feat: automatically respect rate limit headers in retry logic (#7246)

## Description

This PR adds automatic support for respecting rate limit headers
(`retry-after-ms` and `retry-after`) in the retry logic, following the
same approach used by both the [Anthropic client
SDK](https://github.com/anthropics/anthropic-sdk-typescript/blob/main/src/client.ts#L747-L775)
and [OpenAI client
SDK](https://github.com/openai/openai-node/blob/master/src/client.ts#L753-L781).

Currently, when providers like Anthropic and OpenAI return 429 rate
limit errors with `retry-after` headers, the SDK doesn't automatically
respect these headers and instead uses its standard exponential backoff.
This can lead to unnecessary retries that fail because they happen
before the rate limit window has passed.

## Changes

- Added `getRetryDelay` function that parses rate limit headers and
determines the appropriate retry delay
- Modified retry logic to use the parsed rate limit delay when
reasonable (0-60 seconds), falling back to exponential backoff otherwise
- This is a behind-the-scenes improvement with no API changes required

The implementation follows the exact logic used by both Anthropic and
OpenAI SDKs:
1. First checks for `retry-after-ms` header (milliseconds)
2. Falls back to `retry-after` header (seconds or HTTP date)
3. Only uses the header value if it's reasonable (between 0 and 60
seconds)
4. Falls back to exponential backoff if no valid headers or if the delay
is unreasonable

## Testing

Added comprehensive tests covering:
- Reasonable rate limit delays (under 60 seconds)
- Unreasonable delays that fall back to exponential backoff
- Invalid header values
- Multiple retry scenarios
- Mocked responses from Anthropic and OpenAI
- HTTP date format parsing
- Preference for `retry-after-ms` over `retry-after`

All tests pass successfully.

## Issue

Closes #7247

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: jeffbarg

.changeset/wise-headers-respect.md

@@ -0,0 +1,7 @@
+---
+'ai': patch
+---
+
+feat: automatically respect rate limit headers in retry logic
+
+Added automatic support for respecting rate limit headers (`retry-after-ms` and `retry-after`) in the SDK's retry logic. When these headers are present and contain reasonable values (0-60 seconds), the retry mechanism will use the server-specified delay instead of exponential backoff. This matches the behavior of Anthropic and OpenAI client SDKs and improves rate limit handling without requiring any API changes.

packages/ai/src/util/prepare-retries.ts

@@ -1,7 +1,7 @@
 import { InvalidArgumentError } from '../../src/error/invalid-argument-error';
 import {
   RetryFunction,
-  retryWithExponentialBackoff,
+  retryWithExponentialBackoffRespectingRetryHeaders,
 } from '../../src/util/retry-with-exponential-backoff';
 
 /**
@@ -37,6 +37,8 @@ export function prepareRetries({
 
   return {
     maxRetries: maxRetriesResult,
-    retry: retryWithExponentialBackoff({ maxRetries: maxRetriesResult }),
+    retry: retryWithExponentialBackoffRespectingRetryHeaders({
+      maxRetries: maxRetriesResult,
+    }),
   };
 }