Track per-run token usage and cost in meta.json

[RandomOscillations]

Summary

• Thread actual token counts from API responses through provider → facade → reasoning → engine → meta.json
• Each provider's simple_call_async now returns (dict, TokenUsage) with real token counts from the API response
• Two-pass reasoning captures tokens from both pivotal (Pass 1) and routine (Pass 2) calls, engine accumulates across chunks and computes estimated USD cost via pricing.py

Changes

• providers/base.py — TokenUsage dataclass, updated abstract signature
• providers/openai.py — Extract usage from Responses API (input_tokens/output_tokens) and Chat Completions API (prompt_tokens/completion_tokens)
• providers/claude.py — Extract usage from response.usage
• llm.py — Pass-through tuple return from simple_call_async
• models/simulation.py — Token fields on ReasoningResponse
• reasoning.py — BatchTokenUsage dataclass, capture tokens in two-pass flow, accumulate in batch_reason_agents
• engine.py — Running totals, _compute_cost() using pricing.py, write cost block to meta.json

meta.json output

"cost": {
  "pivotal_input_tokens": 1234567,
  "pivotal_output_tokens": 456789,
  "routine_input_tokens": 234567,
  "routine_output_tokens": 89012,
  "total_input_tokens": 1469134,
  "total_output_tokens": 545801,
  "pivotal_model": "gpt-5",
  "routine_model": "gpt-5-mini",
  "estimated_usd": 5.1234
}

Test plan

• All 618 existing tests pass with updated return types
• 6 new provider token extraction tests (OpenAI Responses, Chat Completions, Claude, null usage)
• 3 new engine tests (chunk accumulation, meta.json cost output, unknown model handling)
• ruff check and ruff format clean
• Manual: run a small simulation, verify meta.json contains correct cost block

Closes #59

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.