[Router] Add composable scoring and eligibility policies (#37731)

Co-authored-by: inkcherry <mingzhi.liu@amd.com>

experimental/sgl-router/src/config/cli.rs

@@ -12,9 +12,9 @@ use std::num::NonZeroU32;
 use crate::config::{
     default_cb_cool_down, default_proxy_request_timeout_secs, default_stale_request_timeout_secs,
     resolve_mode, ActiveLoadConfig, CacheAwareConfig, CircuitBreakerConfig, Config,
-    DiscoveryBackend, K8sDiscoveryConfig, KvIndexerEndpointConfig, LogFormat, ModelConfig,
-    ObservabilityConfig, PolicyKind, ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig,
-    StickyConfig,
+    DiscoveryBackend, EligibilityConfig, FilterKind, FusedTerm, K8sDiscoveryConfig,
+    KvIndexerEndpointConfig, LogFormat, ModelConfig, ObservabilityConfig, PolicyKind, ProxyConfig,
+    ServerConfig, StaticUrlsDiscoveryConfig, StickyConfig, StickyFallbackKind, DEFAULT_FUSE,
 };
 
 const DEFAULT_KV_INDEXER_QUERY_TIMEOUT_MS: u64 = 100;
@@ -73,19 +73,32 @@ pub struct Cli {
     /// Multiplicative load spread gating the absolute balance check.
     #[arg(long)]
     pub balance_rel_threshold: Option<f32>,
-    /// External KV indexer gRPC endpoint used as the authoritative cache signal.
-    /// Needs an explicit scheme, e.g. `http://10.0.0.1:50051`.
+    /// External KV indexer gRPC endpoint used as the cache signal.
     #[arg(long)]
     pub kv_indexer_endpoint: Option<String>,
     /// KV Indexer query timeout in milliseconds. Requires
     /// `--kv-indexer-endpoint`; defaults to 100.
     #[arg(long)]
     pub kv_indexer_query_timeout_ms: Option<u64>,
-    /// Maximum concurrent KV Indexer queries issued by this Router. Requires
+    /// Maximum concurrent KV Indexer queries. Requires
     /// `--kv-indexer-endpoint`; defaults to 32.
     #[arg(long)]
     pub kv_indexer_query_max_inflight: Option<usize>,
 
+    /// Weighted terms for `--policy fused_score`.
+    #[arg(long, value_delimiter = ',')]
+    pub fuse: Vec<FusedTerm>,
+
+    /// Ordered hard constraints applied before policy selection.
+    #[arg(long, value_delimiter = ',')]
+    pub filter: Vec<FilterKind>,
+    /// Router-local in-flight limit for `--filter overloaded`.
+    #[arg(long)]
+    pub max_in_flight: Option<usize>,
+    /// Minimum cached prompt share for `--filter prefix_cache`.
+    #[arg(long)]
+    pub prefix_cache_min_share: Option<f32>,
+
     // ---- sticky-session policy (only used by `--policy sticky`) ----
     /// Request header carrying the routing key for sticky-session routing.
     /// Defaults to `x-sgl-routing-key` when `--policy sticky` is set.
@@ -96,7 +109,7 @@ pub struct Cli {
     /// `round_robin` / `random` / `power_of_two` / `load_based`. Defaults
     /// to `round_robin`.
     #[arg(long, value_enum)]
-    pub sticky_fallback_policy: Option<PolicyKind>,
+    pub sticky_fallback_policy: Option<StickyFallbackKind>,
     /// Evict a sticky assignment after it has been idle (unreferenced) this
     /// many seconds. Defaults to 600.
     #[arg(long)]
@@ -179,7 +192,8 @@ impl Cli {
             || self.kv_indexer_query_max_inflight.is_some();
         if tuned_cache_aware && self.policy != PolicyKind::CacheAwareZmq {
             return Err(anyhow!(
-                "cache-aware tuning flags require --policy cache_aware_zmq"
+                "--cache-threshold / --balance-abs-threshold / --balance-rel-threshold \
+                 require --policy cache_aware_zmq"
             ));
         }
         if self.kv_indexer_query_timeout_ms == Some(0) {
@@ -202,6 +216,57 @@ impl Cli {
                 "--kv-indexer-query-max-inflight requires --kv-indexer-endpoint"
             ));
         }
+
+        if !self.fuse.is_empty() && self.policy != PolicyKind::FusedScore {
+            return Err(anyhow!("--fuse requires --policy fused_score"));
+        }
+        let fused = if self.policy == PolicyKind::FusedScore {
+            let terms = if self.fuse.is_empty() {
+                DEFAULT_FUSE
+                    .iter()
+                    .map(|&kind| FusedTerm { kind, weight: None })
+                    .collect()
+            } else {
+                self.fuse.clone()
+            };
+            for (i, t) in terms.iter().enumerate() {
+                if terms[..i].iter().any(|p| p.kind == t.kind) {
+                    return Err(anyhow!("--fuse: `{}` is listed more than once", t.kind));
+                }
+            }
+            Some(terms)
+        } else {
+            None
+        };
+
+        for (i, kind) in self.filter.iter().enumerate() {
+            if self.filter[..i].contains(kind) {
+                return Err(anyhow!("--filter: `{kind}` is listed more than once"));
+            }
+        }
+        let has = |k: FilterKind| self.filter.contains(&k);
+        if self.max_in_flight.is_some() != has(FilterKind::Overloaded) {
+            return Err(anyhow!(
+                "--max-in-flight and `--filter overloaded` require each other"
+            ));
+        }
+        if self.prefix_cache_min_share.is_some() != has(FilterKind::PrefixCache) {
+            return Err(anyhow!(
+                "--prefix-cache-min-share and `--filter prefix_cache` require each other"
+            ));
+        }
+        if self
+            .prefix_cache_min_share
+            .is_some_and(|s| !(s > 0.0 && s <= 1.0))
+        {
+            return Err(anyhow!("--prefix-cache-min-share must be in (0, 1]"));
+        }
+        let eligibility = (!self.filter.is_empty()).then(|| EligibilityConfig {
+            filters: self.filter.clone(),
+            max_in_flight: self.max_in_flight,
+            min_prefix_share: self.prefix_cache_min_share,
+        });
+
         let tuned_sticky = self.routing_key_header.is_some()
             || self.sticky_fallback_policy.is_some()
             || self.sticky_idle_secs.is_some()
@@ -213,11 +278,10 @@ impl Cli {
             ));
         }
 
-        // Build (and validate) the sticky config exactly when the sticky
+        // Build and validate the sticky config exactly when the sticky
         // policy is selected. The header name must parse as an HTTP header
         // name so a typo fails at startup rather than silently never
-        // matching any request header; the fallback must be a
-        // dependency-free policy the factory can build standalone.
+        // matching any request header.
         let sticky = if self.policy == PolicyKind::Sticky {
             let d = StickyConfig::default();
             let header_name = self.routing_key_header.unwrap_or(d.header_name);
@@ -225,15 +289,6 @@ impl Cli {
                 anyhow!("--routing-key-header {header_name:?} is not a valid HTTP header name: {e}")
             })?;
             let fallback_policy = self.sticky_fallback_policy.unwrap_or(d.fallback_policy);
-            if matches!(
-                fallback_policy,
-                PolicyKind::Sticky | PolicyKind::CacheAwareZmq
-            ) {
-                return Err(anyhow!(
-                    "--sticky-fallback-policy must be one of round_robin / random / \
-                     power_of_two / load_based; cache_aware_zmq and sticky are not allowed"
-                ));
-            }
             let idle_secs = self.sticky_idle_secs.unwrap_or(d.idle_secs);
             let eviction_interval_secs = self
                 .sticky_eviction_interval_secs
@@ -315,6 +370,8 @@ impl Cli {
                 circuit_breaker,
                 cache_aware,
                 sticky,
+                fused,
+                eligibility,
             },
             discovery,
             proxy: ProxyConfig {
@@ -401,7 +458,7 @@ fn join_selector(terms: &[String]) -> Option<String> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::config::{DiscoveryBackend, K8sDiscoveryMode};
+    use crate::config::{DiscoveryBackend, K8sDiscoveryMode, ScoreTermKind};
 
     /// Parse argv (without the leading binary name) into a `Config`.
     fn into_config(args: &[&str]) -> Result<Config> {
@@ -703,6 +760,40 @@ mod tests {
         );
     }
 
+    #[test]
+    fn policy_accepts_only_routing_strategies() {
+        for value in ["prefix_cache", "overloaded"] {
+            let err = into_config_owned(with_model(&[
+                "--worker-urls",
+                "http://x:30000",
+                "--policy",
+                value,
+            ]))
+            .expect_err("score terms and filters are not top-level policies")
+            .to_string();
+            assert!(err.contains(value), "{value}: {err}");
+        }
+    }
+
+    #[test]
+    fn filters_and_fuse_terms_reject_non_members() {
+        let cases = [
+            (vec!["--filter", "load_based"], "load_based"),
+            (
+                vec!["--policy", "fused_score", "--fuse", "sticky"],
+                "sticky",
+            ),
+        ];
+        for (args, value) in cases {
+            let err = into_config_owned(with_model(
+                &[&["--worker-urls", "http://x:30000"], &args[..]].concat(),
+            ))
+            .expect_err("the option must reject a kind from another layer")
+            .to_string();
+            assert!(err.contains(value), "{value}: {err}");
+        }
+    }
+
     /// `--policy load_based` parses to the load-based selector.
     #[test]
     fn parses_load_based_policy() {
@@ -967,11 +1058,35 @@ mod tests {
         assert_eq!(c.model.policy, PolicyKind::Sticky);
         let s = c.model.sticky.expect("sticky config built");
         assert_eq!(s.header_name, "x-sgl-routing-key");
-        assert_eq!(s.fallback_policy, PolicyKind::RoundRobin);
+        assert_eq!(s.fallback_policy, StickyFallbackKind::RoundRobin);
         assert_eq!(s.idle_secs, 600);
         assert_eq!(s.eviction_interval_secs, 60);
     }
 
+    #[test]
+    fn sticky_fallback_help_lists_only_dependency_free_policies() {
+        use clap::CommandFactory;
+
+        let mut command = Cli::command();
+        let mut help = Vec::new();
+        command.write_long_help(&mut help).unwrap();
+        let help = String::from_utf8(help).unwrap();
+        let (_, after) = help
+            .split_once("--sticky-fallback-policy <STICKY_FALLBACK_POLICY>")
+            .expect("sticky fallback option is documented");
+        let choices = after
+            .split_once("--sticky-idle-secs")
+            .expect("sticky fallback precedes its tuning")
+            .0;
+
+        for value in ["round_robin", "random", "power_of_two", "load_based"] {
+            assert!(choices.contains(value), "missing {value}: {choices}");
+        }
+        for value in ["fused_score", "cache_aware_zmq", "sticky"] {
+            assert!(!choices.contains(value), "unexpected {value}: {choices}");
+        }
+    }
+
     #[test]
     fn sticky_flags_override_defaults() {
         let c = into_config_owned(with_model(&[
@@ -991,11 +1106,74 @@ mod tests {
         .unwrap();
         let s = c.model.sticky.expect("sticky config built");
         assert_eq!(s.header_name, "x-session-id");
-        assert_eq!(s.fallback_policy, PolicyKind::LoadBased);
+        assert_eq!(s.fallback_policy, StickyFallbackKind::LoadBased);
         assert_eq!(s.idle_secs, 120);
         assert_eq!(s.eviction_interval_secs, 15);
     }
 
+    #[test]
+    fn filter_builds_the_eligibility_config_in_order_and_is_off_by_default() {
+        let c = into_config_owned(with_model(&[
+            "--worker-urls",
+            "http://x:30000",
+            "--policy",
+            "round_robin",
+            "--filter",
+            "overloaded,prefix_cache",
+            "--max-in-flight",
+            "64",
+            "--prefix-cache-min-share",
+            "0.6",
+        ]))
+        .unwrap();
+        let e = c.model.eligibility.expect("--filter must build the config");
+        assert_eq!(
+            e.filters,
+            vec![FilterKind::Overloaded, FilterKind::PrefixCache],
+            "order is priority, so it must survive parsing",
+        );
+        assert_eq!((e.max_in_flight, e.min_prefix_share), (Some(64), Some(0.6)));
+        assert_eq!(
+            c.model.policy,
+            PolicyKind::RoundRobin,
+            "not gated on --policy"
+        );
+
+        let bare = into_config_owned(with_model(&["--worker-urls", "http://x:30000"])).unwrap();
+        assert!(bare.model.eligibility.is_none(), "no --filter, no layer");
+    }
+
+    #[test]
+    fn filter_misconfigurations_fail_at_startup() {
+        let cases: [(&[&str], &str); 6] = [
+            (&["--filter", "overloaded"], "require each other"),
+            (&["--max-in-flight", "64"], "require each other"),
+            (&["--filter", "prefix_cache"], "require each other"),
+            (&["--prefix-cache-min-share", "0.6"], "require each other"),
+            (
+                &["--filter", "overloaded,overloaded", "--max-in-flight", "64"],
+                "listed more than once",
+            ),
+            (
+                &[
+                    "--filter",
+                    "prefix_cache",
+                    "--prefix-cache-min-share",
+                    "0.0",
+                ],
+                "must be in (0, 1]",
+            ),
+        ];
+        for (extra, want) in cases {
+            let mut args = vec!["--worker-urls", "http://x:30000"];
+            args.extend_from_slice(extra);
+            let err = into_config_owned(with_model(&args))
+                .unwrap_err()
+                .to_string();
+            assert!(err.contains(want), "for {extra:?} got: {err}");
+        }
+    }
+
     #[test]
     fn non_sticky_policy_leaves_sticky_none() {
         let c = into_config_owned(with_model(&[
@@ -1049,10 +1227,7 @@ mod tests {
         ]))
         .unwrap_err()
         .to_string();
-        assert!(
-            err.contains("--sticky-fallback-policy must be one of"),
-            "got: {err}"
-        );
+        assert!(err.contains("invalid value"), "got: {err}");
     }
 
     #[test]
@@ -1067,10 +1242,7 @@ mod tests {
         ]))
         .unwrap_err()
         .to_string();
-        assert!(
-            err.contains("--sticky-fallback-policy must be one of"),
-            "got: {err}"
-        );
+        assert!(err.contains("invalid value"), "got: {err}");
     }
 
