feat(SERVE-COMPLETION-LONGTAIL): session_id is not carried from the request to the engine, and no handler reads headers

Row: SERVE-COMPLETION-LONGTAIL

Found by wave PORTQ-2 re-deriving PORT-NOW entry 60 of 5559679229..e126687a9a
(#2646). Upstream:
f57123aa2d vllm#48048, "feat(frontend): session id plumbing into requests".
Previously read by #2524's sample (.agents/sync/2026-09-01-cdefd9d.md §13 item
9, "API-only"); re-verified at e24ec8bfd, still holds.

Absent, checked twice and in this tree's own vocabulary

session_id, SessionId, X-Session-ID and session-id return zero lines
across include/, src/ and tests/. Searching the CONCEPT rather than the
upstream spelling gives the same answer: the request carries only request_id
and priority (include/vllm/v1/engine/types.h:77,88, :176), and the server
reads no request headers at all —
src/vllm/entrypoints/openai/api_server.cpp:1176-1179 binds
[this, write](const httplib::Request& req, httplib::Response& res) { write(handle_chat_completions(req.body), res); },
discarding the header-bearing object. The only get_header_value in the tree is
in the HuggingFace HTTP client (src/vllm/transformers_utils/downloader.cpp:65).

Unlike most of this tranche, every surface the commit touches is ported here
(the OpenAI request structs, serving_chat / serving_completion,
EngineCoreRequest, v1::Request), which is why this is portable work rather
than an absent surface.

What the port needs

1. An optional session_id on ChatCompletionRequest / CompletionRequest in
   include/vllm/entrypoints/openai/protocol.h and its parse.
2. The resolution order: body field, then the X-Session-ID header, then
   vllm_xargs["session_id"]. Our vllm_xargs is itself deferred
   (protocol.h:37), so only the first two arms are reachable and the third is
   owed rather than ported.
3. A header-plumbing seam. api_server.cpp:1176-1179 drops
   httplib::Request today, so the handler signature has to carry headers. This
   is the largest single piece and it does NOT exist upstream, where
   raw_request is already threaded.
4. std::optional<std::string> session_id on EngineCoreRequest
   (include/vllm/v1/engine/types.h) and v1::Request
   (include/vllm/v1/request.h), copied alongside src/vllm/v1/request.cpp:109.
5. A reader, or the field lands dead under AGENTS.md §"Nothing lands dead".
   LogRequestReceived (include/vllm/entrypoints/openai/request_logger.h:32-35)
   is the natural one and is what makes the field reachable from the server's
   default configuration.

Roughly 150-220 lines over ~10 files plus one test; the header seam is the part
to size carefully.

Open question

Whether upstream has any CONSUMER of session_id beyond carrying it. At this
commit it is pure plumbing. If it is carry-only upstream too, our port must add a
reader anyway to be reachable — settled by grepping session_id in vLLM at a
later revision for a read site.

Nothing was executed for this finding.

Start with the request types in include/vllm/entrypoints/openai/protocol.h and the handler in src/vllm/entrypoints/openai/api_server.cpp, then trace propagation through include/vllm/v1/engine/types.h, include/vllm/v1/request.h, and src/vllm/v1/request.cpp. Add the documented body/header plumbing and a reader in include/vllm/entrypoints/openai/request_logger.h; done includes the one planned test passing and no dead session_id field.