Classify SessionServer endpoints before proxying

[matrix72c]

Summary

This PR makes SessionServer classify requests from the HTTP method and normalized endpoint before inspecting or mutating their JSON envelopes.

It fixes Anthropic /v1/messages/count_tokens requests being mistaken for model generations because they contain prompt-shaped fields such as messages, and it closes the catch-all proxy path that could otherwise forward unrelated requests to LMDeploy management endpoints.

This endpoint-policy change was split out of #2059 so that proxy routing and trace-reliability behavior can be reviewed independently.

Endpoint behavior

Request	Behavior	Training trace
POST /v1/messages	Forward as Anthropic generation	Yes, when requested
POST /v1/chat/completions	Forward as OpenAI generation	Yes, when requested
POST /v1/responses	Return the existing native 501 response	No
POST /v1/messages/count_tokens	Transparently forward, removing only a legacy SessionServer-owned session_id	No
HEAD /api/hello	Return 204 locally	No
Any other method or endpoint	Return native-format 404 locally	No

In particular, GET /v1/models and LMDeploy management routes such as /terminate, /update_weights, /sleep, and /wakeup never reach the worker.

Implementation

• Add a small private classifier based on method + normalized path.
  • Leading slashes are normalized.
  • Query strings do not participate in classification.
  • Known endpoints may appear behind a base-path prefix.
• Separate generation processing from auxiliary forwarding before request hooks run.
  • Only generation endpoints run on_request, response cleaning, stream parsing, and on_response.
  • count_tokens preserves messages, system, tools, and other client fields. Only the compatibility-only session_id field is removed.
• Handle HEAD /api/hello, unsupported methods, and unsupported endpoints locally before opening an upstream connection.
• Log rejected requests using only the method and normalized path; bodies, headers, query strings, and credentials are not logged.

The classifier is intentionally private and local to SessionServer; this PR does not add a shared enum, dataclass, or cross-project abstraction.

Tests

Added focused classifier and local aiohttp fake-upstream tests covering:

• leading slash, base-path prefix, query string, and incorrect method handling;
• Messages, Chat Completions, Responses, token counting, and hello endpoints;
• transparent token-count request/response forwarding with no trace hooks;
• unchanged generation hook execution; and
• local rejection of unknown and LMDeploy management endpoints with zero upstream calls.

Validation performed on this topic alone:

pytest -q tests/rl/test_session_server_endpoints.py
6 passed

ruff check xtuner/v1/rl/rollout/session_server.py tests/rl/test_session_server_endpoints.py
All checks passed

git diff --check upstream/main...fix/session-server-endpoint-classification

The topic also merges cleanly with #2059 in the local integration branch; the combined endpoint and trace suites pass all 12 tests.

Non-goals

• Trace recovery after malformed successful responses or early client disconnects remains in Preserve SessionServer traces across response failures #2059.
• No LMDeploy code or management route is changed.
• Model discovery remains disabled.
• Claude Code WebFetch, plugin-marketplace, and general tool network policy are outside this proxy path and are unchanged.