feat: add bounded native Qwen3 CPU reranking

AGENTS.md

@@ -28,6 +28,8 @@ share `/data/target` across agent lanes. See
 the precise threat model. Ordinary compilation is not a hardware-test permit.
 Every Cargo invocation, including lockfile generation and formatting, uses the
 runner; an ambient compiler wrapper must not escape the isolated lane.
+Host CPU compile admission is separate: where required, the approved admission
+entrypoint wraps this runner rather than relying on ambient PATH shims.
 
 ## Key patterns
 
@@ -62,6 +64,7 @@ runner; an ambient compiler wrapper must not escape the isolated lane.
 | STT pipeline | `crates/ekphrasis/` |
 | Tokenizer | `crates/tokenize/` |
 | Native text request pipeline | `crates/text/` |
+| Shared bounded template rendering | `crates/templates/` |
 | Shared synthetic GGUF test data | `crates/test-fixtures/` (dev-only) |
 
 The private planning corpus governs future scope and sequencing. This repository's

ARCHITECTURE.md

@@ -61,6 +61,18 @@ semantically respects that boundary.
   Stella API and tensor/encoder graph; direct consumers disable default
   features to exclude that accelerator-capable graph. Consumers still use the
   unchanged `core::EmbeddingModel` contract.
+- `rerank` consumes the same CPU decoder/tokenizer graph and `templates` for
+  native Qwen3 pair scoring. Its default `modernbert` feature preserves the
+  existing encoder implementation; disabling default features removes that
+  accelerator-capable graph without changing the `Reranker` contract.
+- `templates` owns bounded, capability-free artifact-template rendering for
+  `text` and `rerank`. It has no model, GGUF, tokenizer or device dependency;
+  pipelines retain artifact binding, typed message roles and token policy.
+- `tokenize` owns exact vocabulary/special-ID verification and refusal of
+  configured tokenizer padding or truncation. Native text, embedding and
+  reranking setup invoke that guard; ordinary tokenizer consumers retain their
+  configured behavior. Disabling automatic special tokens alone does not
+  disable padding or truncation.
 - `taxis` depends locally on `hipcore`.
 - `kernels/gpu` enables the local `hipcore` and `taxis` dependencies and GPU
   launcher modules, including their nested parity references. Standalone
@@ -186,6 +198,26 @@ instruction, and advertised dimensions do not imply Matryoshka qualification.
 Synthetic family and pipeline tests do not establish deployed-artifact parity,
 retrieval quality, reindex authority, serving or GPU qualification.
 
+## Native reranking ownership
+
+`decoders::Qwen3RankWeights` admits a distinct rank profile over the shared
+private Qwen3 body. Rank pooling, exact `[yes, no]` labels and a two-row
+`cls.output.weight` are required; embedding admission still refuses extra
+heads. The decoder returns raw terminal-token classifier logits after final
+RMS normalization. `rerank` owns their signed `yes - no` reduction and returns
+one relevance logit per input index, not probabilities or sorted results.
+
+The pipeline renders its verified artifact's embedded template using typed
+system/query/document messages and an explicit setup instruction. It encodes
+the whole render with automatic special tokens disabled. Separate byte,
+token and batch bounds reject oversized inputs without truncation. The shared
+`templates` owner retains strict undefined values, fuel, recursion and output
+bounds, without external or named template resolution. Public batch structs
+are revalidated at the prediction boundary; failure publishes no partial map.
+These controls do not bound all tokenizer/template intermediates or total
+process memory. Synthetic execution does not resolve exact-artifact conversion
+provenance, template parity, retrieval quality, deployment or hardware gates.
+
 ## cfg flags
 
 - `logismos_no_gpu_kernels` - build path without compiled HIP kernels. Implemented GPU operations

Cargo.toml

@@ -95,6 +95,7 @@ tokenize = { path = "crates/tokenize" }
 cache = { path = "crates/cache" }
 decode = { path = "crates/decode" }
 text = { path = "crates/text" }
+templates = { path = "crates/templates" }
 core = { path = "crates/core" }
 transformers = { path = "crates/transformers" }
 encoders = { path = "crates/encoders" }

README.md

@@ -8,7 +8,7 @@ targeting AMD gfx1100, with owned HIP/WMMA kernels and progressively owned execu
 **Status:** HIP primitives, Stella CPU golden-fixture parity, action-free placement,
 process-local admission/residency coordination, and bounded instruction emulation exist.
 GGUF inspection, digest-bound mixed-weight CPU projections, bounded hybrid CPU
-text generation and native Qwen3 CPU embeddings are foundations, not serving or
+text generation and native Qwen3 CPU embeddings/reranking are foundations, not serving or
 hardware qualification. The W7900
 is available; the RX 7900 XTX is a
 planned second device and requires its own qualification. The experimental below-HIP
@@ -101,6 +101,21 @@ establish exact deployed-artifact parity, retrieval quality or reindex authority
 Direct `embed` consumers disable default features for the HIP-free native path;
 the default `stella` feature preserves the existing Stella API and dependencies.
 
