Pricing & USD-denominated budgets

A Agentic SpendGuard tenant can spend a single USD budget across multiple LLM providers (OpenAI, Anthropic, Bedrock, Azure OpenAI, Gemini). The flow:

1. Pricing table — pricing_table in canonical DB stores (provider, model, token_kind) → $/1M tokens. Loaded from deploy/demo/init/pricing/seed.yaml at compose-up time.
2. Pricing freeze — bundle build picks the latest pricing_version and embeds the snapshot hash in contract_bundle.metadata.json. Sidecar reads it at startup; production deploys validate via cosign.
3. Adapter conversion — adapter computes µUSD per call: (input_tokens × $/1M_input + output_tokens × $/1M_output) × 10^6.
4. Sidecar reserve — adapter submits BudgetClaim(amount_atomic=N µUSD). Sidecar reserves under the usd_micros unit (scale=6).
5. Sidecar commit — adapter sees the actual provider response, recomputes µUSD, sidecar commits via CommitEstimated.

The multi_provider_usd demo mode shows this end-to-end with real OpenAI + Anthropic in the same session against a shared USD budget.

See pricing.py for the helper that does the conversion.