[taciturnaxolotl]

Retry decisions are made by matching on what an error looks like: a ProviderError, a net.Error, or a transport error. An error that knows it is worth retrying, but is none of those, has to dress up as one. Crush's request timeout was posing as a net.Error just to get retried, and even then a retryable timeout that wrapped a 400 ProviderError lost to the provider error's own verdict.

• Adds a RetryableError interface with IsRetryable() bool. The outermost one in the chain decides. ProviderError already has that method, so it satisfies the interface and its separate branch is gone.
• An error that wraps a context error is still never retried. The abort check runs first, and the interface doc says so.

Used by charmbracelet/crush#3985, which has to point at a release containing this before it can merge.

💘 Generated with Crush