agent_contracts: verifiable inference settlement — model-gated contracts

[grapheneaffiliate]

Add the first contract that uses the ternary VM as a decision model rather than an adding machine. Every v0.1.0 contract reaches its decision by reading a trusted flag from the witness (conditional_payment fires on input[56]==1, multisig counts flag bytes, time_locked compares a supplied clock); the VM is used only for carry/borrow arithmetic. risk_gated_transfer instead computes the gate as a neural-network forward pass on the verifier path, so a machine-learning inference becomes a bit-exact, re-executable artifact that both releases money and resolves disputes — no oracle, no human underwriter.

This works only because TernaryNetwork::forward is bit-exact on every conformant integer host (ARCHITECTURE.md s0.8): a model's output is a deterministic function of (weights_hash, features), so any participant can re-derive the verdict. Floating-point inference cannot do this.

• risk_gated_transfer (agent_contracts/src/inference.rs): transfer gated by credit_risk_model_v1, a 4-layer integer-only ternary MLP computing APPROVE = (income + collateral - debt >= 500) AND (risk_flags <= 10). The AND is a real ReLU(b1+b2-1) non-linearity. program_hash commits to the model's weights_hash, so the model is part of the contract identity.
• inference_agent example: happy path (model APPROVEs, loan settles) + dispute path (model DENIEs a high-risk applicant, malicious executor claims it disbursed, judge re-executes the model -> SlashExecutor).
• 10 new tests (264 workspace total), incl. a feature-grid sweep proving the network is bit-exact vs a plain-Rust ground truth on both decision boundaries. Strict clippy + rustfmt clean.
• docs/use-cases/verifiable-inference-settlement.md: the capability, the build, honest limits, and a portfolio of further model-gated use cases (parametric insurance, verifiable inference markets, oracle-free KYC, deterministic auctions, per-inference model licensing, confidential model-gated payments). INDEX + CHANGELOG updated in the same commit.

Claude-Session: https://claude.ai/code/session_017mNfJDn9fro84YvM3WnCks

What & why

CI gates (all must pass)

• cargo build --workspace --release
• cargo test --workspace --release
• cargo fmt --all -- --check
• cargo clippy --workspace --exclude psl-rust-runner --all-targets -- -D warnings
• ruff check . (if Python touched)
• tools/ci/check_legacy_isolation.sh

Repo invariants

• No new unwrap()/expect() on production paths outside the audited categories (docs/UNWRAP_AUDIT.md updated if added)
• No floating point on the verifier path (fp in tooling/diagnostics is flagged explicitly below)
• No new Lean sorrys in load-bearing theorems
• docs/INDEX.md updated in this PR if any Markdown doc was added, moved, or removed
• Frozen code untouched (legacy/ per ADR-0001, trace-hash v1 per ADR-0008) — or the PR explains why

Tests