable delivery.
• Accept events from every per-channel subscription and wait for EOSE from every main subscription before marking the poller or persisted health connected.
• Rebuild the complete per-channel subscription map on every reconnect.
• Extend the relay regression rig to suppress live fan-out for either known-bad shape: a filter carrying since or a filter carrying multiple #h values.

Root cause and impact

The production relay returns stored events and EOSE for multi-channel filters but does not live-fanout new events to them. Release 26.08.011 removed the independent wire-since trigger, but its single main subscription still carried both configured channels in one #h array. Splitting the live stream by channel removes both relay triggers while preserving the client-side replay floor and exactly-once durable delivery.

The active heartbeat probe intentionally remains one aggregate, bounded REQ with multi-#h, since, and limit: 1: tonight's controlled matrix proved that stored-query/EOSE path works, and the probe does not rely on live fan-out.

Validation

• uv run --frozen pytest -q tests/test_buzz_inbound_poller.py — 7 passed
• uv run --frozen pytest -q tests/test_buzz_*.py — 62 passed
• uv run --frozen pytest -q tests/test_agent_registry.py — 117 passed
• uv run --frozen ruff check . — passed
• git diff --check — passed

The repository-wide local run reached 829 passed / 2 skipped before being stopped in an unrelated environment-bound API test: the live daemon already owned port 8890, so the shared-MCP fixture failed to bind and the manual-dream test escaped its mocked runner into tmux. GitHub CI remains the clean full-suite gate for this exact head.