fix(model-error-classifier): mark OpenAI server_error patterns as retryable (fixes #3799)

OpenAI streaming responses can surface mid-stream errors with type
'server_error' or with the prose message 'An error occurred while processing
your request'. Neither matched any entry in RETRYABLE_MESSAGE_PATTERNS, so
shouldRetryError returned false and the runtime-fallback / fallback-retry
code paths skipped retry. The result was that GPT-5.5 subagent (and main)
turns silently stalled until the stale timeout fired.

The maintainer's diagnosis on issue #3799 explicitly recommends adding
these two patterns to model-error-classifier.ts; this commit does exactly
that and adds two regression tests covering the JSON envelope and the
prose form.

src/shared/model-error-classifier.ts

@@ -80,6 +80,11 @@ const RETRYABLE_MESSAGE_PATTERNS = [
   "请求过于频繁",       // "too many requests"
   "暂时不可用",         // "temporarily unavailable"
   "服务不可用",         // "service unavailable"
+  // OpenAI streaming server_error events surface either as a literal "server_error"
+  // type or as the prose error sentence below. Without these patterns subagent
+  // streams stall instead of being retried (issue #3799).
+  "server_error",
+  "an error occurred while processing",
 ]
 
 /**