feat: add hummbl_model_routing package — free inference, OpenRouter, cost tracking

[hummbl-dev]

Summary

Extract model routing, free inference, and cost tracking from founder-mode as a standalone stdlib-only package. This is the canonical home for these modules; founder-mode will stub to re-export from here (PR #1969 in founder-mode).

Modules

• free_inference: Routes chat requests to best available free LLM endpoint (Ollama, NVIDIA NIM, OpenRouter free, paid fallback). Circuit breaker per provider, retry with exponential backoff, task-class-aware model selection.
• openrouter_client: OpenRouter API client (stdlib urllib, OpenAI-compatible, governed). CostGovernor for pre-call cost estimation.
• cost_tracker: SQLite-backed API cost tracking with budget enforcement, health checks, HMAC auth for sensitive operations. AlertService import deferred to TYPE_CHECKING (optional dependency on founder_mode).

Package

• Name: hummbl-model-routing v0.1.0
• Python: >=3.11, zero third-party runtime dependencies
• Optional extras: [test] (pytest), [governance] (hummbl-governance>=1.3.0)

Test plan

• 26 unit tests pass (routing order, circuit breaker, cost tracker, mocked HTTP)
• Package installs cleanly via pip install -e .
• All public imports work: chat, TaskClass, InferenceResponse, CostTracker, BudgetStatus, UsageRecord, OpenRouterClient, CostGovernor, ChatResponse, OpenRouterNetworkError
• No hard dependency on founder_mode (AlertService is TYPE_CHECKING only)
• CI green on push