[apps][long-polling] make retries configurable & add jitter, exponential backoff and a per-attempt abort

[Ayc0]

Motivation

The dev server's long-poll loop was hardcoded to 10 attempts that fired back-to-back with no delay, and nothing bounded a single attempt — if a connection stalled it just hung. Made the retry behaviour configurable and added the two standard API auto-retry strategies (jitter + exponential backoff) on top.

Changes

• New apps.longPolling config: maxRetries (set to 1 to disable long-poll retrying entirely), jitter, exponentialBackoff, timeoutMs.
• Each attempt now carries an AbortSignal.timeout(timeoutMs). A stalled attempt is abandoned and re-polled against the same receipt rather than failing the action — the receipt stays valid across attempts.
  • timeoutMs defaults to 40s, not 30s: the deadline needs headroom over the server's ~30s window, otherwise a healthy poll racing its own response gets aborted.
• Retry delay is capped at 2s and uses equal jitter (d/2 + random*d/2). done: false is the expected outcome of a healthy poll, not a failure, so the delay is there to de-synchronize concurrent pollers — every ms of it is time with no poll in flight.
• RequestOpts gains a signal that doRequest forwards to fetch (also flows through the OAuth path).

Two things worth flagging for review, both caught after the first pass:

• doRequest calls bail(error) on any fetch-level rejection, so the naive version of this (just passing a signal) made things worse than no signal at all — a stall returned a 500 on attempt 1 instead of retrying. The request is now wrapped so aborts continue and non-abort errors rethrow.
• The abort check can't use instanceof Error. Node rejects with a DOMException built in undici's realm, which fails instanceof across realm boundaries (vm contexts, and the Jest env) — so the retry silently wouldn't happen. It matches on name structurally instead.

QA Instructions

No visual change — dev-server behaviour only.

Covered by unit tests, including the paths that were broken: stall-then-retry against a real abort, maxRetries: 1 fail-fast, and non-abort errors surfacing instead of being retried away.

Blast Radius

apps dev server only (/__dd/executeAction), which is alpha. Defaults keep the previous 10-attempt behaviour, so the only change for existing users is that retries are now spaced out and a hung poll recovers instead of hanging.

The signal addition to RequestOpts is opt-in and unset everywhere else, so no other product plugin changes behaviour.