Gateway compatibility: add embeddings, image, audio, and batch surfaces

[tcballard]

Why

Enterprise clients use more than chat, but each modality has distinct capability, storage, size, streaming, and cost risks.

Scope

• Define normalized capability contracts for embeddings, image generation/input, audio, and batches.
• Implement each surface incrementally behind explicit provider support.
• Add modality-specific request, response, concurrency, timeout, and cost bounds.
• Extend eligibility and allowlists to prevent unsupported routing.
• Define batch persistence, cancellation, expiry, and reconciliation semantics before enabling it.

Acceptance criteria

• No modality is advertised until its provider adapter and parity fixtures pass.
• Unsupported modalities and parameters fail before provider delivery.
• Binary and batch payloads have strict size and retention limits.
• Usage and cost accounting is modality-correct.
• The implementation can ship surfaces independently rather than requiring all four at once.

Parent: #146

Invariant

The scored model decision remains offline, deterministic, keyless, and explainable.