[tyt2y3]

Follow-up to #47, fixing two problems in the error paths of the per-shard tail anchoring.

Problems

1. Cluster mode broke at startup. Cluster::run assigns all shards to the first node and relies on MOVED errors from the first read to redistribute them (move_shards). The anchoring XREVRANGE runs before the first XREAD and its errors were propagated raw via ?, so a MOVED reply became RedisErr::Moved, which the node loop treats as fatal. A RealTime + Latest (the default) consumer against a Redis cluster died with ConsumerDied whenever its keys did not all live on the first connected node — the common case with several streams. Ask/ClusterDown/MasterDown were similarly fatal here despite being retried on the XREAD path.

2. A transient error at first read silently re-introduced the starvation bug. first_read was cleared before the fallible anchoring loop. If an XREVRANGE failed with an error the node loop retries (Io → reconnect, TryAgain → sleep), the retry skipped the anchoring block entirely, and every still-unanchored shard fell back to sending $ on each XREAD — exactly the quiet-stream message loss #47 fixed, for the consumer's whole lifetime, with nothing surfaced.

Fix

• Anchoring errors now go through the same four-way classification as the XREAD reply: MOVED → move_shards (cluster reassigns; the destination node re-anchors since AddShard re-arms its first_read), Io → reconnect, cluster errors → retry, anything else → send_error so the consumer sees the actual error instead of a generic ConsumerDied.
• first_read is cleared only after every shard is anchored. Error paths return with it still true, and a retry resumes where it left off — already-anchored shards are skipped via id.is_some(), so the loop is idempotent.
• A malformed id from XREVRANGE is also routed through send_error for consistency.

Tests

New unit tests for parse_stream_id (valid/edge/invalid ids), and five new scenarios in realtime-multi.rs on top of the three from #47:

• quiet_history_not_starved_nor_replayed — a quiet stream with pre-existing history anchors at its own tail: no replay, and the message landing in the starvation window is not lost.
• absent_streams_all_delivered — streams that don't exist yet anchor at 0-0; the very first entries are delivered.
• multiple_rounds — the anchor keeps advancing after first delivery across repeated starvation cycles.
• earliest_gets_everything — Earliest is unaffected by anchoring: full history plus starvation-window messages.
• wrong_type_key_errors — a stream key occupied by a plain string (WRONGTYPE on XREVRANGE) surfaces the actual backend error to the consumer. Verified this fails on main (generic ConsumerDied) and passes with the fix. Requires the redis crate's blocking client as a dev-dependency for setup.

All pass locally on both runtime-tokio and runtime-smol against Redis 8. The MOVED and Io-retry paths aren't covered by CI — they need a Redis cluster / fault injection; verified by code inspection against the XREAD path, which handles the identical errors.

🤖 Generated with Claude Code