Count attempts on the message, not on the consumer, or a restart resets everyone's budget:

const MAX_ATTEMPTS = 5
export async function handle(message: QueueMessage) {  try {    await process(message.body)    await message.ack()  } catch (error) {    if (message.attempts >= MAX_ATTEMPTS) {      await deadLetter.put(message, { reason: String(error) })      await message.ack()      return    }    await message.retry({ delayMs: backoff(message.attempts) })  }}

Exponential with jitter, capped. Without jitter the whole batch that failed together retries together, which is how one bad deploy turns into five identical spikes.

There is one already, in the Dead letter replay script thread. It reads from the same store this writes to, so nothing new is needed here beyond making sure reason is populated: replay filters on it.