[Router] 2381 unified interface for semantic_similarity and model classifier

[SlawomirNowaczyk]

Implements the semantic_similarity classifier for the Lemonade Router and unifies it with the generic classifier type: both now produce a label → score map and are addressed by label + min_score/max_score band. reference_phrases becomes a concept → phrases map, so one classifier can score many concepts in a single embedding pass.

Rationale for the breaking change is here: https://github.com/lemonade-sdk/lemonade/blob/router/2381-semantic_similarity_classifier-unified/semantic_classifier_unification_rationale.md

What changed

• Unified scoring contract. semantic_similarity fills Score::labels with one entry per concept (max cosine over that concept's reference phrases), read back via score_of(label) exactly like classifier. The condition layer (make_classifier_band_condition, leaf-factory label/default_label resolution) has no semantic-vs-model branch.
• Multi-concept support. reference_phrases is now an object mapping each concept name (the classifier's output label) to its exemplar phrases. Input is embedded once and scored against every concept; reference embeddings are computed once and cached (mutex-guarded, since classifiers are shared across concurrent requests).
• Removed the magic empty key. The old Score.labels[""] special case is gone. Score::primary() is now strict: it returns the lone entry of a genuinely label-less classifier and 0.0 otherwise, so a condition that omits label can never silently match an arbitrary label of a multi-label score.
• Shared parsing. make_classifier factors out parse_labels / parse_default_label; the only type-specific step is parse_reference_phrases, which derives labels from the concept keys. semantic_similarity rejects an explicit labels field (concept names are authoritative).
• Schema. route_policy.schema.json updates reference_phrases to a non-empty object of concept → non-empty string[], and forbids labels on semantic_similarity.

Frozen-semantics / compatibility

• reference_phrases shape changes from string[] to {concept: string[]} — a breaking change to the (unreleased) routing schema, intentional and the whole point of the unification.
• Score::primary() redefined (lone-entry-or-0.0); the ""-key convention is no longer special anywhere in the engine. Worth a row in the frozen-semantics table in README.md.

Validation

• C++ unit tests: test_routing_policy_semantic.cpp (multi-concept scoring, max-cosine, caching, on_error, failure paths) plus updates to test_routing_policy_registry.cpp, test_routing_policy_contract.cpp, test_routing_policy_evaluator.cpp, and the l2_semantic.json fixture / fake_classifier_services.h.
• Run: ctest --output-on-failure -R "RoutingPolicy(Contract|Evaluator|Registry|Semantic)Test".

Closes

Closes #2381.

[eddierichter-amd]

Looks great. Two small changes requested.

[ramkrishna2910]

Decision: we're consolidating on this unified approach for #2381 over the single-concept variant in #2482. Reasoning is timing — the routing schema is still unreleased, so collapsing semantic_similarity onto the generic classifier contract (and dropping the magic empty key) is free now and a breaking, corpus-and-migration cost after release. The rationale doc captures it well, and the multi-concept capability is a real bonus for no ongoing complexity.

Two things to land before merge (both mechanical, per Eddie's review):

1. CI — add RoutingPolicySemanticTest to the ctest filter (RoutingPolicy(Contract|Evaluator|Registry|Semantic)Test) in both spots in cpp_server_build_test_release.yml.
2. schema-lock — route_policy.schema.json changed but schema-lock.json wasn't refreshed. Regenerate it in this PR (python test/test_schema_lock.py --update) so the lock guard passes. The schema is still released: false, so this is a clean refresh, not a back-compat break.

Heads-up on sequencing: #2380 (deterministic conditions) also touches route_policy.schema.json + schema-lock.json (tightening keywords_any/keywords_all to minItems: 1). I'll rebase #2380 onto this once it merges and regenerate the lock there, so we don't collide on the hash.