     /// A zero eviction interval would panic `tokio::time::interval` at
@@ -1110,4 +1282,93 @@ mod tests {
             "got: {err}"
         );
     }
+
+    /// `argv` is space-split, so a case reads as the command line an operator
+    /// would type. Model + worker URL are supplied.
+    fn cfg_of(argv: &str) -> Result<Config> {
+        let extra: Vec<&str> = argv.split_whitespace().collect();
+        into_config_owned(with_model(
+            &[&["--worker-urls", "http://10.0.0.1:30000"], &extra[..]].concat(),
+        ))
+    }
+
+    fn fuse_err(argv: &str) -> String {
+        cfg_of(argv).unwrap_err().to_string()
+    }
+
+    /// Resolved terms as `(kind, weight)` pairs; `None` when the policy is
+    /// not `fused_score` and so builds no term list at all.
+    fn fused_of(argv: &str) -> Option<Vec<(ScoreTermKind, Option<f32>)>> {
+        let ts = cfg_of(argv).unwrap().model.fused?;
+        Some(ts.iter().map(|t| (t.kind, t.weight)).collect())
+    }
+
+    fn fuse_ok(argv: &str) -> Vec<(ScoreTermKind, Option<f32>)> {
+        fused_of(argv).expect("fused_score builds a term list")
+    }
+
+    /// `--policy fused_score` alone composes the useful pair, and `--fuse`
+    /// overrides it with names + optional per-term weights.
+    #[test]
+    fn fuse_defaults_to_the_useful_pair_and_parses_weights() {
+        use ScoreTermKind::{LoadBased, PrefixCache, Random};
+        let pair = [(PrefixCache, None), (LoadBased, None)];
+        assert_eq!(fuse_ok("--policy fused_score"), pair);
+        // Comma-separated, order preserved, weight optional per term.
+        assert_eq!(
+            fuse_ok("--policy fused_score --fuse load_based=0.3,random"),
+            [(LoadBased, Some(0.3)), (Random, None)],
+        );
+        assert!(fused_of("").is_none(), "round_robin builds no term list");
+    }
+
+    /// Non-finite and negative weights are refused, naming the term.
+    ///
+    /// `nan`/`inf` matter more than they look: `str::parse::<f32>` accepts
+    /// both, and a NaN weight makes every worker's fused total NaN, so argmax
+    /// discards them all and the router silently degrades to least-load.
+    #[test]
+    fn fuse_rejects_non_finite_and_negative_weights() {
+        for bad in ["nan", "NaN", "inf", "-inf", "-0.5", "banana"] {
+            let err = fuse_err(&format!("--policy fused_score --fuse load_based={bad}"));
+            assert!(err.contains("load_based"), "{bad}: names the term: {err}");
+            assert!(
+                err.contains("must be finite and >= 0") || err.contains("is not a number"),
+                "{bad}: {err}",
+            );
+        }
+        for good in ["0", "0.3", "2", "1e3"] {
+            let got = fuse_ok(&format!("--policy fused_score --fuse load_based={good}"))[0].1;
+            assert_eq!(got, Some(good.parse::<f32>().unwrap()));
+        }
+    }
+
+    #[test]
+    fn fuse_rejects_malformed_compositions() {
+        let cases: [(&str, &[&str]); 5] = [
+            ("--fuse load_based", &["--fuse requires", "fused_score"]),
+            (
+                "--policy fused_score --fuse fused_score,load_based",
+                &["fused_score", "not a score term"],
+            ),
+            (
+                "--policy fused_score --fuse load_based,load_based",
+                &["load_based", "listed more than once"],
+            ),
+            (
+                "--policy fused_score --fuse not_a_policy",
+                &["not_a_policy", "is not a score term"],
+            ),
+            (
+                "--policy sticky --sticky-fallback-policy prefix_cache",
+                &["prefix_cache", "invalid value"],
+            ),
+        ];
+        for (argv, wants) in cases {
+            let err = fuse_err(argv);
+            for want in wants {
+                assert!(err.contains(want), "{argv}: want {want:?}, got: {err}");
+            }
+        }
+    }
 }

experimental/sgl-router/src/config/mod.rs

@@ -88,6 +88,8 @@ mod tests {
                 circuit_breaker: None,
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
                 urls: urls.iter().map(|s| s.to_string()).collect(),

experimental/sgl-router/src/config/types.rs

@@ -70,7 +70,8 @@ impl Default for ActiveLoadConfig {
 /// policy factory.
 ///
 /// Accepted on the CLI (`--policy`) as `round_robin` / `random` /
-/// `power_of_two` / `load_based` / `cache_aware_zmq` / `sticky`.
+/// `power_of_two` / `load_based` / `fused_score` / `cache_aware_zmq` /
+/// `sticky`.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default, clap::ValueEnum)]
 pub enum PolicyKind {
     #[default]
@@ -83,6 +84,9 @@ pub enum PolicyKind {
     /// Selects the currently least-loaded worker.
     #[value(name = "load_based")]
     LoadBased,
+    /// Weighted sum of `--fuse` terms.
+    #[value(name = "fused_score")]
+    FusedScore,
     /// Cache-aware routing fed by SGLang's ZMQ KV-cache event publisher.
     /// Requires the model to have a tokenizer loaded; cache_aware tuning
     /// lives on `ModelConfig::cache_aware`.
@@ -97,6 +101,78 @@ pub enum PolicyKind {
     Sticky,
 }
 
+impl std::fmt::Display for PolicyKind {
+    /// The CLI spelling for this policy kind.
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let v = <Self as clap::ValueEnum>::to_possible_value(self)
+            .expect("PolicyKind skips no variants");
+        f.write_str(v.get_name())
+    }
+}
+
+/// A hard admission constraint accepted by `--filter`.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, clap::ValueEnum)]
+pub enum FilterKind {
+    /// Router-local in-flight capacity limit.
+    #[value(name = "overloaded")]
+    Overloaded,
+    /// Requires a minimum share of cached prompt blocks.
+    #[value(name = "prefix_cache")]
+    PrefixCache,
+}
+
+impl std::fmt::Display for FilterKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let v = <Self as clap::ValueEnum>::to_possible_value(self)
+            .expect("FilterKind skips no variants");
+        f.write_str(v.get_name())
+    }
+}
+
+/// A soft scoring term accepted by `--fuse`.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, clap::ValueEnum)]
+pub enum ScoreTermKind {
+    /// Independent uniform-random preference.
+    #[value(name = "random")]
+    Random,
+    /// Prefers the least router-local active load.
+    #[value(name = "load_based")]
+    LoadBased,
+    /// Prefers the largest local prefix-cache overlap.
+    #[value(name = "prefix_cache")]
+    PrefixCache,
+}
+
+impl std::fmt::Display for ScoreTermKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let v = <Self as clap::ValueEnum>::to_possible_value(self)
+            .expect("ScoreTermKind skips no variants");
+        f.write_str(v.get_name())
+    }
+}
+
+/// Policy choices that can initialize or handle a keyless sticky request.
+/// These policies have no request-scoped cache or sticky-state dependency.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, clap::ValueEnum)]
+pub enum StickyFallbackKind {
+    #[value(name = "round_robin")]
+    RoundRobin,
+    #[value(name = "random")]
+    Random,
+    #[value(name = "power_of_two")]
+    PowerOfTwo,
+    #[value(name = "load_based")]
+    LoadBased,
+}
+
+impl std::fmt::Display for StickyFallbackKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let v = <Self as clap::ValueEnum>::to_possible_value(self)
+            .expect("StickyFallbackKind skips no variants");
+        f.write_str(v.get_name())
+    }
+}
+
 #[derive(Debug, Clone)]
 pub struct ServerConfig {
     pub host: String,
@@ -146,7 +222,7 @@ pub struct ModelConfig {
     pub tokenizer_path: String,
     pub policy: PolicyKind,
     pub circuit_breaker: Option<CircuitBreakerConfig>,
-    /// Tuning for cache-aware routing. Ignored unless
+    /// Tuning for the cache-aware ZMQ policy. Ignored unless
     /// `policy = "cache_aware_zmq"`. `None` falls back to defaults at
     /// policy construction time.
     pub cache_aware: Option<CacheAwareConfig>,
@@ -155,6 +231,10 @@ pub struct ModelConfig {
     /// The chat handler reads `sticky.header_name` to populate
     /// [`crate::policies::SelectionContext::routing_key`].
     pub sticky: Option<StickyConfig>,
+    /// Terms for `policy = "fused_score"`.
+    pub fused: Option<Vec<FusedTerm>>,
+    /// Hard constraints applied before policy selection.
+    pub eligibility: Option<EligibilityConfig>,
 }
 
 /// External KV Indexer client settings.
@@ -165,6 +245,54 @@ pub struct KvIndexerEndpointConfig {
     pub query_max_inflight: usize,
 }
 
+/// Eligibility filter configuration.
+#[derive(Debug, Clone, Default)]
+pub struct EligibilityConfig {
+    /// Filters in priority order.
+    pub filters: Vec<FilterKind>,
+    /// `overloaded`: in-flight count at which a worker stops being eligible.
+    pub max_in_flight: Option<usize>,
+    /// `prefix_cache` minimum cached prompt share.
+    pub min_prefix_share: Option<f32>,
+}
+
+/// Default `--policy fused_score` terms.
+pub const DEFAULT_FUSE: [ScoreTermKind; 2] = [ScoreTermKind::PrefixCache, ScoreTermKind::LoadBased];
+
+/// One `--fuse` policy and optional weight.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct FusedTerm {
+    pub kind: ScoreTermKind,
+    /// Weight override; `None` keeps the term's own `Criterion::weight()`.
+    pub weight: Option<f32>,
+}
+
+impl std::str::FromStr for FusedTerm {
+    type Err = String;
+    fn from_str(s: &str) -> Result<Self, String> {
+        let (name, weight) = match s.split_once('=') {
+            Some((n, w)) => (n, Some(parse_fuse_weight(n, w)?)),
+            None => (s, None),
+        };
+        let kind = <ScoreTermKind as clap::ValueEnum>::from_str(name, false)
+            .map_err(|_| format!("--fuse: `{name}` is not a score term"))?;
+        Ok(FusedTerm { kind, weight })
+    }
+}
+
+/// Parses a finite, non-negative term weight.
+fn parse_fuse_weight(name: &str, raw: &str) -> Result<f32, String> {
+    let w: f32 = raw
+        .parse()
+        .map_err(|_| format!("--fuse: `{name}` weight `{raw}` is not a number"))?;
+    if !w.is_finite() || w < 0.0 {
+        return Err(format!(
+            "--fuse: `{name}` weight `{raw}` must be finite and >= 0"
+        ));
+    }
+    Ok(w)
+}
+
 /// Per-model cache-aware tuning.
 #[derive(Debug, Clone)]
 pub struct CacheAwareConfig {
@@ -182,8 +310,7 @@ pub struct CacheAwareConfig {
     /// that the absolute check is gated on. Default 1.1 — 10 % relative
     /// difference triggers re-balancing.
     pub balance_rel_threshold: f32,
-    /// Optional external KV Indexer client configuration. When configured, it
-    /// replaces the local ZMQ radix tree as the cache signal.
+    /// Optional external KV Indexer client configuration.
     pub kv_indexer_endpoint: Option<KvIndexerEndpointConfig>,
 }
 
@@ -215,8 +342,7 @@ pub const DEFAULT_STICKY_HEADER: &str = "x-sgl-routing-key";
 
 /// Per-model sticky-session tuning. Built from the `--routing-key-header`
 /// / `--sticky-*` flags by [`crate::config::cli::Cli::into_config`], which
-/// also validates that `header_name` parses as an HTTP header name and
-/// that `fallback_policy` is one of the dependency-free policies.
+/// also validates that `header_name` parses as an HTTP header name.
 #[derive(Debug, Clone)]
 pub struct StickyConfig {
     /// Request header carrying the routing key. Validated to parse as a
@@ -226,9 +352,8 @@ pub struct StickyConfig {
     /// to pick the initial worker when a new key is first seen. One of
     /// `round_robin` / `random` / `power_of_two` / `load_based` — the
     /// dependency-free policies the factory can build standalone (no
-    /// `HashTree` / tokenizer / ZMQ feed). `cache_aware_zmq` and `sticky`
-    /// are rejected at config-build time.
-    pub fallback_policy: PolicyKind,
+    /// `HashTree` / tokenizer / ZMQ feed).
+    pub fallback_policy: StickyFallbackKind,
     /// Evict an assignment after it has been idle (unreferenced) this many
     /// seconds. Bounds the map against unbounded routing-key cardinality.
     pub idle_secs: u64,
@@ -247,7 +372,7 @@ impl Default for StickyConfig {
     fn default() -> Self {
         Self {
             header_name: DEFAULT_STICKY_HEADER.to_string(),
-            fallback_policy: PolicyKind::RoundRobin,
+            fallback_policy: StickyFallbackKind::RoundRobin,
             idle_secs: default_sticky_idle_secs(),
             eviction_interval_secs: default_sticky_eviction_interval_secs(),
         }

experimental/sgl-router/src/policies/cache_aware_zmq.rs

@@ -144,14 +144,12 @@ impl CacheAwareZmqPolicy {
             return None;
         }
 
-        // The index may include unhealthy workers or workers in another pool.
         let best_routable_blocks = matches
             .iter()
             .filter(|m| workers.iter().any(|worker| worker.url == m.address))
             .map(|m| m.matched_prefix_blocks)
             .max()
             .unwrap_or(0);
-
         let match_rate = best_routable_blocks as f32 / signal.query_blocks as f32;
         if let Some(metrics) = self.metrics.get() {
             metrics.observe_overlap_blocks(ctx.model().0.as_str(), best_routable_blocks as u64);
@@ -184,8 +182,6 @@ impl Policy for CacheAwareZmqPolicy {
             return Self::pick_min_load(workers);
         }
 
-        // An external signal is authoritative: an empty/unusable result
-        // degrades only to min-load and never consults the local radix tree.
         if let Some(signal) = ctx.external_prefix() {
             return self
                 .select_external(workers, ctx, signal)
@@ -349,6 +345,8 @@ mod tests {
                 circuit_breaker: None,
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: crate::config::DiscoveryBackend::StaticUrls(
                 crate::config::StaticUrlsDiscoveryConfig {
@@ -400,7 +398,7 @@ mod tests {
     }
 
     #[test]
-    fn external_prefix_signal_skips_unroutable_best_match() {
+    fn external_prefix_signal_selects_the_best_routable_match() {
         let mut config = cfg_default();
         config.cache_threshold = 0.0;
         let policy = CacheAwareZmqPolicy::new(
@@ -438,7 +436,7 @@ mod tests {
     }
 
     #[test]
-    fn external_empty_result_uses_min_load_without_local_tree() {
+    fn external_empty_result_uses_min_load() {
         let tree = Arc::new(HashTree::new());
         let registry = tokenizer_registry_with_tiny();
         let text = "hello world hello world hello world";

experimental/sgl-router/src/policies/factory.rs

@@ -1,7 +1,9 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
 // SPDX-License-Identifier: Apache-2.0
 
-use crate::config::{Config, ModelConfig, PolicyKind};
+use crate::config::{
+    Config, FilterKind, ModelConfig, PolicyKind, ScoreTermKind, StickyFallbackKind,
+};
 use crate::discovery::ModelId;
 use crate::policies::{
     cache_aware_zmq::CacheAwareZmqPolicy,
@@ -10,27 +12,25 @@ use crate::policies::{
     power_of_two::PowerOfTwoChoicesPolicy,
     random::RandomPolicy,
     round_robin::RoundRobinPolicy,
+    scoring::{
+        admission::Overloaded, prefix_cache, prefix_cache::PrefixCachePolicy, FusedScorePolicy,
+        Pipeline,
+    },
     sticky::StickyPolicy,
     Policy, PolicyRegistry,
 };
 use crate::tokenizer::TokenizerRegistry;
-use anyhow::Result;
+use anyhow::{anyhow, Result};
 use std::sync::Arc;
 use std::time::Duration;
 
