Token & cost estimation adapters

[Cryptoteep]

Context

RunResult reports durationMs but nothing about tokens or cost. For eval suites run in CI, you usually want to know:

• how many tokens were sent/received,
• estimated cost per case and per suite,
• totals across a run.

Proposal

Introduce a token-counting adapter that runPrompt/evalPrompt can call optionally:

import { tiktokenCounter } from 'promptkit/tokens'

const result = await runPrompt(
  summarize,
  { text: '…' },
  { call, tokens: tiktokenCounter('gpt-4o-mini') },
)
// result.tokens: { prompt: 42, completion: 18, total: 60 }
// result.cost: { prompt: 0.000063, completion: 0.000027, total: 0.00009 }

Implementation notes

• Define a TokenCounter interface: (text: string, role: 'prompt' | 'completion') => number.
• Define a Pricing map per model, loaded from a JSON file so it can be updated without a release.
• Ship a tiktoken adapter behind a subpath export promptkit/tokens (tiktoken-wasm is heavy; keep it optional / peer dep).
• Extend RunResult and EvalResult with optional tokens and cost fields.
• Add a --cost flag to promptkit eval that prints a cost summary.

Acceptance

• TokenCounter interface + tiktokenCounter adapter.
• Pricing JSON + cost computation.
• RunResult.tokens / RunResult.cost populated when a counter is supplied.
• CLI --cost summary.
• Tests + docs.

This is a larger issue — feel free to break it into multiple PRs (interface first, then tiktoken adapter, then pricing).

/label "help wanted" "enhancement"