feature: declare cross-encoder rerankers in model_catalog

Audited upstream baseline (2026-07-15)

Current main has no general cross-encoder reranker runtime, /v1/rerank endpoint, or SR_CROSS_ENCODER_MODEL_* configuration. PR #2236 proposes that foundation but remains open and unmerged; its branch-level environment variables are not a released compatibility contract.

This issue defines the production configuration and lifecycle contract that should accompany or follow the base runtime, without describing the open PR as current behavior.

Scope

• Add a typed cross-encoder/reranker declaration to the canonical model catalog.
• Validate model identity, architecture/runtime compatibility, artifact source, limits, and credentials.
• Load/unload it through the shared model-runtime lifecycle.
• Expose capability and readiness through model inventory/diagnostics.
• Document migration only if an interim surface actually ships.

Boundaries

• Do not standardize open-PR environment variables as the long-term public API.
• Do not couple the model declaration to one routing signal or endpoint.
• Do not claim a model is ready before artifact/runtime validation succeeds.
• Keep secrets out of config projections, diagnostics, and Replay.

Acceptance criteria

• Rerankers have a canonical, round-trippable config declaration.
• Unknown or incompatible declarations fail with actionable diagnostics.
• Runtime inventory reports configured, loading, ready, and failed states.
• Config/reference/docs contract tests cover the new surface.
• CPU/model-free CI validates schema and lifecycle behavior; model receipts validate real inference separately.

Relevant surfaces

src/semantic-router/pkg/config, src/semantic-router/pkg/modelruntime, src/semantic-router/pkg/modelinventory, capability inventory (#2358), docs, and tests.

Related: PR #2236, architecture #2247, routing integration #2251.