-/// Build a dependency-free policy for use as the sticky-session fallback
-/// (keyless requests + initial pin of a new key). `Cli::into_config`
-/// validates `--sticky-fallback-policy` to one of these four, so the
-/// `CacheAwareZmq`/`Sticky` arms are never reached in practice.
-fn build_sticky_fallback(kind: PolicyKind) -> Arc<dyn Policy> {
+/// Build a dependency-free policy for keyless sticky requests and new pins.
+fn build_sticky_fallback(kind: StickyFallbackKind) -> Arc<dyn Policy> {
     match kind {
-        PolicyKind::RoundRobin => Arc::new(RoundRobinPolicy::new()),
-        PolicyKind::Random => Arc::new(RandomPolicy::new()),
-        PolicyKind::PowerOfTwo => Arc::new(PowerOfTwoChoicesPolicy::new()),
-        PolicyKind::LoadBased => Arc::new(LoadBasedPolicy::new()),
-        PolicyKind::CacheAwareZmq | PolicyKind::Sticky => {
-            unreachable!("sticky fallback is validated to be dependency-free in Cli::into_config")
-        }
+        StickyFallbackKind::RoundRobin => Arc::new(RoundRobinPolicy::new()),
+        StickyFallbackKind::Random => Arc::new(RandomPolicy::new()),
+        StickyFallbackKind::PowerOfTwo => Arc::new(PowerOfTwoChoicesPolicy::new()),
+        StickyFallbackKind::LoadBased => Arc::new(LoadBasedPolicy::new()),
     }
 }
 
@@ -46,20 +46,38 @@ fn build_sticky(model: &ModelConfig) -> Arc<dyn Policy> {
     ))
 }
 
-/// Construct a policy for a single model from its [`ModelConfig`] and the
-/// process-shared `HashTree` + `TokenizerRegistry` + `BlockSizeOracle`.
-///
-/// The tree, tokenizer registry, and oracle are only consulted by the
-/// cache-aware-zmq variant; other policies ignore them. Callers building
-/// all policies for the same process pass the same instances to every
-/// model.
+/// Constructs a policy for one model.
 pub fn build_policy(
     model: &ModelConfig,
     tree: Arc<HashTree>,
     tokenizers: Arc<TokenizerRegistry>,
     block_size_oracle: Arc<BlockSizeOracle>,
-) -> Arc<dyn Policy> {
-    match model.policy {
+) -> Result<Arc<dyn Policy>> {
+    let inner = build_kind(model.policy, model, &tree, &tokenizers, &block_size_oracle)?;
+    let Some(elig) = model.eligibility.as_ref().filter(|e| !e.filters.is_empty()) else {
+        return Ok(inner);
+    };
+    let mut filters = Vec::with_capacity(elig.filters.len());
+    for &kind in &elig.filters {
+        filters.push(build_filter(kind, model, &tree, &block_size_oracle));
+    }
+    Ok(Arc::new(Pipeline::new(filters, inner)?))
+}
+
+/// Build one policy kind with the shared model dependencies.
+fn build_kind(
+    kind: PolicyKind,
+    model: &ModelConfig,
+    tree: &Arc<HashTree>,
+    tokenizers: &Arc<TokenizerRegistry>,
+    block_size_oracle: &Arc<BlockSizeOracle>,
+) -> Result<Arc<dyn Policy>> {
+    let (tree, tokenizers, block_size_oracle) = (
+        Arc::clone(tree),
+        Arc::clone(tokenizers),
+        Arc::clone(block_size_oracle),
+    );
+    Ok(match kind {
         PolicyKind::RoundRobin => Arc::new(RoundRobinPolicy::new()),
         PolicyKind::Random => Arc::new(RandomPolicy::new()),
         PolicyKind::PowerOfTwo => Arc::new(PowerOfTwoChoicesPolicy::new()),
@@ -74,32 +92,90 @@ pub fn build_policy(
             ))
         }
         PolicyKind::Sticky => build_sticky(model),
+        PolicyKind::FusedScore => build_fused(model, &tree, &block_size_oracle)?,
+    })
+}
+
+/// Builds one hard admission filter with the shared model dependencies.
+fn build_filter(
+    kind: FilterKind,
+    model: &ModelConfig,
+    tree: &Arc<HashTree>,
+    block_size_oracle: &Arc<BlockSizeOracle>,
+) -> Arc<dyn Policy> {
+    match kind {
+        FilterKind::Overloaded => {
+            let cap = (model.eligibility.as_ref())
+                .and_then(|e| e.max_in_flight)
+                .unwrap_or(usize::MAX);
+            Arc::new(Overloaded::new(cap))
+        }
+        FilterKind::PrefixCache => {
+            let share = (model.eligibility.as_ref())
+                .and_then(|e| e.min_prefix_share)
+                .unwrap_or(0.0);
+            Arc::new(
+                PrefixCachePolicy::new(
+                    Arc::clone(tree),
+                    Arc::clone(block_size_oracle),
+                    prefix_cache::DEFAULT_WEIGHT,
+                )
+                .with_min_share(share),
+            )
+        }
     }
 }
 
-/// Compatibility shim used by tests + non-cache-aware code paths. Builds
-/// a policy without wiring the cache-aware dependencies; rejects
-/// `CacheAwareZmq` to keep the call sites that don't have a `HashTree` /
-/// `TokenizerRegistry` to hand from accidentally compiling.
-#[cfg(test)]
-pub fn build_policy_kind_only(kind: PolicyKind) -> Arc<dyn Policy> {
+/// Builds one soft scoring term for `--policy fused_score`.
+fn build_score(
+    kind: ScoreTermKind,
+    tree: &Arc<HashTree>,
+    block_size_oracle: &Arc<BlockSizeOracle>,
+) -> Arc<dyn Policy> {
     match kind {
+        ScoreTermKind::Random => Arc::new(RandomPolicy::new()),
+        ScoreTermKind::LoadBased => Arc::new(LoadBasedPolicy::new()),
+        ScoreTermKind::PrefixCache => Arc::new(PrefixCachePolicy::new(
+            Arc::clone(tree),
+            Arc::clone(block_size_oracle),
+            prefix_cache::DEFAULT_WEIGHT,
+        )),
+    }
+}
+
+/// Builds `--policy fused_score`.
+fn build_fused(
+    model: &ModelConfig,
+    tree: &Arc<HashTree>,
+    oracle: &Arc<BlockSizeOracle>,
+) -> Result<Arc<dyn Policy>> {
+    let spec = model.fused.as_deref().unwrap_or_default();
+    if spec.is_empty() {
+        return Err(anyhow!(
+            "--policy fused_score needs at least one --fuse term"
+        ));
+    }
+    let mut terms: Vec<(Arc<dyn Policy>, Option<f32>)> = Vec::with_capacity(spec.len());
+    for t in spec {
+        terms.push((build_score(t.kind, tree, oracle), t.weight));
+    }
+    Ok(Arc::new(FusedScorePolicy::new(terms)?))
+}
+
+/// Builds a policy with test defaults.
+#[cfg(test)]
+pub fn build_policy_kind_only(kind: PolicyKind) -> Result<Arc<dyn Policy>> {
+    Ok(match kind {
         PolicyKind::RoundRobin => Arc::new(RoundRobinPolicy::new()),
         PolicyKind::Random => Arc::new(RandomPolicy::new()),
         PolicyKind::PowerOfTwo => Arc::new(PowerOfTwoChoicesPolicy::new()),
         PolicyKind::LoadBased => Arc::new(LoadBasedPolicy::new()),
-        PolicyKind::CacheAwareZmq => {
-            // Provide an empty tree + empty tokenizer registry + fresh
-            // oracle so the test policy is constructible. Production
-            // callers go through `build_policy` with the real
-            // process-shared instances.
-            Arc::new(CacheAwareZmqPolicy::new(
-                crate::config::CacheAwareConfig::default(),
-                Arc::new(HashTree::new()),
-                Arc::new(TokenizerRegistry::default()),
-                BlockSizeOracle::new(),
-            ))
-        }
+        PolicyKind::CacheAwareZmq => Arc::new(CacheAwareZmqPolicy::new(
+            crate::config::CacheAwareConfig::default(),
+            Arc::new(HashTree::new()),
+            Arc::new(TokenizerRegistry::default()),
+            BlockSizeOracle::new(),
+        )),
         PolicyKind::Sticky => {
             let s = crate::config::StickyConfig::default();
             Arc::new(StickyPolicy::new(
@@ -108,7 +184,10 @@ pub fn build_policy_kind_only(kind: PolicyKind) -> Arc<dyn Policy> {
                 build_sticky_fallback(s.fallback_policy),
             ))
         }
-    }
+        PolicyKind::FusedScore => {
+            return Err(anyhow!("--policy {kind} needs --fuse terms from the model"))
+        }
+    })
 }
 
 pub fn build_registry(
@@ -126,19 +205,12 @@ pub fn build_registry(
             Arc::clone(&tree),
             Arc::clone(&tokenizers),
             Arc::clone(&block_size_oracle),
-        ),
+        )?,
     );
     Ok(reg)
 }
 
-/// Convenience for tests + non-cache-aware callers: builds a registry with
-/// a fresh, empty `HashTree` and an empty `TokenizerRegistry`. The
-/// cache-aware-zmq policy will then degrade to min-load (no tokenizer +
-/// no worker-published block size → fallback) — which is exactly what
-/// the legacy tests assume.
-///
-/// Production callers go through [`build_registry`] with the real
-/// process-shared instances.
+/// Builds a registry with empty cache-aware dependencies.
 pub fn build_registry_with_defaults(cfg: &Config) -> Result<PolicyRegistry> {
     build_registry(
         cfg,
@@ -156,7 +228,90 @@ mod tests {
         StaticUrlsDiscoveryConfig,
     };
 
-    use crate::config::PolicyKind;
+    use crate::config::{
+        EligibilityConfig, FilterKind, PolicyKind, ScoreTermKind, StickyFallbackKind,
+    };
+    use crate::discovery::{WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::SelectionContext;
+    use crate::workers::Worker;
+
+    fn worker(id: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(id.into()),
+            url: format!("http://{id}:30000"),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("modelA".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    #[test]
+    fn filter_overloaded_wires_through_the_factory() {
+        let mut cfg = cfg_with_model("modelA", PolicyKind::LoadBased);
+        cfg.model.eligibility = Some(EligibilityConfig {
+            filters: vec![FilterKind::Overloaded],
+            max_in_flight: Some(2),
+            min_prefix_share: None,
+        });
+        let reg = build_registry_with_defaults(&cfg).unwrap();
+        let p = reg.get(&ModelId("modelA".into())).unwrap();
+
+        let ws = vec![worker("w0"), worker("w1")];
+        let model = ModelId("modelA".into());
+        let ctx = SelectionContext::new(&model, None);
+        let _one = ws[1].load_guard();
+        assert_eq!(
+            p.select(&ws, &ctx).unwrap().id,
+            ws[0].id,
+            "load still ranks"
+        );
+
+        let _fill: Vec<_> = (ws.iter())
+            .flat_map(|w| (0..2).map(|_| w.load_guard()))
+            .collect();
+        assert!(
+            p.select(&ws, &ctx).is_none(),
+            "every worker is over the cap"
+        );
+    }
+
+    #[test]
+    fn a_filter_must_actually_constrain() {
+        let mut cfg = cfg_with_model("modelA", PolicyKind::LoadBased);
+        let built = |c: &Config| {
+            build_registry_with_defaults(c).map(|r| r.get(&ModelId("modelA".into())).unwrap())
+        };
+        assert!(
+            !built(&cfg).unwrap().can_filter(),
+            "no floor, so the term stays a pure preference",
+        );
+
+        cfg.model.eligibility = Some(EligibilityConfig {
+            filters: vec![FilterKind::PrefixCache],
+            max_in_flight: None,
+            min_prefix_share: Some(0.6),
+        });
+        assert!(
+            built(&cfg).unwrap().needs_request_tokens(),
+            "the floor reads the prompt"
+        );
+
+        cfg.model.eligibility = Some(EligibilityConfig {
+            filters: vec![FilterKind::PrefixCache],
+            max_in_flight: None,
+            min_prefix_share: Some(0.6),
+        });
+        let filter = build_filter(
+            FilterKind::PrefixCache,
+            &cfg.model,
+            &Arc::new(HashTree::new()),
+            &BlockSizeOracle::new(),
+        );
+        assert!(
+            filter.can_filter(),
+            "a configured prefix-cache floor is a filter",
+        );
+    }
 
     fn cfg_with_model(id: &str, policy: PolicyKind) -> Config {
         Config {
@@ -172,6 +327,8 @@ mod tests {
                 circuit_breaker: None,
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
                 urls: vec!["http://placeholder:0".into()],
@@ -183,13 +340,68 @@ mod tests {
 
     #[test]
     fn build_policy_kind_only_covers_all_variants() {
-        // Trivially total — the match is exhaustive over `PolicyKind`.
-        let _ = build_policy_kind_only(PolicyKind::RoundRobin);
-        let _ = build_policy_kind_only(PolicyKind::Random);
-        let _ = build_policy_kind_only(PolicyKind::PowerOfTwo);
-        let _ = build_policy_kind_only(PolicyKind::LoadBased);
-        let _ = build_policy_kind_only(PolicyKind::CacheAwareZmq);
-        let _ = build_policy_kind_only(PolicyKind::Sticky);
+        for kind in [
+            PolicyKind::RoundRobin,
+            PolicyKind::Random,
+            PolicyKind::PowerOfTwo,
+            PolicyKind::LoadBased,
+            PolicyKind::CacheAwareZmq,
+            PolicyKind::Sticky,
+        ] {
+            assert!(build_policy_kind_only(kind).is_ok(), "{kind:?}");
+        }
+        assert!(build_policy_kind_only(PolicyKind::FusedScore).is_err());
+    }
+
+    #[test]
+    fn prefix_cache_builds_as_a_score_term() {
+        let p = build_score(
+            ScoreTermKind::PrefixCache,
+            &Arc::new(HashTree::new()),
+            &BlockSizeOracle::new(),
+        );
+        assert!(p.can_fuse(), "prefix_cache must be usable as a --fuse term");
+    }
+
+    #[test]
+    fn fused_score_builds_score_terms_and_rejects_an_empty_config() {
+        let mut cfg = cfg_with_model("m", PolicyKind::FusedScore);
+        let term = |kind, weight| crate::config::FusedTerm { kind, weight };
+
+        for weight in [None, Some(2.5)] {
+            cfg.model.fused = Some(vec![term(ScoreTermKind::PrefixCache, weight)]);
+            assert!(build_registry_with_defaults(&cfg).is_ok(), "{weight:?}");
+        }
+
+        cfg.model.fused = Some(vec![]);
+        assert!(build_registry_with_defaults(&cfg)
+            .unwrap_err()
+            .to_string()
+            .contains("at least one --fuse term"));
+    }
+
+    #[test]
+    fn fused_score_accepts_an_outer_eligibility_pipeline() {
+        let mut cfg = cfg_with_model("modelA", PolicyKind::FusedScore);
+        cfg.model.fused = Some(vec![crate::config::FusedTerm {
+            kind: ScoreTermKind::LoadBased,
+            weight: None,
+        }]);
+        cfg.model.eligibility = Some(EligibilityConfig {
+            filters: vec![FilterKind::Overloaded],
+            max_in_flight: Some(0),
+            min_prefix_share: None,
+        });
+
+        let policy = build_registry_with_defaults(&cfg)
+            .expect("an outer filter must not make fused terms non-fusable")
+            .get(&ModelId("modelA".into()))
+            .unwrap();
+        let workers = vec![worker("w0"), worker("w1")];
+        let model = ModelId("modelA".into());
+        assert!(policy
+            .select(&workers, &SelectionContext::new(&model, None))
+            .is_none());
     }
 
