[inference-out] upstream: resolved-route enrichment middleware with a cost-estimator callback #75

agent/tinyagents/routes.rs (194) is mostly ResolvedRouteMiddleware. It emits generation enrichment, records the resolved route, applies :free pricing and calls pricing::lookup. All of that is generic apart from the cost estimator and the run-context slot.

Add, in tinyagents-harness/src/middleware/library/workload.rs:

• ResolvedRouteMiddleware<C: CostEstimator>;
• trait CostEstimator { fn estimate(&self, model, usage) -> Option<f64> };
• a route sink callback.

The host passes agent::cost::estimate_call_cost_usd and its OpenHumanRunContext.resolved_route slot. Inline turn_required_capabilities / route_fallback_policy at their two call sites (turn_exec.rs:556, turn_plan.rs:54).

Host row unblocked (H9): agent/tinyagents/routes.rs.

Blocked by

None.

Definition of done

The listed APIs exist upstream with tests, are released/tagged, and the matching host-adoption leaf can pin them. Gates: the repo's own cargo test, cargo clippy -D warnings and CI.

Agent brief

• Files to consider: the rows above; the epic's Evaluation section has each row's reason.
• Research: the gaps listed above. Read the cited upstream symbols before writing code.
• Workflow: one implementing-an-epic step. Follow OpenHuman's .claude/skills/migrate-to-submodule: upstream PR to the canonical tinyhumansai/<repo>, then tag/release; the OpenHuman gitlink bump and host deletion happen in the matching host-adoption leaf. tinyinference must not depend on tinyagents or OpenHuman.
• Subagent: tdd-implementer.

Parent: tinyhumansai/openhuman#8092