fix: Slow side-question SSE clients can strand provider runs indefinitely

[sam-saffron-jarvis]

What changed

• Wrapped side-question SSE responses with the existing per-write streaming response writer and its 30-second deadline, matching chat and response-run streams.
• Added a regression test with a writer that blocks unless a write deadline is installed. The test verifies the handler returns, records a bounded nonzero deadline, and cancels the detached side-question provider.

Why this is high-value

The HTTP server intentionally has no global write timeout so long-lived SSE streams remain valid. Previously, side-question SSE was the exception to the per-write deadline rule: a suspended mobile browser or another client that stopped reading could block the handler in Write or Flush forever. Because the provider uses a detached context, that also left the side question active, made the session non-evictable, and caused later side questions to conflict. Bounding each write lets the existing write-error path cancel the provider and release the session.

Validation

• gofmt -w cmd/serve_side_question.go cmd/serve_side_question_test.go
• go build ./...
• XDG_CONFIG_HOME=$(mktemp -d) go test ./...
• go test ./cmd -run 'TestServeSideQuestion(Disconnect|SlowClient)' -count=100
• git diff --check