[ishandhanani]

Summary

Step 1 of splitting lib/llm into smaller crates. dynamo-llm stops carrying its own copy of the HTTP server machinery and depends on dynamo-http-server for it, so the HTTP metrics and connection types that other dynamo-llm modules import now live below the crate instead of inside its largest import cycle.

• Deleted from dynamo-llm: http/service/server.rs, http/service/write_timeout.rs, the generic half of disconnect.rs (connection handle and monitor) and of metrics.rs (the Metrics family, guards, response collector, error predicates), ServiceObserver / ServiceStage / InflightPermit, the request-id echo middleware, get_body_limit, is_json_content_type, and the frontend-load frame types. http/service/metrics.rs goes from 6,381 to 2,068 lines.
• Kept in dynamo-llm: everything that needs the model catalog or protocol types: model-ready and start-time collectors, LoRA metrics, the serving-load collector adapter, MDC and runtime-config metric updates, response metric processing, the frontend disconnect stream monitors, and RouteDoc (it carries the frontend's OpenAPI route kinds).
• Forward-port into dynamo-http-server: the crate was copied on 2026-10-23 and ten commits changed the frontend copy afterwards (perf(llm): batch the per-token output-token counter per request ai-dynamo/dynamo#17067, feat: add Muse Spark Open support and shared Chat/Responses tool APIs ai-dynamo/dynamo#17068, feat(decisions): coordinate bounded scoring with pinned worker placement ai-dynamo/dynamo#15853, feat(decisions): expose compatible decision and systemone endpoints ai-dynamo/dynamo#15854, refactor(frontend): unify native generate handlers behind format plugins ai-dynamo/dynamo#16983, fix(anthropic): align Messages framing, errors, and validation ai-dynamo/dynamo#17367, feat(vllm): add audio transcription serving ai-dynamo/dynamo#16208, feat(planner): collect batch scheduling observations ai-dynamo/dynamo#14196, feat(frontend): refuse media before routing for text-only models ai-dynamo/dynamo#17874, fix(beam): finalize constrained tool calls at confirmed model EOS ai-dynamo/dynamo#17625). Their changes to the shared code are ported: batched output-token counter with flush on finish and drop, System One histograms, the Transcriptions / SystemOne / Decisions endpoints, the UnsupportedModality error type, sample-count ITL flush.
• New in dynamo-http-server: metrics::scrape_router (the Prometheus scrape route; the frontend wraps it to add its RouteDoc), and server::{bind_listener, write_idle_timeout} plus HttpServer::tls_acceptor, which the frontend admin listener needs to share the main listener's TLS settings.
• No re-exports: dynamo-llm does not re-export the moved items. In-repo callers import them from dynamo-http-server: seven test files under lib/llm/tests, ext-proc, global-frontend, the generate-format catalog test, and the SGLang and vLLM generate-format crates. The Python bindings need no source change.

Not in this PR: the error-class to HTTP status mapping (http::service::error), which is the remaining reason other modules import http; it moves next.

Validation

• cargo test -p dynamo-llm -p dynamo-http-server -p dynamo-generate-format-catalog -p dynamo-sglang-generate -p dynamo-vllm-generate -p dynamo-global-frontend --no-fail-fast: exit 0, 63 test binaries, 4,688 passed, 0 failed, 31 ignored.
• Against the pre-change tree, the 47 dynamo-llm integration and doc-test binaries report the same counts; the only edits to those test files are import paths.
• Unit-test accounting: 57 tests left lib/llm/src and all 57 exist in lib/http-server/src by name; before this change 56 of them ran in both crates. dynamo-llm lib 3,956 to 3,900 (one new upstream test offsets the 57), dynamo-http-server lib 56 to 57.
• Item-level comparison of the pre-change frontend files against lib/http-server: 269 shared functions, types and tests are identical after normalizing visibility and paths; 14 differ only by the intended adaptation (ModelLoadView in place of CommittedModelView, the serving-load collector split, set_worker_info_from_parts).
• cargo check --workspace --all-targets: exit 0. cargo check --locked in lib/bindings/python: exit 0. cargo metadata --locked passes for all four lock files.
• cargo clippy --no-deps --all-targets -- -D warnings on dynamo-llm, dynamo-http-server and the five touched caller crates: exit 0.
• In-crate imports of http from other dynamo-llm modules: 26 files in 12 modules before, 14 files in 9 modules after; discovery, migration and frontend_load no longer import it.