     #[test]
@@ -209,9 +421,6 @@ mod tests {
         let tokenizers = Arc::new(TokenizerRegistry::default());
         let reg = build_registry(&cfg, tree, tokenizers, BlockSizeOracle::new()).unwrap();
         let p = reg.get(&ModelId("modelA".into())).unwrap();
-        // Down-cast probe via Debug — cheaper than carrying a type-tag
-        // on the trait. Pinning the debug repr is fine because the field
-        // name is part of the file's public test surface.
         let dbg = format!("{p:?}");
         assert!(
             dbg.contains("CacheAwareZmqPolicy"),
@@ -246,4 +455,22 @@ mod tests {
             "expected StickyPolicy debug repr, got: {dbg}",
         );
     }
+
+    #[test]
+    fn sticky_fallback_builder_covers_all_typed_choices() {
+        let workers = vec![worker("w0"), worker("w1")];
+        let model = ModelId("modelA".into());
+        let ctx = SelectionContext::new(&model, None);
+        for kind in [
+            StickyFallbackKind::RoundRobin,
+            StickyFallbackKind::Random,
+            StickyFallbackKind::PowerOfTwo,
+            StickyFallbackKind::LoadBased,
+        ] {
+            assert!(
+                build_sticky_fallback(kind).select(&workers, &ctx).is_some(),
+                "{kind:?}"
+            );
+        }
+    }
 }

experimental/sgl-router/src/policies/kv_events/tree.rs

@@ -501,6 +501,74 @@ impl TreeState {
         }
     }
 
+    /// Contiguous prefix depth for every worker on the chain, in one descent.
+    /// `insert` marks a worker at every node it descends, so "holds the first
+    /// `d` blocks" means present at each of levels `1..=d`. A worker is frozen
+    /// at the first level that omits it, so a `remove`d interior node stops the
+    /// count at the hole instead of counting past it.
+    fn prefix_depths(
+        &self,
+        parent_hash: Option<i64>,
+        block_hashes: &[i64],
+    ) -> HashMap<KvWorkerId, usize> {
+        if block_hashes.is_empty() {
+            return HashMap::new();
+        }
+        // Same start resolution as `match_prefix`: an ambiguous `parent_hash`
+        // has no worker context to disambiguate, so fall back to root.
+        let start = match parent_hash {
+            None => ROOT_ID,
+            Some(p) => match self.by_hash.get(&p) {
+                Some(set) if set.len() == 1 => *set.iter().next().unwrap(),
+                _ => ROOT_ID,
+            },
+        };
+        // Keys borrow the arena for the walk; ids are cloned once on the way out.
+        let mut depths: HashMap<&KvWorkerId, usize> = HashMap::new();
+        let mut alive: Vec<&KvWorkerId> = Vec::new();
+        let mut current = start;
+        let mut reached = 0usize;
+        let now = now_millis();
+        for &h in block_hashes {
+            let Some(child_id) = self
+                .nodes
+                .get(&current)
+                .and_then(|n| n.children.get(&h).copied())
+            else {
+                break;
+            };
+            let Some(child) = self.nodes.get(&child_id) else {
+                break;
+            };
+            child.last_used.store(now, Ordering::Relaxed);
+            current = child_id;
+            reached += 1;
+            if reached == 1 {
+                // Absent here means never in `alive`: a tail held without
+                // block 0 is reported as holding nothing.
+                alive = child.workers.iter().collect();
+            } else {
+                let mut still = Vec::with_capacity(alive.len());
+                for w in alive {
+                    if child.workers.contains(w) {
+                        still.push(w);
+                    } else {
+                        depths.insert(w, reached - 1);
+                    }
+                }
+                alive = still;
+            }
+            if alive.is_empty() {
+                break;
+            }
+        }
+        // Whoever is still tracked held every level the walk reached.
+        for w in alive {
+            depths.insert(w, reached);
+        }
+        depths.into_iter().map(|(w, d)| (w.clone(), d)).collect()
+    }
+
     /// Approximate count of *non-root* nodes in the tree.
     fn node_count(&self) -> usize {
         // Subtract one for the root sentinel.
@@ -652,6 +720,18 @@ impl HashTree {
         state.match_prefix(parent_hash, block_hashes)
     }
 
+    /// How many leading blocks of `block_hashes` each worker holds contiguously,
+    /// in one descent under one read lock. [`Self::match_prefix`] names only the
+    /// deepest matched node's holders, so it cannot answer this. Absent = none.
+    pub fn prefix_depths(
+        &self,
+        parent_hash: Option<i64>,
+        block_hashes: &[i64],
+    ) -> HashMap<KvWorkerId, usize> {
+        let state = self.state.read();
+        state.prefix_depths(parent_hash, block_hashes)
+    }
+
     /// Approximate number of non-root nodes in the tree (the root sentinel
     /// is not counted). Useful for metrics and to decide when to call
     /// [`HashTree::evict_lru`].
@@ -702,6 +782,36 @@ mod tests {
         ids.iter().map(|w| (*w).clone()).collect()
     }
 
+    /// One descent answers for every worker at its *own* depth, where
+    /// `match_prefix` names only whoever sits at the deepest matched node.
+    /// Also pins that a `remove`d interior block stops the count at the hole.
+    #[test]
+    fn prefix_depths_answers_every_worker_at_its_own_depth() {
+        let chain = [1i64, 2, 3, 4];
+        let (deep, shallow, holed) = (
+            worker("http://a", 0),
+            worker("http://b", 0),
+            worker("http://c", 0),
+        );
+        let tree = HashTree::new();
+        tree.insert(&deep, None, &chain);
+        tree.insert(&shallow, None, &chain[..2]);
+        tree.insert(&holed, None, &chain);
+        tree.remove(&holed, &chain[1..2]);
+
+        let depths = tree.prefix_depths(None, &chain);
+        assert_eq!(depths.get(&deep), Some(&4), "holds the whole chain");
+        assert_eq!(depths.get(&shallow), Some(&2), "holds two of four");
+        assert_eq!(depths.get(&holed), Some(&1), "stops at the cleared block");
+        assert_eq!(depths.get(&worker("http://d", 0)), None, "holds nothing");
+
+        // Contiguity matters: `remove` left `holed` listed at the deepest node,
+        // so `match_prefix` credits it with the full chain scored here at 1.
+        let m = tree.match_prefix(None, &chain);
+        assert_eq!(m.matched_blocks, 4);
+        assert_eq!(m.workers, workers(&[&deep, &holed]));
+    }
+
     #[test]
     fn empty_match_returns_zero_no_workers() {
         let tree = HashTree::new();

experimental/sgl-router/src/policies/load_based.rs

@@ -1,14 +1,12 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
 // SPDX-License-Identifier: Apache-2.0
 
-use crate::policies::{Policy, SelectionContext};
+use crate::policies::scoring::ScoringPolicy;
+use crate::policies::SelectionContext;
 use crate::workers::Worker;
 use std::sync::Arc;
 
-/// Deterministic load-based policy.
-///
-/// Chooses the candidate with the lowest current `Worker::active_load`.
-/// Ties follow the candidate slice order, which is registry-dependent.
+/// Prefers the least-loaded candidate; `select()` is the blanket impl's.
 #[derive(Debug, Default)]
 pub struct LoadBasedPolicy;
 
@@ -16,18 +14,23 @@ impl LoadBasedPolicy {
     pub fn new() -> Self {
         Self
     }
-
-    pub fn pick_min_load(workers: &[Arc<Worker>]) -> Option<Arc<Worker>> {
-        workers
-            .iter()
-            .min_by_key(|w| w.active_load())
-            .map(Arc::clone)
-    }
 }
 
-impl Policy for LoadBasedPolicy {
-    fn select(&self, workers: &[Arc<Worker>], _ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
-        Self::pick_min_load(workers)
+impl ScoringPolicy for LoadBasedPolicy {
+    /// `1.0` for the least loaded down to `0.0` for the most, min-max scaled to
+    /// the CURRENT fleet -- relative, not absolute, so it cannot saturate:
+    /// `1 - load/256` reads a busy fleet as all-`0.0`, tied inside
+    /// `TIE_EPSILON`, so the term dies exactly when load matters most.
+    ///
+    /// Purely a preference: "everybody is busy" is not a reason to refuse to
+    /// route, so this term never constrains. Capacity is `--filter`'s job.
+    fn scores(&self, workers: &[Arc<Worker>], _ctx: &SelectionContext<'_>) -> Vec<f32> {
+        let loads: Vec<usize> = workers.iter().map(|w| w.active_load()).collect();
+        let lo = loads.iter().min().copied().unwrap_or(0);
+        let span = (loads.iter().max().copied().unwrap_or(0) - lo) as f32;
+        // `max(1.0)` is exact: a zero span means every `l - lo` is zero too.
+        let score = |l: usize| 1.0 - (l - lo) as f32 / span.max(1.0);
+        loads.into_iter().map(score).collect()
     }
 }
 
@@ -35,6 +38,8 @@ impl Policy for LoadBasedPolicy {
 mod tests {
     use super::*;
     use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::scoring::argmax::TIE_EPSILON;
+    use crate::policies::Policy;
 
     fn worker(id: &str) -> Arc<Worker> {
         Arc::new(Worker::new(WorkerSpec {
@@ -46,25 +51,38 @@ mod tests {
         }))
     }
 
-    #[test]
+    #[test] // upstream's, retargeted from `pick_min_load` onto blanket `select`
     fn empty_returns_none() {
-        let policy = LoadBasedPolicy::new();
-        let model = ModelId("tiny".into());
-        let ctx = SelectionContext::new(&model, None);
-        assert!(policy.select(&[], &ctx).is_none());
+        let m = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&m, None);
+        assert!(LoadBasedPolicy::new().select(&[], &ctx).is_none());
     }
 
+    /// `select()` alone CANNOT detect a broken score: ARGMAX breaks a tie on
+    /// load, so a constant `scores()` still lands on the minimum and that arm
+    /// passes for the wrong reason. Ranking is therefore asserted on the vector
+    /// itself, strictly outside `TIE_EPSILON` so a saturating curve cannot hide
+    /// in the tie band -- what `300,900` is for. NaN needs its own arm because
+    /// no ORDERING sees it: it makes every comparison false, which on `0,0` is
+    /// the expected answer. Upstream's `picks_lowest_active_load` goes under
+    /// rule 4 -- the unique-minimum 2-worker case, which `0,1` subsumes.
     #[test]
-    fn picks_lowest_active_load() {
-        let policy = LoadBasedPolicy::new();
+    fn scores_rank_strictly_by_load_and_the_choice_lands_on_the_minimum() {
         let model = ModelId("tiny".into());
-        let ctx = SelectionContext::new(&model, None);
-        let w0 = worker("w0");
-        let w1 = worker("w1");
-        let _g0 = w0.load_guard();
-        assert_eq!(
-            policy.select(&[w0, Arc::clone(&w1)], &ctx).unwrap().id,
-            w1.id
-        );
+        let (ctx, p) = (SelectionContext::new(&model, None), LoadBasedPolicy::new());
+        for spec in ["0,1", "2,1,0", "1,0,1", "0,0", "5,2,9,2", "300,900"] {
+            let loads: Vec<usize> = spec.split(',').map(|s| s.parse().unwrap()).collect();
+            let ws: Vec<Arc<Worker>> = (0..loads.len()).map(|i| worker(&format!("w{i}"))).collect();
+            let _held: Vec<_> = (ws.iter().zip(&loads))
+                .flat_map(|(w, n)| (0..*n).map(move |_| w.load_guard()))
+                .collect();
+            let scores = p.scores(&ws, &ctx);
+            for (i, j) in (0..loads.len()).flat_map(|i| (0..loads.len()).map(move |j| (i, j))) {
+                let ok = (scores[i] > scores[j] + TIE_EPSILON, scores[i].is_nan());
+                assert_eq!(ok, (loads[i] < loads[j], false), "{spec} scored {scores:?}");
+            }
+            let got = p.select(&ws, &ctx).expect("non-empty").active_load();
+            assert_eq!(got, *loads.iter().min().expect("non-empty"), "{spec}");
+        }
     }
 }

experimental/sgl-router/src/policies/mod.rs

@@ -10,55 +10,34 @@ pub mod power_of_two;
 pub mod random;
 pub mod registry;
 pub mod round_robin;
+pub mod scoring;
 pub mod sticky;
 
 use crate::discovery::ModelId;
+use crate::policies::scoring::{EligibilityFilter, ScoringPolicy};
 use crate::server::metrics::MetricsRegistry;
 use crate::tokenizer::{adapter, TokenizerRegistry};
 use crate::workers::Worker;
 use dashmap::DashMap;
 use std::sync::Arc;
 
-/// Tokens produced once at ingress for a request. Consumed by the
-/// cache-aware selection decision and, when `engine_equivalent`, forwarded
-/// to the engine as `input_ids` so the engine skips its own prompt
-/// tokenization (the router and engine would otherwise tokenize the same
-/// prompt twice in the same cluster).
+/// Tokens produced at ingress for routing and optional engine forwarding.
 pub struct RequestTokens {
     /// The prompt token ids.
     pub ids: Vec<u32>,
-    /// True only when the ids were produced via the model's chat encoder —
-    /// i.e. they match what the engine would tokenize from the chat
-    /// template. False for the raw-prompt fallback, where the engine must
-    /// tokenize the text itself, so the ids are NOT safe to forward.
+    /// Whether the token ids are safe to forward as engine `input_ids`.
     pub engine_equivalent: bool,
 }
 
-/// External indexer answer prepared by the async ingress path for the
-/// synchronous cache-aware policy.
+/// External indexer answer prepared by the async ingress path for a
+/// cache-aware policy.
 pub struct ExternalPrefixSignal {
     pub outcome: sgl_kv_indexer::PrefixOutcome,
     pub query_blocks: usize,
 }
 
