A command carrying an `idempotencyKey` is now executed at most once: a repeat is
answered with the original reply rather than re-run. That is the precondition
every world verb in Phase 12 is waiting on, and it is what let `uo.broadcast`
stop being un-retryable.
The gate sits in BridgeBoot's inbound dispatch, not in each handler, so it covers
every kind including ones a later protocol adds. A command with no key behaves
exactly as it did before, which leaves the admin screens unchanged.
Four rules, each a decision rather than an implementation detail: reserve on
receipt (so a handler that defers is covered, answering `bridge.busy` to a repeat
in flight); a key that has begun is never released, not even when the handler
throws; a replay is stamped with the REPEAT's correlation id, because the
sidecar's reqId is fresh per call and replaying the original would hang the retry;
and the bound is loud, because an evicted key is the guarantee's one hole.
`champ.boss.killed` rides along because a bump costs a release, a bundle and an
operator update on every shard. It fires from EventSink.CreatureDeath, detected
by type so a boss that popped and died inside one sweep is still reported, and it
carries the damage table that exists at the death and nowhere else.
overlay.toml protocol = 6, in this commit rather than a later one.
Co-Authored-By: Claude <noreply@anthropic.com>
BridgeLink owns a TcpClient to 127.0.0.1 and nothing else touches it. Emit() is
called from the Core thread; it enqueues onto a bounded drop-oldest queue and
returns. A link thread drains the queue and reconnects with backoff; a reader
thread parses inbound lines and marshals each to the Core thread via
Timer.DelayCall. An absent, slow, or wedged sidecar therefore cannot stall the
shard, which is the property the rest of the bridge depends on.
Outbound JSON is written by hand into a StringBuilder because it runs on the
Core thread for every event and the measured budget assumes that cost. Inbound
uses JavaScriptSerializer: commands arrive at human rates, so correctness beats
speed, and parsing happens off the Core thread anyway. That needs a
System.Web.Extensions reference.
server.hello is emitted per connection rather than once at ServerStarted. A
sidecar that restarts independently would otherwise never learn which shard it
is attached to. It carries a bootId, stable across reconnects and fresh on every
shard restart, so the sidecar can tell "I reconnected" from "the shard
restarted" and keep or discard its cache accordingly.
Two defects found by testing and fixed before commit:
- Backoff ceiling was 30s, so a sidecar restart cost up to half a minute of
buffering on a loopback socket. Now 5s.
- A stale reader could kill a fresh connection: reader.Join(1s) can time out,
and the old thread's finally block then set the shared _dead flag, possibly
tearing down the connection that had replaced it. Connections now carry an
epoch and a reader only marks dead the one it owned.
Acceptance evidence recorded in docs/PLAN.md §11: boots with no sidecar, buffers
through the outage and drains on connect, round-trips ping/pong on the Core
thread, survives unknown kinds and malformed JSON, and reconnects unattended.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
