feat(eval): suite-level total cost budget guardrail (core runtime)

[christso]

Canonical Plan

This issue body is the source of truth for implementation.

Objective

We add a run-wide budget guardrail for matrix/trial evaluations to prevent unbounded spend.

Why This Matters

• Matrix + trials scale cost multiplicatively (targets × tests × trials).
• A run can be interrupted late with unexpected cost if we only bound per trial.
• CI and scheduled benchmark runs need a deterministic total-cost ceiling.

Why This Is Core (Not External)

• External tooling can only detect overspend after results are written; it cannot reliably stop provider calls in-flight.
• Budget enforcement is an execution-control primitive, so it belongs in the orchestrator where dispatch decisions are made.
• We need consistent behavior across providers/targets during the run, not post-hoc filtering.

Architecture Boundary

Core runtime change (execution-control primitive).

Deliverable Location

• Core implementation: packages/core/src/evaluation/
• Expected touchpoints:
  • packages/core/src/evaluation/types.ts
  • packages/core/src/evaluation/loaders/config-loader.ts
  • packages/core/src/evaluation/orchestrator.ts
• Tests: packages/core/test/evaluation/
• Docs updates: apps/web/src/content/docs/

Design Latitude

We choose schema/CLI shape and budget-breach behavior details, as long as behavior is non-breaking and explicit.

Acceptance Signals

• We support an optional suite-level total budget cap.
• We enforce budget accounting across targets/tests/trials.
• We surface budget-triggered early termination in outputs/status.

Non-Goals

• We do not replace existing per-trial cost_limit_usd behavior.
• We avoid provider-specific budgeting semantics in core unless unavoidable.