feat: expand signal extractors and improve signal model coverage

Summary

Expand signal extractors and improve signal model coverage while keeping signals registered, observable, evaluated, and versioned through shared contracts.

Parent roadmap: #2287 Related: #2330, #2333, #2347, #2341 Child and related issues: #2360, #2382, #1485, #2154, #2247, #2250, #2251, #2252, #2301

Motivation

The router already has many signal families in config and runtime code: keyword, embedding, domain/category, fact_check, user_feedback, reask, preference, language, context, structure, complexity, modality, authz, jailbreak, pii, kb, conversation, event, and projection. H2 adds two requirements:

1. make the signal surface extensible and observable through #2330;
2. improve signal quality through better models, extractors, calibration, and lifecycle discipline.

Scope

This issue coordinates signal/model quality work across:

• domain classifier accuracy (#1485);
• complexity signal backend improvements (#2154);
• cross-encoder reranking as signal/model support (#2247, #2250, #2251, #2252);
• mmBERT model discovery (#2301);
• multimodal/image routing robustness (#2347);
• signal model lifecycle, datasets, evals, and release provenance (#2360);
• next-generation router base-model and model-family research for rerankers, embeddings, and classifiers under SOTA quality, long-context, high-concurrency, and low-latency constraints (#2382);
• confidence/threshold calibration (#2341).

Non-goals

• Do not accept new signal extractors without registration, observation output, and evaluation hooks.
• Do not silently update default signal model behavior without calibration/provenance.
• Do not mix recipe-local signal references with global custom-signal runtime configuration.

Acceptance criteria

• New signal extractors are registered through the unified signal registry/observation model.
• Signal model accuracy changes include reproducible eval evidence and calibration notes.
• Runtime diagnostics expose matched signal names, confidence/value where available, and extractor errors/fallbacks.
• Signal improvements can feed decision-level eval and dashboard analytics.

Validation

make agent-report ENV=cpu CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/decision src/semantic-router/pkg/extproc src/semantic-router/pkg/modelruntime src/training candle-binding nlp-binding ml-binding website/docs"
make agent-lint CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/decision src/semantic-router/pkg/extproc src/semantic-router/pkg/modelruntime src/training candle-binding nlp-binding ml-binding website/docs"
make agent-ci-gate CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/decision src/semantic-router/pkg/extproc src/semantic-router/pkg/modelruntime src/training candle-binding nlp-binding ml-binding website/docs"