fix: retry transient Anthropic server errors (HTTP 200 with error body)

[Sam Schillace (ramparte)]

Problem

Transient Anthropic API failures kill sessions instead of being retried.

Anthropic returned HTTP 200 with an error body:

```json
{
"type": "error",
"error": {
"type": "api_error",
"message": "Internal server error",
"details": null
}
}
```

The SDK didn't raise an exception (HTTP status was 200), the response wasn't validated, and the session halted immediately with `retryable=false`. The failing request was small (~4.5k tokens) with no tool calls — three earlier, larger requests with tool calls all succeeded fine. Textbook transient server-side failure.

Root Cause

Two gaps in the error handling:

1. No response validation — after the API call returns, the response is passed directly to `_convert_to_chat_response()` without checking if it's actually a valid message. HTTP 200 with an error body slips through.

2. Base `APIError` not caught — the exception chain catches specific `APIStatusError` subclasses (`RateLimitError`, `BadRequestError`, etc.) and `APIStatusError` itself, but not the base `APIError`. The SDK can raise `APIConnectionError`, `APITimeoutError`, or other non-status errors that fall through to the generic `Exception` handler.

Fix

Two defensive layers, both feeding into the existing `retry_with_backoff` loop:

1. Response validation (belt-and-suspenders)

After the API call, before returning the response, validate `response.type == "message"`. If not, extract error details and raise `KernelProviderUnavailableError(retryable=True)`.

2. Base `AnthropicAPIError` catch

New exception handler between `AnthropicAPIStatusError` and `asyncio.TimeoutError` that catches the SDK's base `APIError` class. Treats all as transient/retryable.

Testing

• All 298 existing tests pass
• Syntax verified via `ast.parse()`
• Pure additions — no existing code modified or removed
• Patch tested locally via source override