feat(prefill-router): add learned prefill complexity routing

.agents/skills/switchyard-coding-agent-launchers/SKILL.md

@@ -43,6 +43,7 @@ after.
 | Look up upstream model catalogs | `fetch_model_ids(...)` in `switchyard/cli/model_catalog/model_discovery.py` |
 | Add a plan-execute route | Add `type: plan_execute` in route YAML; `_plan_execute_switchyard` in `route_bundle.py` maps it to `PlanExecuteProfileConfig`. |
 | Add a deterministic route | Add `type: deterministic` in route YAML; `_deterministic_switchyard` maps it to `DeterministicRoutingProfileConfig`. |
+| Add a learned prefill-probe route | Add `type: prefill_probe` in route YAML; `_prefill_probe_switchyard` maps it to `PrefillProbeProfileConfig` while model discovery skips the internal probe. |
 | Add a preset | Put it beside the profile config under `switchyard/lib/profiles/`; presets return typed config objects, not runnable apps. |
 | Inspect saved defaults | `switchyard configure --show` reads `switchyard/cli/config/user_config.py` and renders via `switchyard/cli/status.py`. |
 
@@ -87,6 +88,8 @@ Route YAML and launchers share this model-dispatch path:
   direct strong/weak passthrough entries.
 - `type: deterministic`, `type: escalation_router`, and `type: stage_router`
   register the route key plus direct strong/weak passthrough entries.
+- `type: prefill_probe` registers the route key plus direct strong/weak
+  passthrough entries; its probe target remains internal.
 - `type: plan_execute` registers the route key plus the executor as a direct
   passthrough; the planner is internal routing logic.
 - `type: latency_service` and `type: noop` register the

.agents/skills/switchyard-lib-core/SKILL.md

@@ -39,6 +39,7 @@ right validation set. If the change is driven by a launcher need, also read
 | An OpenAI-compatible provider target such as NVIDIA Inference Hub or OpenRouter | Use the existing OpenAI-compatible backend/profile with `base_url`, `api_key`, and model id wiring. Add a new backend only when the provider has a real wire-format, auth, retry, or health contract that cannot fit that path. |
 | Direct Rust component bindings | Add concrete PyO3 classes under `crates/switchyard-py/src/component_bindings/`, keep config bindings near the component binding that consumes them, and expose them lazily from `switchyard_rust/components.py`. Do not keep growing `core_bindings.rs` or `switchyard_rust/core.py` with concrete component classes. |
 | Route YAML / model dispatch | Use `switchyard/cli/route_bundle.py` and `switchyard/lib/route_table_builders.py`. They build `RouteTable` entries from profile-backed runtimes and keep launchers plus `switchyard serve --routing-profiles` on one path. |
+| A learned prefill-probe route | Use `type: prefill_probe` with `probe`, `strong`, and `weak` targets. The route table exposes and discovers only the completion tiers; the probe remains internal to `PrefillProbeProfileConfig`. |
 | Route sub-agent traffic to a fixed worker target | Set `subagent_target: <target-id>` in any profile's common envelope (consumed like `type` in Rust `SerializedProfileConfig`, `crates/switchyard-components-v2/src/config/parsing.rs`). The Python loader (`switchyard/lib/profiles/loader.py`) wraps the built profile in `SubagentOverrideProfile`; detection is the Rust-bound `is_subagent_request(headers)` from `switchyard_rust.profiles`, a thin wrapper over `Metadata::from_headers` + `Metadata::is_subagent_work` in `crates/protocol/src/metadata.rs` (the canonical lineage fact and work-vs-maintenance policy). On the libsy `Algorithm` path, wrap with the `SubagentOverride` combinator (`crates/libsy/src/algorithms/subagent_override.rs`; Python: `switchyard.libsy.algorithms.subagent_override`). Do not re-implement header sniffing inside individual profiles or algorithms. |
 | Shared/persistent session-affinity pins across workers or pod churn | Configure the latency route with `session_affinity: true` + `affinity_store: redis` + `affinity_store_url` (optional `affinity_store_ttl_seconds`, `affinity_key_prefix`); the escalation_router route takes the same `affinity_store*` keys (no `session_affinity` flag — its latch is always on; default prefix `swyd:esc:`). `SessionAffinity` keeps the Rust `SessionCache` as L1 and reads/writes through the `AffinityPinStore` L2 (`switchyard/lib/redis_pin_store.py`), fail-open behind a 0.1s socket timeout and a 3-failure/10s-cooldown circuit breaker (`switchyard_affinity_l2_breaker_open` gauge). Requires the `switchyard[affinity-redis]` extra. |
 | Stats / telemetry | Reuse `StatsRequestProcessor`, `StatsResponseProcessor`, `StatsLlmBackend`, and `StatsAccumulator`. A profile config should thread one accumulator through all three when stats are enabled. Do not write a parallel collector. |