-/// Produce the routing tokens — and whether they are engine-equivalent —
-/// from an already-parsed request body, using the shared tokenizer registry.
-///
-/// Tokenization is a property of the MODEL (does it have a chat encoder?),
-/// not of the routing policy, so this lives here as a free function the
-/// ingress calls directly with `ctx.tokenizers` — every policy (sticky,
-/// round-robin, cache-aware) shares one tokenize. The cache-aware policy also
-/// calls it as a body-tokenize fallback for callers that didn't pre-tokenize.
-///
-/// Chat requests (`messages`) on a model that has a chat encoder are rendered
-/// through that encoder and tokenized the way the engine does, so the query
-/// hashes match the engine's cached blocks (chat-templated tokens) AND the ids
-/// are safe to hand the engine as `input_ids` (`engine_equivalent = true`).
-/// Everything else — `/v1/completions` (`prompt`), `/generate` (`text`), or a
-/// chat model without an encoder — tokenizes the raw extracted prompt text;
-/// those ids only match the engine after it applies its own template, so they
-/// are NOT engine-equivalent. A failed encoder render/encode falls through to
-/// the raw path rather than failing the request.
+/// Tokenizes a request for routing. Chat-encoder tokens are engine-equivalent;
+/// raw prompt tokens are used only for routing.
 pub fn request_tokens_for(
     tokenizers: &TokenizerRegistry,
     model_id: &ModelId,
@@ -82,11 +61,7 @@ pub fn request_tokens_for(
     })
 }
 
-/// Tokenize `text` for `model_id` via the shared registry. Returns `None` if
-/// no tokenizer is loaded (the model_id may be misconfigured) or if encoding
-/// fails / yields no tokens. An encode error logs at WARN (a loaded-but-erroring
-/// tokenizer silently disables the offload); the no-text / empty-output paths
-/// are expected and stay quiet.
+/// Tokenizes text with the model tokenizer.
 fn tokenize_text(
     tokenizers: &TokenizerRegistry,
     model_id: &ModelId,
@@ -97,13 +72,6 @@ fn tokenize_text(
         Ok(ids) if !ids.is_empty() => Some(ids),
         Ok(_) => None,
         Err(e) => {
-            // WARN, not DEBUG: a tokenizer that is loaded but consistently
-            // erroring silently turns the whole tokenization offload into a
-            // no-op, so the failure must be visible above DEBUG. Sustained
-            // failure logs once per request; the volume signal is the
-            // `sgl_router_ingress_tokenize_errors_total` counter (which the
-            // chat handler bumps on the chat-encode failure), so no
-            // rate-limiter here.
             tracing::warn!(
                 model = %model_id,
                 error = %e,
@@ -114,10 +82,7 @@ fn tokenize_text(
     }
 }
 
-/// Extract a raw prompt-text candidate from an already-parsed JSON request
-/// body. Returns `None` when there's no routable text field; the caller then
-/// skips tokenization. This is the raw path — chat requests on a model with a
-/// chat encoder are rendered via the encoder instead (see [`request_tokens_for`]).
+/// Extracts raw prompt text from a parsed request body.
 ///
 /// Supported shapes (in priority order):
 ///   1. `"prompt": "..."` — `/v1/completions`-style.
@@ -129,7 +94,6 @@ fn tokenize_text(
 ///      multimodal content blocks; text-only blocks concatenated.
 ///   5. `"text": "..."` — SGLang `/generate` native form.
 ///
-/// Anything else yields `None`.
 pub(crate) fn extract_prompt_text_from_value(v: &serde_json::Value) -> Option<String> {
     if let Some(s) = v.get("prompt").and_then(|p| p.as_str()) {
         return Some(s.to_string());
@@ -173,16 +137,7 @@ pub(crate) fn extract_prompt_text_from_value(v: &serde_json::Value) -> Option<St
     None
 }
 
-/// Selection input — carries the request body and the routing tokens
-/// (computed once at ingress) so cache-aware policies can hash prefix
-/// tokens without reshaping the [`Policy`] trait or re-tokenizing.  Today's
-/// load-only policies (round-robin, random, power-of-two, load-based) read
-/// only `workers`; sticky reads `routing_key`.
-///
-/// Constructed via [`Self::new`] / [`Self::with_routing_key`]; the
-/// ingress-computed tokens are attached with [`Self::with_request_tokens`].
-/// Accessors expose immutable references so callers cannot mutate the model
-/// id or swap in a different body without going through the constructor.
+/// Immutable request data consumed by a routing policy.
 pub struct SelectionContext<'a> {
     model: &'a ModelId,
     request_body: Option<&'a [u8]>,
@@ -216,9 +171,7 @@ impl<'a> SelectionContext<'a> {
         }
     }
 
-    /// Attach the ingress-computed routing tokens. When present, the
-    /// cache-aware policy consumes these instead of re-parsing and
-    /// re-tokenizing the body.
+    /// Attaches ingress-computed routing tokens.
     pub fn with_request_tokens(mut self, request_tokens: Option<&'a [u32]>) -> Self {
         self.request_tokens = request_tokens;
         self
@@ -244,8 +197,7 @@ impl<'a> SelectionContext<'a> {
         self.routing_key
     }
 
-    /// Ingress-precomputed routing tokens, if any. `None` means the policy
-    /// must derive tokens itself (e.g. a caller that didn't pre-tokenize).
+    /// Returns ingress-computed routing tokens.
     pub fn request_tokens(&self) -> Option<&[u32]> {
         self.request_tokens
     }
@@ -258,25 +210,33 @@ impl<'a> SelectionContext<'a> {
 pub trait Policy: Send + Sync + std::fmt::Debug {
     fn select(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Option<Arc<Worker>>;
 
-    /// Whether this policy's ROUTING decision needs the request tokens (i.e.
-    /// it routes by prompt prefix). Ingress tokenization itself is no longer
-    /// gated on this — that is a model property (`has_chat_encoder`) decided at
-    /// ingress via [`request_tokens_for`]. This flag is the EXTRA gate that
-    /// keeps the cache-aware policy's RAW-prompt routing path alive: a
-    /// cache-aware model with no chat encoder still wants its `/v1/completions`
-    /// /`text` prompt tokenized for tree matching, which `has_chat_encoder`
-    /// alone would not trigger. Default `false` (load-only + sticky route
-    /// without prefix tokens); only the cache-aware policy overrides it.
+    /// Whether policy selection needs request tokens.
     fn needs_request_tokens(&self) -> bool {
         false
     }
 
-    /// Attach the process metrics registry after construction. Default is a
-    /// no-op — only policies that emit metrics (cache-aware-zmq's
-    /// `sgl_router_overlap_blocks`) override it. Mirrors
-    /// `ActiveLoadRegistry::attach_metrics`: the registry is built after the
-    /// policies, so it is injected here rather than passed to the constructor.
+    /// Attaches the process metrics registry after construction.
     fn attach_metrics(&self, _metrics: Arc<MetricsRegistry>) {}
+
+    /// Returns the optional per-worker scoring view.
+    fn as_scoring(&self) -> Option<&dyn ScoringPolicy> {
+        None
+    }
+
+    /// Returns the optional per-worker eligibility view.
+    fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+        None
+    }
+
+    /// Whether this policy exposes scores for `--fuse`.
+    fn can_fuse(&self) -> bool {
+        self.as_scoring().is_some()
+    }
+
+    /// Whether this policy exposes an eligibility filter.
+    fn can_filter(&self) -> bool {
+        self.as_filter().is_some()
+    }
 }
 
 #[derive(Debug, Default)]
@@ -293,9 +253,7 @@ impl PolicyRegistry {
         self.by_model.get(model).map(|p| p.clone())
     }
 
-    /// Inject the metrics registry into every registered policy. Called once
-    /// at startup (after the registry is built) so metrics-emitting policies
-    /// can record into the shared registry.
+    /// Attaches metrics to each registered policy.
     pub fn attach_metrics(&self, metrics: Arc<MetricsRegistry>) {
         for entry in self.by_model.iter() {
             entry.value().attach_metrics(Arc::clone(&metrics));

experimental/sgl-router/src/policies/random.rs

@@ -1,9 +1,9 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
 // SPDX-License-Identifier: Apache-2.0
 
-use crate::policies::{Policy, SelectionContext};
+use crate::policies::{scoring::ScoringPolicy, SelectionContext};
 use crate::workers::Worker;
-use rand::seq::SliceRandom;
+use rand::Rng;
 use std::sync::Arc;
 
 #[derive(Debug, Default)]
@@ -15,8 +15,52 @@ impl RandomPolicy {
     }
 }
 
-impl Policy for RandomPolicy {
-    fn select(&self, workers: &[Arc<Worker>], _ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
-        workers.choose(&mut rand::thread_rng()).cloned()
+impl ScoringPolicy for RandomPolicy {
+    /// Argmax of n iid uniforms IS a uniform choice: exactly the old `choose`.
+    /// Never constrains: a coin toss is not an eligibility rule.
+    fn scores(&self, workers: &[Arc<Worker>], _ctx: &SelectionContext<'_>) -> Vec<f32> {
+        let mut rng = rand::thread_rng();
+        (0..workers.len()).map(|_| rng.gen()).collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::Policy;
+
+    fn worker(id: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(id.into()),
+            url: format!("http://{id}:30000"),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("tiny".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    /// Distributional: `select()` is not pure. Marginals alone are satisfied by
+    /// a ROTATION -- what a constant `scores()` becomes under ARGMAX's rotating
+    /// tiebreak -- so REPEATS share the band: P(pick==prev) is 1/N iid, 0 rotating.
+    #[test]
+    fn picks_uniformly_over_20k_draws_and_repeats_at_the_iid_rate() {
+        const MEAN: f64 = 5_000.0; // 20_000 draws over 4 workers; repeats too
+        const BAND: f64 = 5.0 * 61.237_244; // 5 sigma, sqrt(20_000 / 4 * 3 / 4)
+        let (policy, model) = (RandomPolicy::new(), ModelId("tiny".into()));
+        let ctx = SelectionContext::new(&model, None);
+        let ws: Vec<Arc<Worker>> = (0..4).map(|i| worker(&format!("w{i}"))).collect();
+        assert!(policy.select(&[], &ctx).is_none(), "empty fleet");
+        let (mut counts, mut repeats, mut prev) = ([0usize; 4], 0usize, None);
+        for _ in 0..20_000 {
+            let got = policy.select(&ws, &ctx).expect("non-empty fleet");
+            let i = ws.iter().position(|w| w.id == got.id).expect("a candidate");
+            counts[i] += 1;
+            repeats += usize::from(prev.replace(i) == Some(i));
+        }
+        for n in counts.iter().chain([&repeats]) {
+            let dev = (*n as f64 - MEAN).abs(); // `> 0` is no-starvation
+            assert!(*n > 0 && dev < BAND, "{counts:?} {repeats}");
+        }
     }
 }

experimental/sgl-router/src/policies/scoring/admission.rs

@@ -0,0 +1,123 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
+// SPDX-License-Identifier: Apache-2.0
+
+//! Router-local in-flight admission control.
+
+use super::{EligibilityFilter, OnEmpty};
+use crate::policies::{Policy, SelectionContext};
+use crate::workers::Worker;
+use std::sync::Arc;
+
+/// Rejects workers at the router-local in-flight limit.
+#[derive(Debug)]
+pub struct Overloaded {
+    max_in_flight: usize,
+}
+
+impl Overloaded {
+    pub fn new(max_in_flight: usize) -> Self {
+        Self { max_in_flight }
+    }
+}
+
+impl EligibilityFilter for Overloaded {
+    fn keep(&self, workers: &[Arc<Worker>], _ctx: &SelectionContext<'_>) -> Vec<bool> {
+        (workers.iter())
+            .map(|w| w.active_load() < self.max_in_flight)
+            .collect()
+    }
+
+    /// Do not route to an over-capacity worker.
+    fn on_empty(&self) -> OnEmpty {
+        OnEmpty::Hold
+    }
+}
+
+impl Policy for Overloaded {
+    fn select(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
+        let eligible: Vec<Arc<Worker>> = (workers.iter())
+            .zip(self.keep(workers, ctx))
+            .filter(|(_, ok)| *ok)
+            .map(|(w, _)| Arc::clone(w))
+            .collect();
+        eligible
+            .iter()
+            .min_by_key(|w| w.active_load())
+            .map(Arc::clone)
+    }
+
+    fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+        Some(self)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::scoring::{admit, refs};
+
+    fn worker(id: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(id.into()),
+            url: format!("http://{id}:30000"),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("tiny".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    #[test]
+    fn the_cap_is_a_strict_ceiling() {
+        let ws = vec![worker("idle"), worker("under"), worker("at")];
+        let _under: Vec<_> = (0..2).map(|_| ws[1].load_guard()).collect();
+        let _at: Vec<_> = (0..3).map(|_| ws[2].load_guard()).collect();
+
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+        assert_eq!(
+            Overloaded::new(3).keep(&ws, &ctx),
+            vec![true, true, false],
+            "load 3 against a cap of 3 is over",
+        );
+    }
+
+    #[test]
+    fn a_full_fleet_refuses_rather_than_picking_the_least_bad() {
+        let ws = vec![worker("a"), worker("b")];
+        let _a: Vec<_> = (0..5).map(|_| ws[0].load_guard()).collect();
+        let _b: Vec<_> = (0..9).map(|_| ws[1].load_guard()).collect();
+
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+        let full: Vec<Box<dyn EligibilityFilter>> = vec![Box::new(Overloaded::new(4))];
+        assert!(
+            admit(refs(&full), &ws, &ctx).is_none(),
+            "both over the cap, and the filter Holds",
+        );
+
+        let some: Vec<Box<dyn EligibilityFilter>> = vec![Box::new(Overloaded::new(6))];
+        let out = admit(refs(&some), &ws, &ctx).expect("a is under the cap");
+        assert_eq!(out.len(), 1);
+        assert_eq!(out[0].url, ws[0].url);
+    }
+
+    #[test]
+    fn hold_does_not_yield_to_a_later_filter() {
+        #[derive(Debug)]
+        struct AdmitAll;
+        impl EligibilityFilter for AdmitAll {
+            fn keep(&self, ws: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+                vec![true; ws.len()]
+            }
+        }
+        let ws = vec![worker("a")];
+        let _busy: Vec<_> = (0..9).map(|_| ws[0].load_guard()).collect();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let chain: Vec<Box<dyn EligibilityFilter>> =
+            vec![Box::new(Overloaded::new(2)), Box::new(AdmitAll)];
+        assert!(admit(refs(&chain), &ws, &ctx).is_none());
+    }
+}

