[Bug]: Scaffolding KV cache hints use an unregistered HTTP endpoint

System info

Source inspected at main commit 7b1bedbbb. This is a client/server HTTP route mismatch; no GPU inference run was performed.

Problem

With KV cache hints enabled, OpenaiWorker posts message-based DropKVCacheTask requests to base_url + "kv_cache_hints". For the usual OpenAI base URL http://localhost:8000/v1/, this targets /v1/kv_cache_hints.

The TRT-LLM resource governor instead registers POST /_control/kv_cache/truncate (and the compatibility alias /_resource_governor/truncate). It does not register /v1/kv_cache_hints, so the request cannot reach the truncate handler and receives 404 on a directly connected server.

Reproduction

Configure OpenaiWorker with kv_cache_hint_enabled=True and an OpenAI client base URL ending in /v1/, then execute a DropKVCacheTask. Inspect the outgoing POST path and compare it with the routes in tensorrt_llm/serve/resource_governor.py.

Expected behavior / proposed fix

Post to /_control/kv_cache/truncate, stripping the optional trailing /v1 and normalizing the trailing slash. Preserve the existing request body and Authorization header.

The token-based truncate method already uses the sibling control-plane endpoint. The message-based path changed to kv_cache_hints in #14397; before that it used the registered resource-governor endpoint.

Start with the message-based DropKVCacheTask path in OpenaiWorker and compare its outgoing URL with the routes in tensorrt_llm/serve/resource_governor.py. Reproduce with kv_cache_hint_enabled=True and an OpenAI base URL ending in /v1/, then verify that the request reaches /_control/kv_cache/truncate with the existing body and Authorization header.