benchmark/routing-profiles/prefill-probe-local.yaml

@@ -0,0 +1,55 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Learned prefill-complexity route with a local Qwen3.6 hidden-state probe.
+# Start the probe as documented in docs/vllm-serve-hidden-state.md, then export:
+#   NVIDIA_API_KEY
+#   VLLM_BASE_URL=http://127.0.0.1:8000
+#   HIDDEN_STATES_DIR=/tmp/vllm-hidden-states
+#   PREFILL_ROUTER_CHECKPOINT_DIR=/absolute/path/to/inference_artifact
+#
+# Serve from the repository root:
+#   switchyard --routing-profiles \
+#     benchmark/routing-profiles/prefill-probe-local.yaml -- serve --port 4000
+# Clients select the learned route by sending model: prefill-complexity-router.
+
+defaults:
+  timeout_secs: 600.0
+
+routes:
+  prefill-complexity-router:
+    type: prefill_probe
+    enable_stats: true
+    fallback_target_on_evict: strong
+    tier_timeout_s: 600.0
+
+    probe:
+      id: probe
+      model: Qwen/Qwen3.6-35B-A3B
+      base_url: ${VLLM_BASE_URL}/v1
+      format: openai
+
+    strong:
+      id: strong
+      model: aws/anthropic/bedrock-claude-opus-4-7
+      api_key: ${NVIDIA_API_KEY}
+      base_url: https://inference-api.nvidia.com/v1
+      format: anthropic
+
+    weak:
+      id: weak
+      model: nvidia/nvidia/nemotron-3-super-120b-long-ctx
+      api_key: ${NVIDIA_API_KEY}
+      base_url: https://inference-api.nvidia.com/v1
+      format: openai
+
+    strong_checkpoint_head: opus-4.7
+    weak_checkpoint_head: nemotron-3-super
+    hidden_states_dir: ${HIDDEN_STATES_DIR}
+    checkpoint_dir: ${PREFILL_ROUTER_CHECKPOINT_DIR}
+
+    routing_policy:
+      type: cost_aware
+      lambda: 0.5
+      weak_cost: 0.0
+      strong_cost: 1.0

crates/switchyard-components/Cargo.toml

@@ -15,16 +15,18 @@ rust-version.workspace = true
 async-stream = "0.3"
 async-trait = "0.1"
 futures-util = "0.3"
+half = "2"
 parking_lot = "0.12"
 rand = "0.8"
 reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls-native-roots", "stream"] }
+safetensors = "0.4"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 switchyard-core = { path = "../switchyard-core" }
 switchyard-libsy = { path = "../libsy" }
 switchyard-protocol = { path = "../protocol" }
 switchyard-translation = { path = "../switchyard-translation" }
-tokio = { version = "1", features = ["rt", "sync"] }
+tokio = { version = "1", features = ["rt", "sync", "time"] }
 tracing = { version = "0.1", default-features = false, features = ["std"] }
 
 [dev-dependencies]

crates/switchyard-components/src/lib.rs

@@ -10,6 +10,7 @@
 pub mod backends;
 pub mod dimension_collector;
 pub mod intake;
+mod prefill_probe;
 pub mod request_processors;
 pub mod response_processors;
 pub mod stage_router;
@@ -29,7 +30,8 @@ pub use intake::{
     RequestMetadata, SubModelCall, SubModelCalls,
 };
 pub use request_processors::{
-    DimensionCollector, IntakeRequestProcessor, RandomRoutingDecision, RandomRoutingEngine,
+    DimensionCollector, IntakeRequestProcessor, PrefillProbeProcessorConfig,
+    PrefillProbeRequestProcessor, RandomRoutingDecision, RandomRoutingEngine,
     RandomRoutingProcessorConfig, RandomRoutingTier, StatsRequestProcessor,
 };
 pub use response_processors::{