+[`rerank`](crates/rerank/src/lib.rs) implements native Qwen3 CPU pair scoring
+through the existing `Reranker` contract. Its checked rank profile shares the
+causal decoder body without weakening embedding admission. The verified
+artifact supplies its template; setup supplies the instruction and independent
+byte/token/batch limits. Each input index receives one raw `yes - no` relevance
+logit, not a probability. Oversized requests fail without truncation.
+Direct native consumers disable default features; the default `modernbert`
+feature preserves the existing encoder implementation. Exact converted-model
+provenance, template/tokenizer parity and retrieval quality remain unqualified.
+
+[`templates`](crates/templates/src/lib.rs) owns bounded template rendering for
+text generation and reranking. It permits no host callbacks, loader or named
+template registration; output, recursion and fuel limits are operational
+controls, not a total-memory sandbox.
+
 [`contracts/runtime-scope.toml`](contracts/runtime-scope.toml) records this product boundary.
 Bounded adaptation remains absent unless a named consumer contract supplies an output owner,
 retention and revocation policy, and rollback. The repository guard validates those declared

crates/decoders/src/error.rs

@@ -402,8 +402,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// A Qwen3 embedding metadata value is missing, mistyped, or unsupported.
-    #[snafu(display("qwen3 embedding metadata `{key}` violates {rule}"))]
+    /// A Qwen3 profile metadata value is missing, mistyped, or unsupported.
+    #[snafu(display("qwen3 profile metadata `{key}` violates {rule}"))]
     Qwen3Metadata {
         /// Exact GGUF key responsible for the refusal.
         key: &'static str,
@@ -414,8 +414,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// A Qwen3 embedding tensor inventory entry is missing or malformed.
-    #[snafu(display("qwen3 embedding tensor `{name}` violates {rule}"))]
+    /// A Qwen3 profile tensor inventory entry is missing or malformed.
+    #[snafu(display("qwen3 profile tensor `{name}` violates {rule}"))]
     Qwen3Tensor {
         /// Exact GGUF tensor name responsible for the refusal.
         name: String,
@@ -426,8 +426,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// A bounded Qwen3 embedding execution request is invalid.
-    #[snafu(display("qwen3 embedding execution request {requested} violates {rule}"))]
+    /// A bounded Qwen3 profile execution request is invalid.
+    #[snafu(display("qwen3 profile execution request {requested} violates {rule}"))]
     Qwen3Execution {
         /// Requested token count or other bounded execution value.
         requested: usize,
@@ -438,8 +438,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// Qwen3 embedding execution could not reserve one named local buffer.
-    #[snafu(display("qwen3 embedding could not reserve {length} values for {target}: {source}"))]
+    /// Qwen3 profile execution could not reserve one named local buffer.
+    #[snafu(display("qwen3 profile could not reserve {length} values for {target}: {source}"))]
     Qwen3Allocation {
         /// Named local buffer.
         target: &'static str,
@@ -452,10 +452,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// Qwen3 embedding execution produced a non-finite scalar.
-    #[snafu(display(
-        "qwen3 embedding arithmetic became non-finite during {stage} at index {index}"
-    ))]
+    /// Qwen3 profile execution produced a non-finite scalar.
+    #[snafu(display("qwen3 profile arithmetic became non-finite during {stage} at index {index}"))]
     Qwen3Arithmetic {
         /// Mathematical stage.
         stage: &'static str,
@@ -466,8 +464,8 @@ pub enum Error {
         location: snafu::Location,
     },
 
-    /// A checked shared CPU operation rejected Qwen3 embedding execution.
-    #[snafu(display("qwen3 embedding CPU operation failed: {source}"))]
+    /// A checked shared CPU operation rejected Qwen3 profile execution.
+    #[snafu(display("qwen3 profile CPU operation failed: {source}"))]
     Qwen3Cpu {
         /// Checked CPU operation failure.
         source: kernels::Error,

crates/decoders/src/lib.rs

@@ -28,9 +28,11 @@ pub mod qwen35_execution;
 pub mod qwen35_recurrent;
 mod qwen35_requirements;
 pub mod qwen35_weights;
+pub mod qwen3_rank;
 
 pub use crate::error::{Error, Result};
 pub use crate::qwen3::{Qwen3Execution, Qwen3Weights};
+pub use crate::qwen3_rank::{Qwen3RankExecution, Qwen3RankWeights};
 pub use crate::qwen35::Qwen35StructuralProfile;
 pub use crate::qwen35_execution::{Qwen35Execution, Qwen35ExecutionPlan, Qwen35LogitSelection};
 pub use crate::qwen35_recurrent::Qwen35RecurrentExecution;