fix: send max_tokens for Claude/OpenRouter + retry SSE connection errors

[teknium1]

Summary

Fixes persistent "Network connection lost" failures when Claude Opus generates large tool call responses (e.g. write_file with a plan file) through OpenRouter.

Root Cause Investigation

OpenRouter has an undocumented ~125s inactivity timeout on their upstream proxy to Anthropic. When Opus thinks for >125s before generating tool call output, the proxy kills the connection.

Proved with controlled experiments:

Path	Tools	Result	Time
OpenRouter, no max_tokens	✅	❌ Network connection lost	128s
OpenRouter, max_tokens=128K	✅	❌ Network connection lost	129s
OpenRouter, no tools (text)	❌	✅ 12K tokens	155s
OpenRouter, small tool call	✅	✅ 2.8K chars	18s
OpenRouter, medium tool call	✅	✅ 19K chars	75s
Direct Anthropic API	✅	✅ 47K chars	173s

Chunk timing trace confirmed: all 26 chunks arrive in a 3-second burst at stream start, then complete silence for 125s while Opus thinks, then the connection dies. The direct Anthropic API handles the same silence and returns successfully at 173s.

Additionally, echo_upstream_body debug revealed OpenRouter defaults to max_tokens: 65536 when we don't send it — only half of Opus 4.6's 128K output capacity.

Changes

1. Send explicit max_tokens for Claude through OpenRouter

When self.max_tokens is not set by the user, we now send the model's actual output limit (from _get_anthropic_max_output()) for Claude models on OpenRouter. This ensures full output capacity instead of OpenRouter's 65K default. Extends the practical success window for medium-length responses.

2. Classify SSE connection errors as retryable

OpenRouter sends {"error":{"message":"Network connection lost."}} as SSE events when the upstream drops. The OpenAI SDK raises APIError from these — but our streaming retry logic only recognized httpx-level errors (ReadTimeout, RemoteProtocolError). Now SSE errors with connection-related messages (no HTTP status code) are retried with fresh connections, same as httpx errors.

3. Actionable error guidance

When stream-drop retries are exhausted, the error message now explains the issue and suggests alternatives (execute_code with Python open(), write in smaller sections).

Tests

• 2 new tests: test_sse_connection_lost_retried_as_transient, test_sse_non_connection_error_falls_back_immediately
• All 220 streaming + run_agent tests pass
• Full suite: 6532 passed

Note for OpenRouter

This is ultimately an OpenRouter-side limitation — their upstream proxy timeout doesn't account for Opus's long thinking phase on complex tool calls. Should be reported to them.