experimental/sgl-router/src/policies/scoring/argmax.rs

@@ -0,0 +1,113 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
+// SPDX-License-Identifier: Apache-2.0
+
+//! Selects one worker from per-worker scores.
+
+use crate::workers::Worker;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::Arc;
+
+/// Scores within this distance of the best are tied.
+pub const TIE_EPSILON: f32 = 1e-6;
+
+pub trait Selector: Send + Sync + std::fmt::Debug {
+    /// Index into `workers` of the chosen candidate, or `None` when there is
+    /// nothing to choose from. `scores[i]` belongs to `workers[i]`.
+    fn pick(&self, workers: &[Arc<Worker>], scores: &[f32]) -> Option<usize>;
+}
+
+/// Highest score wins; ties choose the least-loaded candidate and rotate.
+#[derive(Debug, Default)]
+pub struct Argmax {
+    rotor: AtomicUsize,
+}
+
+/// The default selector, shared by every scoring policy that does not override
+/// [`super::ScoringPolicy::selector`].
+pub static ARGMAX: Argmax = Argmax {
+    rotor: AtomicUsize::new(0),
+};
+
+impl Selector for Argmax {
+    fn pick(&self, workers: &[Arc<Worker>], scores: &[f32]) -> Option<usize> {
+        if workers.is_empty() {
+            return None;
+        }
+        let n = workers.len().min(scores.len());
+        let best = (0..n)
+            .map(|i| scores[i])
+            .filter(|s| !s.is_nan())
+            .fold(None::<f32>, |acc, s| Some(acc.map_or(s, |b| b.max(s))));
+        let mut band: Vec<usize> = match best {
+            Some(b) => (0..n)
+                .filter(|&i| !scores[i].is_nan() && scores[i] >= b - TIE_EPSILON)
+                .collect(),
+            None => Vec::new(),
+        };
+        if band.is_empty() {
+            tracing::debug!(
+                n_workers = workers.len(),
+                n_scores = scores.len(),
+                "no usable score; falling back to load + rotation",
+            );
+            band = (0..workers.len()).collect();
+        }
+        let min_load = band.iter().map(|&i| workers[i].active_load()).min()?;
+        let tied: Vec<usize> = band
+            .into_iter()
+            .filter(|&i| workers[i].active_load() == min_load)
+            .collect();
+        let k = self.rotor.fetch_add(1, Ordering::Relaxed) % tied.len();
+        Some(tied[k])
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use std::collections::HashSet;
+
+    fn worker(id: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(id.into()),
+            url: format!("http://{id}:30000"),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("tiny".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    #[test]
+    fn score_wins_unless_the_gap_is_inside_the_tie_band() {
+        let ws = vec![worker("a"), worker("b")];
+        let sel = Argmax::default();
+        let _loaded = ws[1].load_guard();
+
+        assert_eq!(sel.pick(&ws, &[1.0, 1.0 - 1e-3]), Some(0), "clear winner");
+        let tie = [1.0 - 5e-7, 1.0];
+        assert_eq!(sel.pick(&ws, &tie), Some(0), "tie -> less load");
+        assert_eq!(sel.pick(&[], &[]), None, "nothing to choose from");
+    }
+
+    #[test]
+    fn nan_never_wins_from_either_position() {
+        let ws = vec![worker("a"), worker("b")];
+        let sel = Argmax::default();
+        assert_eq!(sel.pick(&ws, &[f32::NAN, 0.0]), Some(1));
+        assert_eq!(sel.pick(&ws, &[0.0, f32::NAN]), Some(0));
+        assert!(sel.pick(&ws, &[f32::NAN, f32::NAN]).is_some());
+    }
+
+    #[test]
+    fn a_total_tie_rotates_over_every_candidate() {
+        let ws = vec![worker("a"), worker("b"), worker("c")];
+        let sel = Argmax::default();
+        let picks: HashSet<usize> = (0..3).filter_map(|_| sel.pick(&ws, &[1.0; 3])).collect();
+        assert_eq!(
+            picks.len(),
+            3,
+            "three tied picks must cover all three: {picks:?}"
+        );
+    }
+}

experimental/sgl-router/src/policies/scoring/mod.rs

@@ -0,0 +1,579 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
+// SPDX-License-Identifier: Apache-2.0
+
+//! Candidate eligibility and scoring policies.
+
+pub mod admission;
+pub mod argmax;
+pub mod prefix_cache;
+
+use crate::policies::{Policy, SelectionContext};
+use crate::workers::Worker;
+use argmax::{Selector, ARGMAX};
+use std::sync::Arc;
+
+/// What a filter means when it has rejected every worker it was shown.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum OnEmpty {
+    /// Ignore this filter when it rejects every candidate.
+    Abstain,
+    /// Keep the rejection: no worker is admissible.
+    Hold,
+}
+
+/// A hard constraint applied before scoring.
+pub trait EligibilityFilter: Send + Sync + std::fmt::Debug {
+    /// Returns one admission flag per worker; `true` keeps the candidate.
+    fn keep(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<bool>;
+
+    /// Whether this constraint reads `ctx.request_tokens()`.
+    fn needs_tokens(&self) -> bool {
+        false
+    }
+
+    /// Controls the result when this filter rejects every candidate.
+    fn on_empty(&self) -> OnEmpty {
+        OnEmpty::Abstain
+    }
+}
+
+/// A soft preference for eligible candidates.
+pub trait ScoringPolicy: Send + Sync + std::fmt::Debug {
+    /// What each candidate is worth, parallel to `workers`; higher is better.
+    fn scores(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<f32>;
+
+    /// Default multiplier as a fused term; `--fuse name=weight` overrides it.
+    fn weight(&self) -> f32 {
+        1.0
+    }
+
+    /// Whether scoring needs request tokens.
+    fn needs_tokens(&self) -> bool {
+        false
+    }
+
+    /// Optional eligibility view for policies that provide both signals.
+    fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+        None
+    }
+
+    /// Selects a winner from the score vector.
+    fn selector(&self) -> &dyn Selector {
+        &ARGMAX
+    }
+}
+
+/// Applies ordered filters. `None` means a holding filter rejected all candidates.
+pub fn admit<'f>(
+    filters: impl IntoIterator<Item = &'f dyn EligibilityFilter>,
+    workers: &[Arc<Worker>],
+    ctx: &SelectionContext<'_>,
+) -> Option<Vec<Arc<Worker>>> {
+    let mut alive: Vec<Arc<Worker>> = workers.to_vec();
+
+    for filter in filters {
+        if alive.is_empty() {
+            break;
+        }
+        let flags = filter.keep(&alive, ctx);
+        let on_empty = filter.on_empty();
+        if flags.len() != alive.len() {
+            tracing::debug!(
+                filter = ?filter,
+                n_workers = alive.len(),
+                n_flags = flags.len(),
+                "eligibility filter returned the wrong arity",
+            );
+            if on_empty == OnEmpty::Hold {
+                return None;
+            }
+        }
+        let untouched = alive.len() == workers.len();
+
+        let next: Vec<Arc<Worker>> = (alive.iter().enumerate())
+            .filter(|(i, _)| flags.get(*i).copied().unwrap_or(true))
+            .map(|(_, w)| Arc::clone(w))
+            .collect();
+
+        if next.is_empty() {
+            match on_empty {
+                OnEmpty::Hold => return None,
+                OnEmpty::Abstain if untouched => {
+                    tracing::warn!(
+                        filter = ?filter,
+                        n_workers = workers.len(),
+                        "eligibility filter rejected every worker on its own; a filter \
+                         with no signal should abstain (all true), not veto the fleet",
+                    );
+                    continue;
+                }
+                OnEmpty::Abstain => {
+                    tracing::debug!(
+                        filter = ?filter,
+                        n_alive = alive.len(),
+                        "eligibility filter conflicts with a higher-priority one; yielding",
+                    );
+                    continue;
+                }
+            }
+        }
+        alive = next;
+    }
+
+    Some(alive)
+}
+
+/// Selects the best-scoring worker.
+impl<T: ScoringPolicy> Policy for T {
+    fn select(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
+        let scores = self.scores(workers, ctx);
+        let i = self.selector().pick(workers, &scores)?;
+        workers.get(i).map(Arc::clone)
+    }
+
+    fn needs_request_tokens(&self) -> bool {
+        ScoringPolicy::needs_tokens(self) || self.as_filter().is_some_and(|f| f.needs_tokens())
+    }
+
+    fn as_scoring(&self) -> Option<&dyn ScoringPolicy> {
+        Some(self)
+    }
+
+    fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+        ScoringPolicy::as_filter(self)
+    }
+}
+
+/// A weighted sum of scoring terms.
+#[derive(Debug)]
+pub struct FusedScorePolicy {
+    /// Terms and optional `--fuse name=weight` overrides.
+    terms: Vec<(Arc<dyn Policy>, Option<f32>)>,
+}
+
+fn view(t: &(Arc<dyn Policy>, Option<f32>)) -> (&dyn ScoringPolicy, f32) {
+    let s = t.0.as_scoring().expect("checked by FusedScorePolicy::new");
+    (s, t.1.unwrap_or_else(|| s.weight()))
+}
+
+impl FusedScorePolicy {
+    /// Reject non-scoring terms during construction.
+    pub fn new(terms: Vec<(Arc<dyn Policy>, Option<f32>)>) -> anyhow::Result<Self> {
+        for (p, _) in &terms {
+            anyhow::ensure!(p.can_fuse(), "policy {p:?} does not support fusion");
+        }
+        Ok(Self { terms })
+    }
+}
+
+/// Applies eligibility filters before an inner policy.
+#[derive(Debug)]
+pub struct Pipeline {
+    filters: Vec<Arc<dyn Policy>>,
+    inner: Arc<dyn Policy>,
+}
+
+impl Pipeline {
+    /// Reject policies that do not expose an eligibility view.
+    pub fn new(filters: Vec<Arc<dyn Policy>>, inner: Arc<dyn Policy>) -> anyhow::Result<Self> {
+        for f in &filters {
+            anyhow::ensure!(f.can_filter(), "policy {f:?} imposes no eligibility rule");
+        }
+        Ok(Self { filters, inner })
+    }
+
+    fn views(&self) -> impl Iterator<Item = &dyn EligibilityFilter> {
+        (self.filters.iter()).map(|p| p.as_filter().expect("checked by Pipeline::new"))
+    }
+}
+
+impl Policy for Pipeline {
+    fn select(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
+        let eligible = admit(self.views(), workers, ctx)?;
+        self.inner.select(&eligible, ctx)
+    }
+
+    fn needs_request_tokens(&self) -> bool {
+        self.inner.needs_request_tokens() || self.views().any(|f| f.needs_tokens())
+    }
+
+    fn attach_metrics(&self, metrics: Arc<crate::server::metrics::MetricsRegistry>) {
+        self.inner.attach_metrics(metrics);
+    }
+}
+
+impl ScoringPolicy for FusedScorePolicy {
+    fn scores(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<f32> {
+        let mut total = vec![0.0f32; workers.len()];
+        for (term, w) in self.terms.iter().map(view) {
+            for (acc, s) in total.iter_mut().zip(term.scores(workers, ctx)) {
+                *acc += w * s;
+            }
+        }
+        total
+    }
+
+    fn needs_tokens(&self) -> bool {
+        self.terms.iter().map(view).any(|(t, _)| t.needs_tokens())
+    }
+}
+
+/// Owned boxes as the borrowed views [`admit`] consumes. Shared by the tests
+/// in this module and its siblings.
+#[cfg(test)]
+pub(crate) fn refs(
+    fs: &[Box<dyn EligibilityFilter>],
+) -> impl Iterator<Item = &dyn EligibilityFilter> {
+    fs.iter().map(|f| &**f)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::round_robin::RoundRobinPolicy;
+
+    fn worker(id: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(id.into()),
+            url: format!("http://{id}:30000"),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("tiny".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    fn fleet() -> Vec<Arc<Worker>> {
+        vec![worker("a"), worker("b"), worker("c")]
+    }
+
+    fn urls(ws: &[Arc<Worker>]) -> Vec<String> {
+        ws.iter().map(|w| w.url.clone()).collect()
+    }
+
+    fn term(p: impl Policy + 'static, w: Option<f32>) -> (Arc<dyn Policy>, Option<f32>) {
+        (Arc::new(p), w)
+    }
+
+    #[derive(Debug)]
+    struct ByIndex(f32, bool, bool);
+
+    fn by(w: f32) -> ByIndex {
+        ByIndex(w, false, false)
+    }
+
+    impl ScoringPolicy for ByIndex {
+        fn scores(&self, workers: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<f32> {
+            (0..workers.len()).map(|i| i as f32).collect()
+        }
+        fn weight(&self) -> f32 {
+            self.0
+        }
+        fn needs_tokens(&self) -> bool {
+            self.2
+        }
+        fn selector(&self) -> &dyn Selector {
+            if self.1 {
+                &PICK_FIRST
+            } else {
+                &ARGMAX
+            }
+        }
+    }
+
+    #[derive(Debug)]
+    struct Keep(Vec<&'static str>, OnEmpty);
+
+    impl EligibilityFilter for Keep {
+        fn keep(&self, workers: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+            (workers.iter())
+                .map(|w| self.0.iter().any(|n| w.url.contains(n)))
+                .collect()
+        }
+        fn on_empty(&self) -> OnEmpty {
+            self.1
+        }
+    }
+
+    impl Policy for Keep {
+        fn select(&self, ws: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Option<Arc<Worker>> {
+            (ws.iter().zip(self.keep(ws, ctx)))
+                .find(|(_, ok)| *ok)
+                .map(|(w, _)| Arc::clone(w))
+        }
+        fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+            Some(self)
+        }
+    }
+
+    fn keep(names: &[&'static str], on_empty: OnEmpty) -> Box<dyn EligibilityFilter> {
+        Box::new(Keep(names.to_vec(), on_empty))
+    }
+
+    fn boxed(f: impl EligibilityFilter + 'static) -> Box<dyn EligibilityFilter> {
+        Box::new(f)
+    }
+
+    #[derive(Debug)]
+    struct PickFirst;
+    static PICK_FIRST: PickFirst = PickFirst;
+    impl Selector for PickFirst {
+        fn pick(&self, workers: &[Arc<Worker>], _: &[f32]) -> Option<usize> {
+            (!workers.is_empty()).then_some(0)
+        }
+    }
+
+    #[test]
+    fn selector_dispatch_uses_the_policys_own_selector() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let scores = by(1.0).scores(&ws, &ctx);
+        assert_eq!(ARGMAX.pick(&ws, &scores), Some(2));
+        assert_eq!(PICK_FIRST.pick(&ws, &scores), Some(0));
+
+        assert_eq!(by(1.0).select(&ws, &ctx).unwrap().id, ws[2].id);
+        let first = ByIndex(1.0, true, false);
+        assert_eq!(first.select(&ws, &ctx).unwrap().id, ws[0].id);
+    }
+
+    #[test]
+    fn can_fuse_is_derived_and_gates_construction() {
+        let fused = FusedScorePolicy::new(vec![term(by(1.0), None)]).unwrap();
+        let fusable: Vec<Arc<dyn Policy>> = vec![Arc::new(by(1.0)), Arc::new(fused)];
+        for p in &fusable {
+            assert!(p.can_fuse());
+            assert!(p.as_scoring().is_some(), "the flag agrees with the view");
+        }
+        let rr: Arc<dyn Policy> = Arc::new(RoundRobinPolicy::new());
+        assert!(!rr.can_fuse());
+        assert!(rr.as_scoring().is_none());
+        let err = FusedScorePolicy::new(vec![term(RoundRobinPolicy::new(), None)])
+            .expect_err("round_robin has no per-worker preference to contribute");
+        assert!(err.to_string().contains("does not support fusion"), "{err}");
+    }
+
+    #[test]
+    fn fusion_nests_and_the_override_replaces_the_terms_own_weight() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let inner = FusedScorePolicy::new(vec![term(by(2.0), None)]).unwrap();
+        assert_eq!(inner.scores(&ws, &ctx), vec![0.0, 2.0, 4.0], "its own 2i");
+
+        let outer =
+            FusedScorePolicy::new(vec![term(inner, None), term(by(3.0), Some(10.0))]).unwrap();
+        assert_eq!(outer.scores(&ws, &ctx), vec![0.0, 12.0, 24.0], "2i + 10i");
+        assert_eq!(outer.select(&ws, &ctx).unwrap().id, ws[2].id);
+    }
+
+    #[test]
+    fn composer_propagates_needs_tokens_from_any_term() {
+        let plain = FusedScorePolicy::new(vec![term(by(1.0), None)]).unwrap();
+        assert!(!plain.needs_request_tokens());
+        let hungry = FusedScorePolicy::new(vec![
+            term(by(1.0), None),
+            term(ByIndex(1.0, false, true), None),
+        ])
+        .unwrap();
+        assert!(hungry.needs_request_tokens(), "any one term is enough");
+
+        #[derive(Debug)]
+        struct Hungry;
+        impl EligibilityFilter for Hungry {
+            fn keep(&self, ws: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+                vec![true; ws.len()]
+            }
+            fn needs_tokens(&self) -> bool {
+                true
+            }
+        }
+        impl Policy for Hungry {
+            fn select(&self, ws: &[Arc<Worker>], _: &SelectionContext<'_>) -> Option<Arc<Worker>> {
+                ws.first().map(Arc::clone)
+            }
+            fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+                Some(self)
+            }
+        }
+        let filtered = Pipeline::new(
+            vec![Arc::new(Hungry)],
+            Arc::new(FusedScorePolicy::new(vec![term(by(1.0), None)]).unwrap()),
+        )
+        .unwrap();
+        assert!(filtered.needs_request_tokens(), "the filter is hungry");
+    }
+
+    #[test]
+    fn a_rejected_worker_cannot_be_out_weighed() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let fused = Pipeline::new(
+            vec![Arc::new(Keep(vec!["a", "b"], OnEmpty::Abstain))],
+            Arc::new(FusedScorePolicy::new(vec![term(by(1.0), Some(1e9))]).unwrap()),
+        )
+        .unwrap();
+        assert_eq!(
+            fused.select(&ws, &ctx).unwrap().url,
+            ws[1].url,
+            "the best ELIGIBLE, not the best"
+        );
+
+        let open = Pipeline::new(
+            vec![Arc::new(Keep(vec!["a", "b", "c"], OnEmpty::Abstain))],
+            Arc::new(FusedScorePolicy::new(vec![term(by(1.0), Some(1e9))]).unwrap()),
+        )
+        .unwrap();
+        assert_eq!(open.select(&ws, &ctx).unwrap().url, ws[2].url);
+    }
+
+    #[test]
+    fn a_conflict_yields_the_later_filter_and_keeps_the_earlier_narrowing() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let chain = vec![
+            keep(&["a", "b"], OnEmpty::Abstain),
+            keep(&["c"], OnEmpty::Abstain),
+        ];
+        let out = admit(refs(&chain), &ws, &ctx).expect("Abstain never holds");
+        assert_eq!(
+            urls(&out),
+            urls(&ws[..2]),
+            "the second filter yields; falling back to the raw fleet would read [a, b, c]",
+        );
+
+        let rev = vec![
+            keep(&["c"], OnEmpty::Abstain),
+            keep(&["a", "b"], OnEmpty::Abstain),
+        ];
+        assert_eq!(urls(&admit(refs(&rev), &ws, &ctx).unwrap()), urls(&ws[2..]));
+    }
+
+    #[test]
+    fn a_filter_after_a_conflict_still_applies() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let chain = vec![
+            keep(&["a", "b"], OnEmpty::Abstain),
+            keep(&["c"], OnEmpty::Abstain),
+            keep(&["b", "c"], OnEmpty::Abstain),
+        ];
+        assert_eq!(
+            urls(&admit(refs(&chain), &ws, &ctx).unwrap()),
+            vec![ws[1].url.clone()]
+        );
+    }
+
+    #[test]
+    fn a_holding_filter_refuses_instead_of_yielding() {
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+
+        let held = vec![
+            keep(&["a", "b"], OnEmpty::Abstain),
+            keep(&["c"], OnEmpty::Hold),
+        ];
+        assert!(
+            admit(refs(&held), &ws, &ctx).is_none(),
+            "no eligible worker, and the filter said Hold",
+        );
+
+        let ok = vec![
+            keep(&["a", "b"], OnEmpty::Abstain),
+            keep(&["b"], OnEmpty::Hold),
+        ];
+        assert_eq!(
+            urls(&admit(refs(&ok), &ws, &ctx).unwrap()),
+            vec![ws[1].url.clone()]
+        );
+
+        let fused = Pipeline::new(
+            vec![
+                Arc::new(Keep(vec!["a", "b"], OnEmpty::Abstain)),
+                Arc::new(Keep(vec!["c"], OnEmpty::Hold)),
+            ],
+            Arc::new(FusedScorePolicy::new(vec![term(by(1.0), None)]).unwrap()),
+        )
+        .unwrap();
+        assert!(fused.select(&ws, &ctx).is_none());
+    }
+
+    #[test]
+    fn a_short_flag_vector_degrades_instead_of_panicking() {
+        #[derive(Debug)]
+        struct Short;
+        impl EligibilityFilter for Short {
+            fn keep(&self, _: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+                vec![false]
+            }
+        }
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+        let out = admit(refs(&[boxed(Short)]), &ws, &ctx).expect("the tail was admitted");
+        assert_eq!(urls(&out), urls(&ws[1..]), "only index 0 rejected");
+    }
+
+    #[test]
+    fn a_short_hold_filter_fails_closed() {
+        #[derive(Debug)]
+        struct ShortHold;
+        impl EligibilityFilter for ShortHold {
+            fn keep(&self, _: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+                vec![true]
+            }
+
+            fn on_empty(&self) -> OnEmpty {
+                OnEmpty::Hold
+            }
+        }
+
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+        assert!(admit(refs(&[boxed(ShortHold)]), &ws, &ctx).is_none());
+    }
+
+    #[test]
+    fn a_dual_role_term_exposes_its_filter_half_through_policy() {
+        #[derive(Debug)]
+        struct Dual;
+        impl EligibilityFilter for Dual {
+            fn keep(&self, ws: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<bool> {
+                ws.iter().map(|w| !w.url.contains('c')).collect()
+            }
+            fn needs_tokens(&self) -> bool {
+                true
+            }
+        }
+        impl ScoringPolicy for Dual {
+            fn scores(&self, ws: &[Arc<Worker>], _: &SelectionContext<'_>) -> Vec<f32> {
+                vec![0.0; ws.len()]
+            }
+            fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+                Some(self)
+            }
+        }
+
+        let p: Arc<dyn Policy> = Arc::new(Dual);
+        assert!(p.can_fuse(), "it still scores");
+        let f = p.as_filter().expect("and it filters");
+        let ws = fleet();
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None);
+        assert_eq!(f.keep(&ws, &ctx), vec![true, true, false]);
+        assert!(
+            p.needs_request_tokens(),
+            "hunger comes from the filter half"
+        );
+    }
+}

