Design & implement cost/token budgets

[AlexChesser]

Summary

Design per-step and per-run token limits, cost caps, and budget-aware execution. This is NOT IN SPEC and requires spec authoring.

Parent issue: #105 — Tier 3, Priority #10

Why

In production, runaway costs from long context or retry loops are a real operational risk. A budget: field per step or per pipeline would enable cost-aware model selection and prevent blowouts. This is especially critical once iterative loops (#116) and retry (#108) are implemented.

Design Decisions Needed

• Budget unit — tokens? Dollars? Both?
• Granularity — per-step, per-pipeline, per-run, or all three?
• YAML syntax — budget: block with max_tokens:, max_cost:?
• Behavior when budget exceeded — abort? switch to cheaper model? warn and continue?
• How to track token usage across steps (runner must report usage)
• Interaction with model cascading (multi-provider routing) — auto-downgrade when budget is low?
• How budget state is exposed in template variables (e.g. {{ budget.remaining_tokens }})

Spec Work Required

New spec section needed for budget/cost control.

Acceptance Criteria

• Spec section authored
• Budget limits can be declared at pipeline and/or step level
• Token/cost tracking per step
• Configurable behavior when budget is exceeded
• Budget state visible in turn log