[FR] Expose retry-attempt observability (hook/event) — retried 503s and per-attempt latency are invisible to callers

[milcho0604]

Is your feature request related to a problem?

The HTTP clients retry transient failures — per the code comment in api-request, "Retries up to 4 times on connection reset and timeout errors as well as 503 errors" — which is great, but the retry lifecycle is completely invisible to the caller. You observe only the final result and total elapsed time.

In production this matters a lot for FCM:

• A sendEach() call that "took 15s" is indistinguishable from one that made three 5s attempts. Capacity planning, SLO attribution, and incident analysis need to tell these apart.
• 503s that the SDK retries away never appear anywhere. During an FCM backend hiccup, our error rate looked flat while the wire was full of retried 503s — we only learned this after instrumenting below the SDK.

What we did as a workaround (and what it revealed)

We attached listeners at the HTTP/2 layer to record, per attempt, the :status header the SDK already receives, plus per-attempt request→response duration. Two things became visible immediately:

1. Swallowed 503s during backend episodes (invisible at the SDK surface).
2. Batches with ~15s total latency decomposed cleanly into 3 × ~5s attempts — i.e. the tail was retry behavior, not slow single requests.

We validated the accounting at scale: in a ~650k-message load run at ~2,000 rps, wire-level attempt counts reconciled exactly with SDK-level result counts (596,663 = 596,663).

Monkey-patching works but is version-fragile and clearly not the intended way.

Describe the solution you'd like

Any of these would solve it (in rough order of preference):

1. An onRetryAttempt(info) callback / EventEmitter on the messaging or app options, with { attempt, statusCode?, errorCode?, elapsedMs, willRetry }.
2. Attempt metadata attached to the final response/error (e.g. attempts: [{status, elapsedMs}, ...]).
3. At minimum, a debug logging hook for retry decisions.

Describe alternatives you've considered

• Runtime-patching the HTTP/2 request path to observe response headers (works, but couples us to SDK internals).
• enableLegacyHttpTransport() + external proxy metrics (gives up HTTP/2).

Additional context

Related: #1615 (custom RetryConfig) — configuration and observability of the same mechanism. Verified against 12.7.0 and 14.1, on Node 16 and 24.