feat: core engine — TokenTracker, ModelPricing, BudgetCircuitBreaker

[hermanngeorge15]

Summary

PR 2 of budget-breaker v0.0.1 Phase 1.

• TokenTracker: AtomicLong pair (prompt + completion), thread-safe, limit checks. Exposes model, hardLimitTokens, softLimitTokens from budget config.
• ModelPricing: Hardcoded defaults for Claude/GPT/Gemini/Ollama + custom overrides via Map. Companion object convenience method.
• BudgetReport: Data class for post-run summary (tokens, cost, duration, breach counts).
• BudgetScope: trackCall() DSL with soft/hard limit checks. Uses AtomicInteger for breach count, AtomicBoolean for soft limit fired flag (no var).
• BudgetCircuitBreaker: Coroutine supervisor with withBudget() API. Callback + SharedFlow for soft limit events. ConcurrentHashMap for reports.

Test plan

• 7 TokenTracker tests: separate tracking, accumulation, soft/hard limit detection, percent used, thread safety, zero start
• 6 ModelPricing tests: Claude Sonnet/Opus, combined cost, unknown model, custom overrides, override precedence
• 2 BudgetReport tests: all fields, totalTokens consistency
• 8 BudgetCircuitBreaker tests: normal completion, hard limit, soft limit callback, soft limit flow event, report generation, report after exception, cost estimation, concurrent agents
• 23 total tests, 0 failures — clean build from scratch

Constraints verified

• No var anywhere — uses AtomicInteger, AtomicBoolean, AtomicLong
• No !! without justifying comment (one in test, explained)
• No Thread.sleep — pure coroutines
• All source in budget-breaker/src/main/kotlin/dev/unityinflow/budget/
• All tests in budget-breaker/src/test/kotlin/dev/unityinflow/budget/