experimental/sgl-router/src/policies/scoring/prefix_cache.rs

@@ -0,0 +1,264 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
+// SPDX-License-Identifier: Apache-2.0
+
+//! Prefix-cache scores from the KV-event [`HashTree`].
+
+use super::{EligibilityFilter, ScoringPolicy};
+use crate::policies::kv_events::{
+    compute_block_hashes, compute_block_hashes_bigram, BlockSizeOracle, HashTree,
+};
+use crate::policies::SelectionContext;
+use crate::workers::Worker;
+use std::sync::Arc;
+
+/// Score for a miss or unavailable prefix signal.
+const NO_HOLDING: f32 = 0.0;
+
+/// Default fused-term weight.
+pub const DEFAULT_WEIGHT: f32 = 1.0;
+
+pub struct PrefixCachePolicy {
+    tree: Arc<HashTree>,
+    block_size_oracle: Arc<BlockSizeOracle>,
+    weight: f32,
+    /// Minimum cached share for eligibility; zero disables filtering.
+    min_share: f32,
+}
+
+impl std::fmt::Debug for PrefixCachePolicy {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PrefixCachePolicy")
+            .field("weight", &self.weight)
+            .field("min_share", &self.min_share)
+            .field("tree_nodes", &self.tree.node_count())
+            .finish()
+    }
+}
+
+impl PrefixCachePolicy {
+    pub fn new(tree: Arc<HashTree>, block_size_oracle: Arc<BlockSizeOracle>, weight: f32) -> Self {
+        Self {
+            tree,
+            block_size_oracle,
+            weight,
+            min_share: 0.0,
+        }
+    }
+
+    /// Require a cached share for eligibility.
+    pub fn with_min_share(mut self, share: f32) -> Self {
+        self.min_share = share;
+        self
+    }
+}
+
+impl PrefixCachePolicy {
+    /// Returns each worker's cached prompt share.
+    fn shares(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<f32> {
+        let flat = || vec![NO_HOLDING; workers.len()];
+
+        let Some(tokens) = ctx.request_tokens().filter(|t| !t.is_empty()) else {
+            return flat();
+        };
+        let Some(block_size) = self.block_size_oracle.get() else {
+            return flat();
+        };
+        let hashes = if self.block_size_oracle.is_bigram() {
+            compute_block_hashes_bigram(tokens, block_size as usize)
+        } else {
+            compute_block_hashes(tokens, block_size as usize)
+        };
+        if hashes.is_empty() {
+            return flat();
+        }
+
+        let depths = self.tree.prefix_depths(None, &hashes);
+        let total = hashes.len() as f32;
+        workers
+            .iter()
+            .map(|w| {
+                depths
+                    .iter()
+                    .filter(|(kw, _)| kw.url == w.url)
+                    .map(|(_, &d)| d)
+                    .max()
+                    .map_or(NO_HOLDING, |d| d as f32 / total)
+            })
+            .collect()
+    }
+}
+
+impl ScoringPolicy for PrefixCachePolicy {
+    fn scores(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<f32> {
+        self.shares(workers, ctx)
+    }
+
+    fn weight(&self) -> f32 {
+        self.weight
+    }
+
+    fn as_filter(&self) -> Option<&dyn EligibilityFilter> {
+        (self.min_share > 0.0).then_some(self as &dyn EligibilityFilter)
+    }
+
+    fn needs_tokens(&self) -> bool {
+        true
+    }
+}
+
+impl EligibilityFilter for PrefixCachePolicy {
+    fn keep(&self, workers: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<bool> {
+        (self.shares(workers, ctx).into_iter())
+            .map(|share| share >= self.min_share)
+            .collect()
+    }
+
+    fn needs_tokens(&self) -> bool {
+        true
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+    use crate::policies::kv_events::KvWorkerId;
+
+    const BLOCK: usize = 4;
+
+    fn worker(url: &str) -> Arc<Worker> {
+        Arc::new(Worker::new(WorkerSpec {
+            id: WorkerId(url.into()),
+            url: url.into(),
+            mode: WorkerMode::Plain,
+            model_ids: vec![ModelId("tiny".into())],
+            bootstrap_port: None,
+        }))
+    }
+
+    fn tokens() -> Vec<u32> {
+        (0..(BLOCK as u32 * 4)).collect()
+    }
+
+    fn insert(tree: &HashTree, url: &str, rank: u32, from: usize, blocks: usize) {
+        let all = compute_block_hashes(&tokens(), BLOCK);
+        let parent = if from == 0 { None } else { Some(all[from - 1]) };
+        tree.insert(
+            &KvWorkerId::new(url.into(), rank),
+            parent,
+            &all[from..from + blocks],
+        );
+    }
+
+    fn policy(tree: Arc<HashTree>) -> PrefixCachePolicy {
+        let oracle = BlockSizeOracle::new();
+        oracle
+            .try_set(BLOCK as u32)
+            .expect("a fresh oracle accepts the first block size");
+        PrefixCachePolicy::new(tree, oracle, 1.0)
+    }
+
+    fn shares(p: &PrefixCachePolicy, ws: &[Arc<Worker>], ctx: &SelectionContext<'_>) -> Vec<f32> {
+        assert!(
+            ScoringPolicy::as_filter(p).is_none(),
+            "no floor configured, so this term must not be a filter at all",
+        );
+        p.scores(ws, ctx)
+    }
+
+    #[test]
+    fn depth_is_a_fraction_and_a_tail_without_block_zero_misses() {
+        let tree = Arc::new(HashTree::new());
+        insert(&tree, "deep", 0, 0, 3);
+        insert(&tree, "tail", 0, 2, 2);
+        let ws = vec![worker("deep"), worker("tail"), worker("cold")];
+
+        let model = ModelId("tiny".into());
+        let ids = tokens();
+        let ctx = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+
+        let scores = shares(&policy(tree), &ws, &ctx);
+        assert_eq!(scores[0], 0.75, "3 of 4 blocks held, not a neutral 1.0");
+        assert_eq!(scores[1], 0.0, "tail without block 0 holds nothing");
+        assert_eq!(scores[2], 0.0, "never seen");
+    }
+
+    #[test]
+    fn several_dp_ranks_of_one_worker_collapse_to_the_deepest() {
+        let tree = Arc::new(HashTree::new());
+        insert(&tree, "dp", 0, 0, 1);
+        insert(&tree, "dp", 1, 0, 3);
+        let ws = vec![worker("dp")];
+
+        let model = ModelId("tiny".into());
+        let ids = tokens();
+        let ctx = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+        assert_eq!(
+            shares(&policy(tree), &ws, &ctx),
+            vec![0.75],
+            "3 of 4, not 1"
+        );
+    }
+
+    #[test]
+    fn without_tokens_every_worker_scores_the_same() {
+        let tree = Arc::new(HashTree::new());
+        insert(&tree, "deep", 0, 0, 3);
+        let ws = vec![worker("deep"), worker("cold")];
+        let model = ModelId("tiny".into());
+        let policy = policy(tree);
+
+        let ids = tokens();
+        let with = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+        assert_eq!(
+            shares(&policy, &ws, &with),
+            vec![0.75, 0.0],
+            "signal is live"
+        );
+
+        let without = SelectionContext::new(&model, None);
+        assert_eq!(
+            shares(&policy, &ws, &without),
+            vec![0.0, 0.0],
+            "and inert here"
+        );
+    }
+
+    #[test]
+    fn without_a_block_size_no_worker_looks_like_a_hit() {
+        let tree = Arc::new(HashTree::new());
+        insert(&tree, "deep", 0, 0, 3);
+        let ws = vec![worker("deep"), worker("cold")];
+
+        let model = ModelId("tiny".into());
+        let ids = tokens();
+        let ctx = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+        let cold = PrefixCachePolicy::new(tree, BlockSizeOracle::new(), 1.0);
+        assert_eq!(shares(&cold, &ws, &ctx), vec![0.0, 0.0]);
+    }
+
+    #[test]
+    fn the_bigram_branch_queries_a_different_chain() {
+        let ids = tokens();
+        let unigram = compute_block_hashes(&ids, BLOCK);
+        let bigram = compute_block_hashes_bigram(&ids, BLOCK);
+        assert_ne!(unigram, bigram, "the two hashers must disagree, else this");
+
+        let tree = Arc::new(HashTree::new());
+        insert(&tree, "deep", 0, 0, 4);
+        let ws = vec![worker("deep")];
+        let model = ModelId("tiny".into());
+        let ctx = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+
+        let oracle = BlockSizeOracle::new();
+        oracle.try_set(BLOCK as u32).unwrap();
+        oracle.set_bigram(true);
+        let p = PrefixCachePolicy::new(Arc::clone(&tree), oracle, 1.0);
+        assert_eq!(
+            shares(&p, &ws, &ctx),
+            vec![0.0],
+            "bigram query, unigram tree"
+        );
+        assert_eq!(shares(&policy(tree), &ws, &ctx), vec![1.0]);
+    }
+}

