server(sse): stop sending Connection: keep-alive on the streaming endpoints

[jamesburton]

Connection is a hop-by-hop, connection-specific header field. RFC 9113 §8.2.2 forbids it over
HTTP/2 and HTTP/3 (an endpoint MUST treat a message containing connection-specific header fields as
malformed), and over HTTP/1.1 it is merely redundant since persistent connections are the default.
Both OpenAI-compatible streaming endpoints were setting it.

Kestrel filters it on HTTP/2, so nothing observable breaks in the common configuration — which is
why it went unnoticed. It matters when a strict intermediary is in the path, and it reads as
intentional to anyone copying the pattern into a new endpoint, which is how it reached the Anthropic
endpoint in #326.

Changes

• ChatCompletionEndpoint.HandleStreamingAsync and CompletionEndpoint.HandleStreamingAsync no
  longer set Connection.
• Both now configure the response through a new shared internal SseResponse.ApplyHeaders, which
  sets the two headers SSE actually needs (text/event-stream, Cache-Control: no-cache) and
  carries the note on why Connection is deliberately absent. One place for it to not be re-added,
  rather than a comment repeated per endpoint.
• tests/DotLLM.Tests.Unit/Server/SseResponseHeaderTests.cs — regression guard asserting the header
  is absent and the other two are set, in the pattern used by server: Anthropic-compatible Messages API (/v1/messages) (#325) #326's AnthropicStreamingTests.
  Written self-contained since that branch is not merged.

Verification

dotnet test tests/DotLLM.Tests.Unit --filter "FullyQualifiedName~Tests.Unit.Server" -v q --nologo
Passed! - Failed: 0, Passed: 20, Skipped: 0, Total: 20

Closes #422

[Copilot]

Pull request overview

Removes the illegal/redundant Connection: keep-alive header from the OpenAI-compatible SSE streaming endpoints to avoid malformed responses over HTTP/2+ (and reduce the chance the pattern gets copied into new endpoints). Consolidates SSE header setup into a single shared helper and adds a regression test.

Changes:

• Removed Connection header emission from /v1/chat/completions and /v1/completions streaming responses.
• Introduced SseResponse.ApplyHeaders(HttpContext) as the single shared place to apply SSE-required headers (Content-Type, Cache-Control) and document why Connection must not be set.
• Added a unit test guarding the expected SSE headers and asserting Connection is absent.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
src/DotLLM.Server/Endpoints/ChatCompletionEndpoint.cs	Routes streaming response header setup through SseResponse.ApplyHeaders and stops setting Connection.
src/DotLLM.Server/Endpoints/CompletionEndpoint.cs	Routes streaming response header setup through SseResponse.ApplyHeaders and stops setting Connection.
src/DotLLM.Server/Endpoints/SseResponse.cs	New shared internal helper that applies SSE headers and documents the Connection header prohibition.
tests/DotLLM.Tests.Unit/Server/SseResponseHeaderTests.cs	Regression tests asserting required SSE headers are set and Connection is not emitted.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.