feat(relay): add isolated Switchyard routing plugin

Signed-off-by: Bryan Bednarski <bbednarski@nvidia.com>

Author: bbednarski9

CHANGELOG.md

@@ -6,6 +6,20 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Added
+
+- **NeMo Relay native plugin** — a dynamically loaded integration that runs
+  libsy's weighted-random, LLM-classifier, escalation, and stage-router
+  algorithms in process while Switchyard owns provider HTTP dispatch,
+  credentials, translation, retries, and fallback. Managed calls require NeMo
+  Relay 0.7 or newer and do not depend on `switchyard-server`.
+
+- **NeMo Relay routing-model usage marks** — classifier judges, escalation
+  judges and discarded weak candidates, and failed routing candidates now emit
+  `switchyard.routing.llm_call` ATOF marks with normalized token usage and
+  latency. The final serving call remains represented only by Relay's outer LLM
+  lifecycle event to prevent double-counting.
+
 ### Removed
 
 - **Latency-aware router** — the `latency_service` route type and its
@@ -31,6 +45,14 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Fixed
 
+- **NeMo Relay stage and escalation integration** — preserved same-protocol
+  request bodies are now re-encoded after tier prompts or handoff notes mutate
+  the normalized request; Relay's synthetic `gateway-gateway` identity no
+  longer shares escalation latch state across unrelated raw gateway requests;
+  stage decision marks now retain picker-default tiers, decision sources, and
+  hard-override confidence. Target bindings also accept non-secret
+  `extra_body` defaults for provider-specific judge controls.
+
 - **Response `model` now names the model that actually served the request**, on
   every serving path and wire format. Streamed Anthropic and Responses replies,
   and every libsy-served reply, previously echoed the model id the client

README.md

@@ -21,6 +21,7 @@ algorithm you write yourself.
 - **Protocol Translation**: convert between OpenAI Chat, Anthropic Messages, and OpenAI Responses formats
 - **Multi-Backend Routing**: random routing, LLM-as-classifier routing, signal-driven stage-router, or your own algorithm
 - **Operational Metrics**: Prometheus metrics cover requests, errors, latency, tokens, and routing overhead
+- **NeMo Relay Plugin**: run random, classifier, escalation, or stage routing in Relay while Switchyard owns provider HTTP dispatch
 
 ## Maturity
 
@@ -154,6 +155,7 @@ configured LLM client selects one upstream format.
 - **[`switchyard-libsy`](crates/libsy/README.md)**: embed routing algorithms in a Rust application
 - **[`switchyard-protocol`](crates/protocol/README.md)**: provider-neutral request, response, and streaming types
 - **[`switchyard-translation`](crates/switchyard-translation/README.md)**: request, response, and stream translation
+- **[`switchyard-nemo-relay-plugin`](crates/switchyard-nemo-relay-plugin/README.md)**: install Switchyard as a native NeMo Relay plugin
 
 ## Community