RFC: Token Budget Observability — live utilization metrics per process

[zatchbell1311-wq]

Hi @ingim and @shsym — now that #382 is merged, I'd like to take on something more substantial.

Problem
Pie manages token budgets internally via the context manager and admission gating, but operators currently have no visibility into live utilization per process. When a daemon deployment is under load, there's no way to know how close individual processes are to their budget limits without digging into logs.

Proposed contribution
Add a lightweight /metrics or /status endpoint to pie-server that exposes:

1. Current token budget utilization per active process.
2. Admission gate status (accepted/rejected counts).
3. KV cache pressure indicators.

Why I'm interested
My own research (DSPM) is directly about token budget management in LLM deployments, so I have deep context on what operators actually need to observe. Happy to start with a design discussion before writing any code.

Is this something the project would welcome, or is it already on the roadmap?