fix(relay): add a retry hint when shared admission is unavailable - #7405

When Redis admission fails, the relay sends rate-limited: shared admission unavailable without a retry hint. Mobile treats a rate-limited CLOSED with no hint as a ten-second session-wide gate, even if the Redis failure was brief.

Add retry in 1s to this WebSocket rejection. This is a retry delay, not a claim that Redis will recover within a second. Admission still fails closed, configured quota resets keep their existing hints, and EVENT/REQ/COUNT rejections keep their existing correlation fields. HTTP responses do not change.

The three regression assertions drive enforce_ws_admission with deliberately unreachable Redis and inspect the actual outbound EVENT, REQ, and COUNT responses. All three failed before the production change. All seven rejection tests pass after it.

Validation beyond the seven rejection tests:

• Full just ci passed on a local checkout combining this PR, fix(mobile): close retired relay handshakes #7386 , and perf(mobile): parallelize independent channel startup reads #7397 .
• All 1,040 relay library tests passed with isolated PostgreSQL 17 and Redis 7 (89 ignored). A mesh echo test timed out on the first run, then passed alone and in the complete repeat. The six earlier media fixture failures passed once PostgreSQL was available.
• All nine relay boot-lifecycle integration tests passed.
• just test remains red at buzz-db::observability_source::p0_pool_acquisitions_use_typed_operation_pairs_without_other, which reports event production path bypasses operation attribution with .fetch_all(pool). The same test fails with the same message from an untouched archive of base commit f038cbbb0. This PR changes none of its source inputs. That failure stops the combined integration run before later packages, so the relay tests above ran separately.