fix(rabbitmq): recover consumer from channel death and broker cancel

[gitpancake]

Problem

image-engine sat online for ~27h with zero consumers on flash-engine.flash-received and a 17,240 message backlog. Process healthy, TCP to RabbitMQ ESTABLISHED, no log output, images_processed_total = 1 over two days. dataimpulse proxy and Pinata were both fine — the consumer had silently unsubscribed and could not recover without a full restart.

Cause

FlashcastrConsumer only recovered from connection-level close. Three paths left it subscribed to nothing:

• channel close with a live connection — handler only logged. A broker-side channel error unsubscribed the consumer permanently.
• broker-side consumer cancel — amqplib delivers null to the consume callback; swallowed by the if (!msg) return guard.
• hung AMQP handshake — connect() had no deadline. A socket that opens but never completes the handshake parks the reconnect loop forever with reconnecting stuck true.

Fix

• channel close and broker cancel both trigger reconnect()
• connectWithTimeout() races an explicit 20s handshake deadline (amqplib's timeout only covers TCP connect), and closes a late-arriving connection rather than leaking it
• reconnect() calls teardown() first — detaches listeners before closing so close events don't re-enter the loop
• new closing flag suppresses recovery during intentional shutdown so close() doesn't race the reconnect loop

Publisher was checked and left alone: it is lazy-connect, nulls its channel on close, and reconnects on next publish.

Verification

• consumer.ts typechecks clean standalone
• Production restart confirmed the diagnosis: consumer re-registered, backlog drained at ~254 msg/min, ipfs_failures_total 0, DLQ 0 — messages genuinely processed, not nacked
• ⚠️ no unit test — libs/rabbitmq has no specs or vitest config; adding one means bootstrapping test infra for the lib
• ⚠️ full tsc project build and nx build are broken on main already (stale dist/ TS6305, unresolvable @nx/next/plugin) — pre-existing, unrelated to this change

The recovery paths only fire on real channel death, so this deploys unexercised. Worth watching consumer count for a day.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HioDdCpmbbk6vrmkLt9Psg