fix: keep long extended-thinking streams intact

[kokojacket]

Summary

• size the claude transformer's per-event scanner to the event itself instead of bufio.Scanner's default 64 KiB limit, and pass the original event through if scanning ever fails
• replace the upstream client's 300s total Timeout with a 10-minute idle timeout (no headers / no body data), so healthy long streams are never cut off while stalled upstreams are still aborted
• lift the server WriteTimeout (10 min) for the proxy route only, so long streams are not truncated on the client side
• keep /v1/models fetches bounded by their own 300s timeout, since the shared client no longer has a total timeout

Why

With long extended thinking (e.g. Opus at high effort), Claude Code sessions behind ccNexus broke permanently:

1. Lost thinking signatures. The signature_delta line grows with thinking length and goes over 64 KiB after roughly 4 minutes of thinking. ClaudeTransformer.TransformResponseWithContext used a default bufio.Scanner and ignored scanner.Err(), so that line was silently dropped (only event: content_block_delta was forwarded). The client stored a thinking block without a signature, and every following request failed with 400 messages.N.content.0.thinking: each thinking block must contain thinking. The session could not continue until the transcript was edited by hand. fix: support multi-megabyte SSE events #141 raised the limit in internal/proxy, but this scanner in internal/transformer/cc still used the default.
2. Streams cut at exactly 5 minutes. http.Client.Timeout covers reading the response body, so any stream longer than 300s ended with Connection lost mid-response. The server WriteTimeout would cut them off again at 10 minutes.

Tests

• new TestClaudeTransformerForwardsLargeSignatureDelta (fails before the fix: 27 of 102525 bytes forwarded)
• new TestProxiedStreamOutlivesServerWriteTimeout: end-to-end through handleProxy, fails before the fix with unexpected EOF
• new idle-timeout tests: an active stream outlives the idle timeout, and stalled streams and missing headers are aborted
• go test ./... -count=1 -race

🤖 Generated with Claude Code