experimental/sgl-router/src/server/app_context.rs

@@ -116,6 +116,8 @@ impl AppContext {
                     circuit_breaker: None,
                     cache_aware: None,
                     sticky: None,
+                    fused: None,
+                    eligibility: None,
                 },
                 discovery: crate::config::DiscoveryBackend::StaticUrls(
                     crate::config::StaticUrlsDiscoveryConfig {

experimental/sgl-router/src/server/routes/models.rs

@@ -54,6 +54,8 @@ mod tests {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         };
         let app = crate::server::app::build_router(std::sync::Arc::new(ctx));
         let res = app

experimental/sgl-router/src/server/routes/tokenize.rs

@@ -121,6 +121,8 @@ mod tests {
                 circuit_breaker: None,
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: crate::config::DiscoveryBackend::StaticUrls(
                 crate::config::StaticUrlsDiscoveryConfig {

experimental/sgl-router/src/tokenizer/mod.rs

@@ -235,6 +235,8 @@ mod tests {
                 circuit_breaker: None,
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: crate::config::DiscoveryBackend::StaticUrls(
                 crate::config::StaticUrlsDiscoveryConfig {

experimental/sgl-router/src/workers/manager.rs

@@ -481,6 +481,8 @@ mod tests {
                 }),
                 cache_aware: None,
                 sticky: None,
+                fused: None,
+                eligibility: None,
             },
             discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
                 urls: vec!["http://test:30000".into()],

experimental/sgl-router/tests/component/discovery/static_urls.rs

@@ -133,6 +133,8 @@ async fn static_urls_pd_role_resolved_end_to_end() {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec![url.clone()],

experimental/sgl-router/tests/component/policies/cache_aware_zmq.rs

@@ -72,6 +72,8 @@ async fn zmq_indexer_routes_to_publishing_worker_e2e() {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: sgl_router::config::DiscoveryBackend::StaticUrls(
             sgl_router::config::StaticUrlsDiscoveryConfig {

experimental/sgl-router/tests/component/policies/fused_score.rs

@@ -0,0 +1,75 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 The SGLang Authors
+// SPDX-License-Identifier: Apache-2.0
+
+//! Two REAL scoring policies whose terms DISAGREE, composed and routed. The
+//! in-crate fusion tests sum `ByIndex` stubs that rank the fleet the SAME way,
+//! so their `select()` half lands on ws[2] whichever term you read; this one's
+//! half discriminates. (Their `scores()` half does catch a dropped term —
+//! verified by mutation, so this file does not claim otherwise.)
+//!
+//! NOT pinned here: how `load_based` scales load — W2's min-max scale-free
+//! defect is unruled, and both candidate curves put the busiest worker at 0.0
+//! and the idlest at 1.0, so every assertion below holds either way.
+
+use sgl_router::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
+use sgl_router::policies::kv_events::{
+    compute_block_hashes, BlockSizeOracle, HashTree, KvWorkerId,
+};
+use sgl_router::policies::load_based::LoadBasedPolicy;
+use sgl_router::policies::scoring::{prefix_cache::PrefixCachePolicy, FusedScorePolicy};
+use sgl_router::policies::{Policy, SelectionContext};
+use sgl_router::workers::Worker;
+use std::sync::Arc;
+
+const BLOCK: usize = 4;
+
+fn worker(id: &str) -> Arc<Worker> {
+    Arc::new(Worker::new(WorkerSpec {
+        id: WorkerId(id.into()),
+        url: id.into(),
+        mode: WorkerMode::Plain,
+        model_ids: vec![ModelId("tiny".into())],
+        bootstrap_port: None,
+    }))
+}
+
+#[test]
+fn the_weight_override_steers_a_two_term_fusion_past_either_term_alone() {
+    let ids: Vec<u32> = (0..(BLOCK as u32 * 4)).collect();
+    let tree = Arc::new(HashTree::new());
+    tree.insert(
+        &KvWorkerId::new("hot".into(), 0),
+        None,
+        &compute_block_hashes(&ids, BLOCK),
+    );
+    let oracle = BlockSizeOracle::new();
+    oracle.try_set(BLOCK as u32).expect("fresh oracle");
+
+    // "hot" holds the whole prompt AND is the busiest: the two terms disagree.
+    let ws = vec![worker("hot"), worker("cold")];
+    let _held: Vec<_> = (0..3).map(|_| ws[0].load_guard()).collect();
+
+    let model = ModelId("tiny".into());
+    let ctx = SelectionContext::new(&model, None).with_request_tokens(Some(&ids));
+    let cache = || PrefixCachePolicy::new(Arc::clone(&tree), Arc::clone(&oracle), 1.0);
+
+    // Vacuity guard: if the terms agreed, no weight could change the answer and
+    // everything below would pass against a composer that read only one of them.
+    assert_eq!(cache().select(&ws, &ctx).unwrap().id, ws[0].id, "cache→hot");
+    assert_eq!(
+        LoadBasedPolicy::new().select(&ws, &ctx).unwrap().id,
+        ws[1].id,
+        "load→cold"
+    );
+
+    // Same two terms, same fleet, same request — only the override differs.
+    for (load_weight, want) in [(0.25, &ws[0]), (4.0, &ws[1])] {
+        let fused = FusedScorePolicy::new(vec![
+            (Arc::new(cache()) as Arc<dyn Policy>, None),
+            (Arc::new(LoadBasedPolicy::new()), Some(load_weight)),
+        ])
+        .expect("both terms are fusable");
+        let got = fused.select(&ws, &ctx).expect("non-empty fleet");
+        assert_eq!(got.id, want.id, "--fuse load_based={load_weight}");
+    }
+}

experimental/sgl-router/tests/component/policies/mod.rs

@@ -4,6 +4,7 @@
 mod zmq_helpers;
 
 mod cache_aware_zmq;
+mod fused_score;
 mod kv_events_hash_parity;
 mod kv_events_tree_concurrent;
 mod kv_events_two_subscribers;

experimental/sgl-router/tests/proxy/cache_aware_input_ids.rs

@@ -12,10 +12,18 @@
 //!   doesn't render tool schemas, so its ids would diverge from the engine).
 //! * A request with multimodal (array) content → `input_ids` omitted (a text
 //!   tokenizer can't represent image content).
+//!
+//! The model id contains `deepseek-v4` so the tokenizer registry auto-attaches
+//! the built-in V4 chat encoder — the engine-equivalent path — without a
+//! template fixture.
 
 use axum::body::Body;
 use axum::http::{Request, StatusCode};
 use serde_json::{json, Value};
+use sgl_router::config::{
+    ActiveLoadConfig, CacheAwareConfig, Config, DiscoveryBackend, ModelConfig, ObservabilityConfig,
+    PolicyKind, ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig,
+};
 use sgl_router::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
 use sgl_router::policies::factory::build_registry;
 use sgl_router::policies::kv_events::{BlockSizeOracle, HashTree};
@@ -28,9 +36,35 @@ use std::sync::Arc;
 use std::time::Duration;
 use tower::ServiceExt;
 
-use crate::common::cache_aware_fixture::{config, MODEL};
 use crate::common::mock_worker::MockWorker;
 
+const MODEL: &str = "deepseek-v4-tiny";
+
+fn config() -> Config {
+    Config {
+        server: ServerConfig {
+            host: "0".into(),
+            port: 0,
+        },
+        observability: ObservabilityConfig::default(),
+        model: ModelConfig {
+            id: MODEL.into(),
+            tokenizer_path: "tests/fixtures/tiny_tokenizer.json".into(),
+            policy: PolicyKind::CacheAwareZmq,
+            circuit_breaker: None,
+            cache_aware: Some(CacheAwareConfig::default()),
+            sticky: None,
+            fused: None,
+            eligibility: None,
+        },
+        discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
+            urls: vec!["http://placeholder:0".into()],
+        }),
+        proxy: ProxyConfig::default(),
+        active_load: ActiveLoadConfig::default(),
+    }
+}
+
 fn build_ctx(url: String) -> Arc<AppContext> {
     let cfg = config();
     let tokenizers = Arc::new(TokenizerRegistry::load_from_config(&cfg).unwrap());

experimental/sgl-router/tests/proxy/chat_routing.rs

@@ -37,6 +37,8 @@ fn config_for(_worker_url: &str) -> Config {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/common/cache_aware_fixture.rs

@@ -29,6 +29,8 @@ pub fn config() -> Config {
             circuit_breaker: None,
             cache_aware: Some(CacheAwareConfig::default()),
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/failover.rs

@@ -41,6 +41,8 @@ async fn failover_when_one_worker_dies() {
             }),
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec![w1.url.clone(), w2.url.clone(), w3.url.clone()],

experimental/sgl-router/tests/proxy/graceful_shutdown.rs

@@ -47,6 +47,8 @@ fn build_ctx_with_worker(worker_url: &str) -> Arc<AppContext> {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/header_forwarding.rs

@@ -34,6 +34,8 @@ async fn forwards_whitelisted_headers_strips_others() {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/pd_bootstrap_injection.rs

@@ -49,6 +49,8 @@ fn config() -> Config {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/pd_pool_isolation.rs

@@ -48,6 +48,8 @@ fn config() -> Config {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/roundrobin_input_ids.rs

@@ -45,6 +45,8 @@ fn config() -> Config {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/sticky_input_ids.rs

@@ -25,7 +25,7 @@ use axum::http::{Request, StatusCode};
 use serde_json::{json, Value};
 use sgl_router::config::{
     ActiveLoadConfig, Config, DiscoveryBackend, ModelConfig, ObservabilityConfig, PolicyKind,
-    ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig, StickyConfig,
+    ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig, StickyConfig, StickyFallbackKind,
 };
 use sgl_router::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
 use sgl_router::policies::factory::build_registry_with_defaults as build_policy_registry;
@@ -60,10 +60,12 @@ fn config() -> Config {
             // mid-test; round-robin fallback for the initial pin of a key.
             sticky: Some(StickyConfig {
                 header_name: HEADER.to_string(),
-                fallback_policy: PolicyKind::RoundRobin,
+                fallback_policy: StickyFallbackKind::RoundRobin,
                 idle_secs: 3600,
                 eviction_interval_secs: 3600,
             }),
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/sticky_routing.rs

@@ -8,7 +8,7 @@
 
 use sgl_router::config::{
     ActiveLoadConfig, Config, DiscoveryBackend, ModelConfig, ObservabilityConfig, PolicyKind,
-    ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig, StickyConfig,
+    ProxyConfig, ServerConfig, StaticUrlsDiscoveryConfig, StickyConfig, StickyFallbackKind,
 };
 use sgl_router::discovery::{ModelId, WorkerId, WorkerMode, WorkerSpec};
 use sgl_router::policies::factory::build_registry_with_defaults as build_policy_registry;
@@ -46,10 +46,12 @@ fn build_sticky_ctx(header_name: &str, worker_urls: &[String]) -> Arc<AppContext
             cache_aware: None,
             sticky: Some(StickyConfig {
                 header_name: header_name.to_string(),
-                fallback_policy: PolicyKind::RoundRobin,
+                fallback_policy: StickyFallbackKind::RoundRobin,
                 idle_secs: 3600,
                 eviction_interval_secs: 3600,
             }),
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

experimental/sgl-router/tests/proxy/timeout.rs

@@ -41,6 +41,8 @@ fn config(_worker_url: &str) -> Config {
             circuit_breaker: None,
             cache_aware: None,
             sticky: None,
+            fused: None,
+            eligibility: None,
         },
         discovery: DiscoveryBackend::StaticUrls(StaticUrlsDiscoveryConfig {
             urls: vec!["http://placeholder:0".into()],

scripts/ci/cuda/ci_install_dependency.sh

@@ -324,6 +324,20 @@ setup_cargo_cache() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
+invalidate_torch_rust_cache() {
+    if [ "${SGLANG_BUILD_RUST_EXTS:-}" = "none" ]; then
+        mark_step_done "${FUNCNAME[0]}"
+        return
+    fi
+
+    # uv's editable build uses a temporary torch path. Rebuild these units
+    # under the lock so Cargo does not reuse that path in a later job.
+    cargo clean --release --manifest-path "${REPO_ROOT}/rust/sglang-radix-tree/Cargo.toml" \
+        -p torch-sys -p sglang-radix-tree
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
 release_cargo_cache_lock() {
     if [ "${CARGO_TARGET_LOCK_HELD:-0}" = "1" ]; then
         flock --unlock 9
@@ -904,6 +918,7 @@ main() {
     install_pytorch_stack
     install_cuda12_deepep_wheel
     setup_cargo_cache
+    invalidate_torch_rust_cache
     install_sglang
     release_cargo_cache_lock
     # Diffusion B200 CI imports torch inside install_sglang_kernel after removing

test/registered/unit/tools/test_cuda_ci_install_dependency.py

@@ -0,0 +1,46 @@
+import importlib.util
+import re
+import unittest
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[4]
+CI_REGISTER_PATH = REPO_ROOT / "python" / "sglang" / "test" / "ci" / "ci_register.py"
+INSTALL_SCRIPT = REPO_ROOT / "scripts" / "ci" / "cuda" / "ci_install_dependency.sh"
+
+
+def _load_module(name, path):
+    spec = importlib.util.spec_from_file_location(name, path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+register_cpu_ci = _load_module("ci_register", CI_REGISTER_PATH).register_cpu_ci
+register_cpu_ci(est_time=0, suite="base-a-test-cpu")
+
+
+class TestCudaCiInstallDependencyTorchCache(unittest.TestCase):
+    def test_rebuilds_torch_extensions_before_editable_install(self):
+        script = INSTALL_SCRIPT.read_text()
+
+        self.assertRegex(
+            script,
+            re.compile(
+                r"setup_cargo_cache\s*\n"
+                r"\s*invalidate_torch_rust_cache\s*\n"
+                r"\s*install_sglang"
+            ),
+        )
+        self.assertIn('"${SGLANG_BUILD_RUST_EXTS:-}" = "none"', script)
+        self.assertRegex(
+            script,
+            re.compile(
+                r"cargo clean --release --manifest-path "
+                r"\"\$\{REPO_ROOT\}/rust/sglang-radix-tree/Cargo\.toml\"\s*\\\n"
+                r"\s*-p torch-sys -p sglang-radix-tree"
+            ),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()