feat: harden multimodal and image-routing signal robustness

Summary

Harden multimodal and image-routing signal behavior so image-modality routing can be evaluated, calibrated, discovered, and tested reliably across router APIs and embedding backends.

Parent roadmap: #2287 Related signal roadmap: #2330, #2335, #2333

Motivation

Image-routing is becoming a broader signal surface: request-time routing, HTTP classify/similarity endpoints, model discovery, threshold packs, and CI coverage all depend on multimodal embeddings behaving consistently. Several open issues show that the current surface has correctness and operational gaps:

• text-bearing images can over-fire image embedding meta-rules without a stage-2 text-modality check;
• image-routing thresholds need recalibration after encoder correctness fixes;
• ONNX and Candle image preprocessing may diverge;
• HTTP APIs do not yet consistently support image inputs across classify/batch/similarity;
• clients cannot discover whether multimodal embedding support is loaded;
• model-gated multimodal tests do not run in CI.

This parent issue groups that work so it is visible in the H2 roadmap instead of remaining scattered.

Child issues

• #2057: Text-bearing image content needs a stage-2 text-modality check.
• #2165: Recalibrate image-routing thresholds after encoder correctness fixes.
• #2166: Resolve ONNX/Candle image preprocessing parity.
• #2317: Extend HTTP API image input to classify/batch and similarity endpoints.
• #2318: Expose multimodal model capability discovery.
• #2319: Run or explicitly gate multimodal model tests in CI.

Scope

This track should define and validate:

• stage-2 OCR/text-modality checks for text-bearing images;
• measured threshold calibration with committed provenance;
• binding parity or fail-closed behavior for unsupported multimodal paths;
• API capability discovery for multimodal models;
• endpoint behavior for image classification and similarity;
• CI or manual receipt policy for model-dependent multimodal tests;
• replay/diagnostic output for multimodal signal decisions.

Non-goals

• Do not make multimodal support mandatory for all deployments.
• Do not silently mix incompatible embedding spaces.
• Do not treat image embedding similarity as sufficient for rendered-text sensitivity.
• Do not broaden the API surface without validation and capability discovery.

Acceptance criteria

• Multimodal/image-routing child issues have explicit owners or sequencing.
• Image-routing thresholds have measurement provenance.
• Text-bearing image routing has a stage-2 design or implementation path.
• Multimodal support is discoverable by clients.
• CI or release validation covers at least one meaningful multimodal encode path, or the manual receipt policy is documented.
• The H2 roadmap links this parent issue as the multimodal signal robustness track.

Validation

make agent-report ENV=cpu CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/apiserver candle-binding onnx-binding config website/docs bench .github"
make agent-lint CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/apiserver candle-binding onnx-binding config website/docs bench .github"
make agent-ci-gate CHANGED_FILES="src/semantic-router/pkg/classification src/semantic-router/pkg/config src/semantic-router/pkg/apiserver candle-binding onnx-binding config website/docs bench .github"