[jhssilva]

What does this PR do?

Fixes two client-side issues in @dd/core/helpers/request surfaced during high-volume sourcemap uploads in web-ui (companion concurrency PR: https://github.com/ddoghq/web-ui/pull/28126):

1. Retry transient fetch() network errors (ECONNRESET, EPIPE, ETIMEDOUT, UND_ERR_*):

   • Previously, doRequest() wrapped both getData() and await fetchImpl(...) in a single try/catch that called bail(error).
   • When fetch() rejected with TypeError: fetch failed (wrapping ECONNRESET, EPIPE, ETIMEDOUT, or UND_ERR_SOCKET), doRequest() called bail(error) on attempt 1, skipping all NB_RETRIES (5) retries and failing the upload immediately with status_code:unknown.
   • doRequest() now re-throws transient fetch network errors so async-retry retries them with backoff, while still bailing immediately on getData() errors, AbortSignal timeouts (AbortError / TimeoutError), and non-network TypeErrors (such as locked/disturbed streams).

2. Materialize compressed payloads in createRequestData() before opening the HTTP request:

   • Previously, createRequestData() returned a lazy req.body.pipeThrough(new CompressionStream('gzip')) stream. When passed to fetch(), Undici immediately sent the HTTP request headers and the 10-byte gzip header (network.bytes_read: 10)—starting the intake edge's 30s/60s request-body receive timers—before the file-backed blobs were read from disk or compressed on Node's libuv threadpool. Under high concurrency, streams queued behind disk/zlib work sent 0 body bytes after the 10-byte gzip header for 60s and timed out with HTTP 408 (response_flags: SI, stream_idle_timeout).
   • createRequestData() now buffers the compressed payload in memory and sets Content-Length before returning, ensuring disk reads and gzip compression complete before the HTTP request headers are sent on the wire.

Related Links

• Companion web-ui PR by @matissjanis-aboltins_ddog: https://github.com/ddoghq/web-ui/pull/28126
• Investigation Notebook: https://app.datadoghq.com/notebook/15973703